It's all sarcastic. It's poking fun at how trendy advances might not necessarily be better than what existed before, as well as how much complexity is involved in them. It makes the most sense if you're already familiar with the concepts in the article (Docker, Heroku, CoreOS, etcd, and so on), and happenings in the technology industry related to them.
As usual, if a joke needs to be explained (to a person) then it's not funny (to them). I found this amusing since it aligned with my experiences:
> -Since no-one understands Paxos, this guy Diego…
> Oh, you know him?
> -No, he works at CoreOS. Anyway, Diego built Raft for his PhD thesis cause Paxos was too hard. Wicked smart dude. And then he wrote etcd as an implementation, and Aphyr said it wasn’t shit.
> What’s Aphyr?
> -Aphyr is that guy who wrote, ‘Call Me Maybe.’ You know, the distributed systems and BDSM guy?
> What? Did you say BDSM?
There are several jokes here based on cultural references related to the aforementioned topics.
The idea is you can deploy any app to any infrastructure of your choice (inside Docker containers). This means that you are not locked into Heroku and it gives you much more flexibility.
It's basically a hosted Rancher http://rancher.com/ service with a focus on a specific stack.
I think in the future, there will be a lot of services like Baasil.io (specializing in various stacks/frameworks) and managed by various open source communities.
Docker and Kubernetes WILL become more accessible to developers - I would bet my life on it.
I'm currently building a CLI tool to allow deploying in a single command - So you can get the simplicity of Heroku while not losing any flexibility/control over your architecture.
But all you've done is shift the problem to, er, you. If someone is uncomfortable taking out a dependency on say Azure or AWS - the two leading Docker hosting platforms, then they sure as hell aren't going to take out a dependency on "baasil.io" are they?
Um no, Rancher is open source can run and manage ANY infrastructure (including Amazon EC2) - You can run it on your own datacenter or even on your own local machine.
Also Baasil.io is essentially just a control panel/dashboard (Rancher-as-a-service), you can quit Baasil.io at any time and switch to your own hosted Rancher instance and you don't have to change any of your application code or change infrastructure providers.
The main benefit of Baasil.io is that it was built by an open source community using open source software and so we can offer the best possible support for apps built on top of our own open source stacks.
I took one look at "baasil.io" and saw the typical landing page with a "Plans" link at the top and questioned why anyone would want to take a dependency on this. If there is some OSS project behind the plan-based charade then that's fine.
"Also Baasil.io is essentially just a control panel/dashboard (Rancher-as-a-service), you can quit Baasil.io at any time and switch to your own hosted Rancher instance and you don't have to change any of your application code or change infrastructure providers."
But so you can with AWS and Azure - especially with their Docker offerings. So I'm not sure what problem baasil.io actually solves? If anything it just adds to the list of dependencies and points of failure.
The main benefit of Baasil.io is that it offers developers a boilerplate/framework which they can extend with their own code (to build scalable realtime apps and services) - Any app/service built on top of this boilerplate can be automatically scaled to 1000 hosts/nodes using a single command.
Users don't have to use Baasil.io, but if they do, they will get the best possible support - For example, a customer can give us access to their Rancher control panel and this would allow us to SSH into their machines to help resolve any problems in a hands-on way.
It's probably more accurate to describe it as "DevOps as a service - With a focus on realtime apps/services".
The value proposition is probably closest to Cloud 66 http://www.cloud66.com/ except more focused on realtime apps.
Another similar service is Zeit.co https://zeit.co/ - Except Zeit.co only runs Node.js. Baasil.io can be extended with components written in any language.
One thing that I liked about Heroku was the ease of deployment for Java applications (or any language for that matter as long as you had the right build pack)
Since they upped the cost of their small tier, I moved to Digital Ocean and installed Dokku, which gives me that Heroku-like deployment experience so managing my (admittedly very small) website isn't that much of a hassle.
We at Boxfuse (https://boxfuse.com) provide the same ease of deployment for running Java apps (as well as Go and Node.js ones) on AWS. All you need to type is literally: boxfuse run myapp-1.0.jar -env=prod
And you automatically get things like auto-scaling, database auto-provisioning. easy debugging and more.
What do you mean? I'm just running bytecode on a virtual machine ontop of a virtualized container on top of virtualized hardware on a CPU where the instruction set is virtualized in microcode...
show me an easy way to push a rails application to aws (with docker) that uses RDS ?
is there ANY way i can spin up a server, add the ssh keys to some configuration file somewhere and just "docker-magic push" and have my rails application running ?
or do "docker-magic bundle exec db:migrate" and have that command run on the server.
Or push a Procfile with worker definitions and have the PAAS automatically pick it up, add it to supervisord/systemd and run it ?
Ive been looking for the same thing. Tried dokku and it was painfully slow. Tried rolling my own solution via git hooks but other devs had problems with execution. I would love to hear more about people running dokku on Digital Ocean (I was on a beefy azure server)
In a very specific case, Heroku is the best solution for my problem. Sounds like it is for you too.
hey - not bad. The documentation looks very similar to heroku.. i cant find if they do stuff to handle workers in Procfiles (which need to be added to the job manager like supervisord/systemd).
But interesting..
Once you finish taking my Scaling Docker on AWS course, you'll have access to a magic command that deploys your app and you're free to optionally run migrations.
You should mention it costs $20 for information that can be found for free anywhere on the web. While other users answer the question, you simply refer to your piggy bank, shameless.
> You should mention it costs $20 for information that can be found for free anywhere on the web. While other users answer the question, you simply refer to your piggy bank, shameless.
Yep, it costs $20. Basically the cost of chinese food for 2, to ensure guaranteed victory in learning the essentials of AWS' platform while having a guided tour on how to deploy a fault tolerant web app with Amazon ECS from start to finish.
You can definitely learn everything for free, but the value in a course is that you're getting a cohesive learning path that was carefully planned and tested. You get a system that you can apply to your own projects and plenty of source code to reference.
You're paying the $20 so you can avoid spending 6 months trying to figure out everything on your own while stringing together a bunch of half-assed blog posts and tutorials.
You pay the small fee for certainty and it's well worth it because your time (and sanity) is not infinite.
I'm sorry you feel that way. Unfortunately time is something we all share, and it brings me a lot of joy to have students say that my courses have saved them a ton of time and helped them meet their goals.
I see that you recently left Amazon after making 200k/year there (a comment you made 5 days ago). I can see why you don't like people promoting Amazon products, I fully understand.
i have grappled with Amazon ECS and im in no mood to go through that again. Convox (YC S15) is a pretty good alternative - but it will take some time for them to tie the "magic" together.
Cloud Foundry is probably what you want. To the point that the Ruby buildpack code is a soft fork of Heroku's (source: I work on the Cloud Foundry buildpacks team). It'll run on AWS, vSphere, OpenStack, Azure, GCP is coming and others to follow.
There is, however, still a hump to get over in installation -- you need to learn what BOSH is, install BOSH, then install Cloud Foundry with BOSH. In the long run, for a production deployment, this is what you want. But it certainly doesn't feel that way when you just want to kick some damn tires.
If you just want to tinker, you can try PCFDev[0]. It's a fully-functional Cloud Foundry installation in a single VM.
Disclosure: I work for Pivotal, we donate the majority of engineering on Cloud Foundry.
thanks for this ! how do you look at deis.io vs Cloud Foundry. Do you see yourself subscribing to one of the popular camps out there... or will you stick to Cloud Foundry in the long term ?
I'm also with Pivotal... CF is a long term deal. Deis is a young startup that began on Dokku before pivoting to K8S, CF is in its third generation runtime and its use of containers predates Docker to 2012.
CF is already a relatively successful business with hundreds of million dollars in annual revenues across a pile of companies. Kubernetes and Docker are small in comparison as "businesses" but of course the momentum there is surging in terms of both pure open source adoption and contribution. It's likely going to be a big market with a lot of choice like Deis, or plain k8s, or RedHat OpenShift, or IBM bluemix, Pivotal , Docker Datacenter, or Mesos/Marathon, etc. It's a bit of a market war brewing and that competition will make for better solutions.
What is great about open source vs. past tech "gold rushes" is that these experiments and feedback loops exist across communities that are otherwise competing and overlapping. Mesos adopted the Docker image format independently of the Docker runtime ; Kubernetes introduced pods independently but also reused parts of Docker. Docker container networking and volumes are being used compatibly in the latest incubated CF releases. RedHat submitted a way to get CF style buildpacks working on K8S. Someone found a way to make CF run on Mesos; I could see a similar attempt on Kubernetes some day. It's a confusing and busy time but also an explosion of activity. And even if there is competition for dollars in the end among all these players that will lead to tension , the work is out in the open mainly.
Most of Cloud Foundry is built the way I like software to be built. Pair programming, TDD, small balanced teams, prioritising for user valu.
That style of development is actually baked into the Cloud Foundry Foundation rules. Companies who join the Foundation are expected to send engineers to ramp up on developing in this style. And voting rights are based on the number of full-time engineers you have assigned to the effort.
The reason I mention all this is that I trust the way we build Cloud Foundry. We still get production bugs and oversights and mistakes. It's around 4 million lines of code that turns into a distributed system of ~50 different interacting processes. We built a fully-featured, robust PaaS, using containers, in about 3 years, starting from scratch.
Nobody outside Google had built a container platform of this level before. Nobody but Heroku had built a fully-featured PaaS of this level before. We are, to my knowledge, the first system to do both of these things. Certainly the first opensource one.
The reason you never hear about Cloud Foundry is because we've already built all the components other folks are trying to roll up into full PaaSes. "It just works, already" is a boring story.
But again, quite seriously: I am obviously very biased.
This is awesome! So you're telling me that not telling the world that you have a high quality, tested, opensource, fully functional, container based Heroku-clone is deliberate ?
Please excuse me while I take an hour to digest that.
Right now the companies who provide the most engineering -- Pivotal and IBM -- are laser-focused on capturing enterprise dollars.
Which from a business perspective makes perfect sense. Pivotal's commercial distribution of Cloud Foundry (PivotalCF) holds the record for fastest-growing sales of any opensource product. Ever.
But there's not much effort on promoting to devs and startups (IBM are starting to do this more with BlueMix, which is their CF distribution). But it's early days.
So basically you'll usually find that I and a sprinkling of my colleagues show up in threads like these out of the goodness of our hearts and fondness for our work.
Obviously I have a financial interest. Pivotal makes money from PivotalCF, I work for Pivotal I'm on an options plan as well. So YMMV.
But I think Cloud Foundry is just ... way ahead, in terms of actually getting work done.
Edit: and since I'm musing aloud about business-y things, I should emphasise again that nothing I say is in an official capacity, consult your lawyer, financial planner and astrologer, etc etc.
fair enough, but I posit X users started consuming this API when SOAP was prevalent and Y users started when ReST was prevalent and Y >> X. Furthermore, SOAP is hard to maintain these days because it's so ancient. i.e. the libraries are not new and/or actively maintained.
As such, I maintain SOAP should be gone for the good of the running system.
In Python you simply don't have good SOAP libraries. They were all started at the tail end of its popularity and then all died quiet deaths when attention shifted to ReST before they were actually production ready, and if you now want to talk to a SOAP service… well, better don't do it in Python. 2, that is. Forget about 3.
Nope. We needed one last September, zeep didn't yet exist back then.
And going by the bugtracker, it's running into quite a few problems with almost-but-not-quite compliant servers/WSDL files, which is a real issue when you're trying to interface ass-old legacy APIs (we're talking "not upgraded since 2006"-old) made by $BigEnterprise. Maybe this time the project won't die before they work out all the little kinks.
I'm working with a very well-known American company with over $4b annual revenue that shall remain nameless and is currently developing a new SOAP API to replace the existing "dump a CSV on an FTP server" integration.
If that was ever true it certainly doesn't seem to be true now. All the tools support WSDL-first. All the tools are compatible with each other. Fill in the URL, let it autogenerate the interface, write your code and it all just works.
Because it was the latest trend 10+ years ago, and now people have made it just work because their applications are all built on it and people need actually good tools to use these architectures. It's always about tools, it's not like TCP is the best protocol or anything, it just has the best tooling, ditto for C, POSIX, etc. anything can be a good standard after 15+ years of work on it. Containers will be like that in a couple of years. It's all just cycles man.
Read it and passed it around to the guys at the office and I can tell you that it goes beyond the Heroku v Docker debate. For us, here at some kind of bank development shop it is about languages and frameworks. I would advice that you read the whole thing. Pretty funny I must add.
https://www.gitignore.io runs on Heroku and it gets 40k+ visitors a month on a free Dyno. I use Heroku because I don't want an IAAS solution, I want a PAAS solution. If I wasn't using Heroku, I would probably find another PAAS, before switching to Docker (even though I do love Docker).
That's 1 request per minute, you could slap that into a raspberry pi on a 3G connection, and do the same for your next 400 apps that has 1 request per minute.
People seem to underestimate just how powerful modern machines really are. And I don't get why people seem to think it's hard to deploy simple web applications. Just write a 4-line shell script that rsync's, runs whatever DB migrations you may have and restarts the thing.
40k visitors per month isn't one per minute. It's anywhere from one per minute to 40k per second. Division of requests by time is the worst possible mistake in calculating load.
I have 200k visitors per month generating 8m page views and about 50m hits on the servers (with CDNs taking another few hundred million hits).
These all peak during the UK weekdays and wind down at nights and weekends.
Divisions over time aren't going to work, but neither is translating visitors into requests, especially as it's only the page views that have a beyond trivial computation cost.
Thank you for the correction, but that was not the point I was trying to make. I was referring to his statement:
> [...] before switching to Docker (even though I do love Docker).
This seems like a very drastic solution to problems he does not yet have. I've been responsible for similar thoughts before, but shortly after realized just how damn stupid I was.
I definitely understand that modern machines are powerful, but I used to STIG RHEL 4 and RHEL 5 boxes for 4 years as my primary day job. I've done everything from create creating kickstart files to manually locking down a whole Linux instances to creating RPM files. At this point in my career I just don't care about the extra cycles I get by using <insert your infrastructure tech>.
If my product needed the extra performance, trust me I would switch.
Thanks for providing that - I just notices that your shell instructions don't include completion. I use the following in zsh: https://gist.github.com/lorenzhs/ad6c009f5748d333b73376e07ae...
With that, I can do "gi <tab>" to get a list of all possibilities.
I'm literally in the process of switching to zshell over the next few weeks. Since it's more than a 1-liner, could I just add that to the Advanced CLI instructions[1] section and credit you?
> So I just need to split my simple CRUD app into 12 microservices, each with their own APIs which call each others’ APIs but handle failure resiliently, put them into Docker containers, launch a fleet of 8 machines which are Docker hosts running CoreOS, “orchestrate” them using a small Kubernetes cluster running etcd, figure out the “open questions” of networking and storage, and then I continuously deliver multiple redundant copies of each microservice to my fleet. Is that it?
exactly. I mean look, if you have a lifestyle business that's only going to support 5-10 people, it's totally a waste of time. if you have some hope of scaling this is the way to go. I get it, just use Heroku. It's easy and convenient. If you're planning on a billion dollar exit, this way is way better.
"-No, look into microservices. It’s the future. It’s how we do everything now. You take your monolithic app and you split it into like 12 services. One for each job you do.
That seems excessive"
A 100 times yes. We tried to split our monolithic Rails app into micro-services built in Go. 2 years and many fires later, we decided to abandon the project. It was mostly because the monitoring and alerting were now split into many different pieces. Also, the team spent too much time debating standards etc. I think micro-services can be valuable, but we definitely didn't do it right, and I think a lot of companies get it wrong. Any positive experiences with micro-services here?
We were looking to break up our monolithic rails app into micro-services so devs could iterate and develop faster. We also thought that the application as a whole would become more failure resistant. Unfortunately, inter-dependancies among the services themselves meant that the failure-resistance didn't pan out as we thought.
I think that splitting into micro services is valuable if and only if you reach a scale where it makes sense to split into micro services. By scale, I mean the number of people on the team (if you have a lot of people, it can make sense to split into micro-services to limit communication bottlenecks between developers) or in term of traffic, in which case microservices can be very useful to better optimize the system piece by piece.
A small team starting a new project should not waste a single second considering microservices unless there's something that is so completely obviously decoupled in a way that not splitting it into a microservice will lead to extra work. It's also way easier to split into microservices after the fact than when you're developing a new app and you don't have a clue how it will look like or what the overall structure of the app will be in a year (most common case for startups).
I think it's more to do with a need rather than going straight just because you have enough people on a team. For instance, if you find that some of your processing/specific request handling can outperform better by using a different framework, programming language than the ones it's currently developed on, then you should definitely consider a microservice approach by decoupling that specific service/functionality from your current stack.
Careful with this one too. Usually adding new features to adapt to a changing marketplace can have new requirements across many of your services that need to be finished quickly. If those services are each in a different language, it can slow everything down by weeks or months.
Multiple platforms is not a problem and generally a good thing as long as it's not excessive. You don't want to be in a case where you have the same number of different platforms as developers or anything like that. I'm guessing there is a rule of thumb here, but I'm not sure what it would be. Max 1 different platform per 5 developers? Something like that.
OSGi makes it easy to end up with a cornucopia of tightly coupled nanoservices all running in the same JVM.
Unless you can coax dOSGi into working (which is tons of fun), then you can have services tightly coupled to other services running on entirely different machines causing frequent (and hilarious) cascades of bundle failures whenever the network hiccups.
OSGi is a trigger word for me now. I've worked on two large OSGi projects (previous job and current job) and it's always the same. Sh*t is always broken (and my lead still insists that OSGi is the one true way to modular bliss). And the OSGi fanboys always say "Your team is using it wrong!" Which very well might be true, but I no longer care. Apparently it's just too damn hard to get a team of code monkeys to respect service boundaries when OSGi makes it so damn easy to ignore them.
If I'm ever in a position of getting to design a new software architecture (hasn't happened in 10 years, but hey I can dream), I'll punch anyone who suggests "OSGi" to me right in the face.
Wholeheartedly agree on OSGi. Disastrous implementations. I worked with servicemix and still have nightmares around class path issues and crazy bundle scoping rules.
A plain old maven built jar with shading works much better in practice, but shading itself is shady :)
Well, I consider that by definition tightly coupled microservices should never be done. If it's not possible to decouple that function then it should not be in a micro service.
In practice micro services mean that you turn a function or method call into a network request. This doesn't really limit communication bottlenecks. It is often more difficult to argee on a network interface than on a simple function or object interface. It's also more difficult to change. You introduce a whole new set of failure modes due to going over the network. Debugging is more difficult since you now can no longer step through your program in a debugger but rather have an opaque network request that you can't step into. You can no longer use editor/IDE features like go to definition. It becomes harder to do integration tests. Version control becomes harder if the different services are in different repositories. A network request is much slower than a function call. You no longer have the advantage of a garbage collector for logical values that now cross network boundaries, and rather need to manually free them. Deployment is more difficult. The list is much longer than this, but I'd be interested in the counter-list: what are the advantages of micro-services?
Yes, but there's another dimension here. If another team breaks your build in a monolithic repo, you may or may not be able to resolve this quickly. You're in a contract with them about the state of the repo and thus your service.
With microservices, the production version of their service would conceivably be stable. It moves the contract from the repo to the state of production services.
> If another team breaks your build in a monolithic repo, you may or may not be able to resolve this quickly.
With a monolithic repo done right, the other teams broke their build of their branch, and it's up to them to resolve it. You, meanwhile, are perfectly happy working on your branch. When their changes are mergeable into trunk, then they may merge them, not before — and likewise for you.
With multiple repos, they break your build, but don't know it. You don't know it either, until you update your copies of their repos — and now you have to figure out what they did, and why, and how to update your logic to handle their new control flow, and then you update again and get to do it again, until finally you ragequit and go live in a log cabin with neither electricity nor running water.
> With multiple repos, they break your build, but don't know it. You don't know it either, until you update your copies of their repos — and now you have to figure out what they did, and why, and how to update your logic to handle their new control flow, and then you update again and get to do it again, until finally you ragequit and go live in a log cabin with neither electricity nor running water.
I don't see how this is a problem if you are pushing frequently and have a CI system. You know within minutes if the build is broken. If it broke, don't pull the project with the breaking changes.
My point is, I don't think one approach is inherently better than the other. Both require effort on the part of the teams to manage changes (or a CM team), and both require defined processes.
> If it broke, don't pull the project with the breaking changes.
I agree with the overall sentiment of your comment, but the quoted part is where I've seen trouble brew. The tendency is to be conservative about pulling updates to dependencies, which can easily get you into a very awkward state when a critical update eventually sits on top of a bunch of updates you didn't take because they broke you. It is usually better to be forced to handle the breakage immediately, one way or another.
I just want to add one disadvantage before I give some advantages. There's a lot of operational complexity involved in routing, monitoring, and keeping every instance of every microservice running. That complexity also makes debugging in production much more difficult, as one must track a relay of network requests through many separate layers to find the point where it actually got stuck.
As for advantages, microservices tend to keep code relatively simple and free from complex inheritance schemes. There's rarely a massive tangled-up engine full of special cases in the mix, as there often is in monolithic apps. This substantially decreases technical debt and learning curve, and can make it simple to understand the function an isolated microservice performs.
There is the obvious advantage that if you have disparate applications executing nearly-identical logic to read or write data to the same location, and the application platforms can't execute the same library code, you can centralize that logic into an HTTP API, which reduces maintenance burden and prevents potentially major bugs.
My opinion is that adopting microservices as a paradigm leads to a slow, difficult-to-debug application, primarily because people take the "micro" in microservices too seriously. One shouldn't be afraid to split functionality out into an ordinary service after it's been shown to be reasonable to do so.
I think that your comment is actually a pretty good test for when not to spin out a micro service.
If you expect to need to step into a function call when debugging, then it's too tightly coupled to spin out. You should be able to look at the arguments to the call and the response and determine if it's correct (and if not, now you have isolated a test case to take to the other service and continue debugging there).
If the interface will change so often that you expect it will be a problem that it's in a separate repository, if you expect that you will always need to deploy in tandem, then it's too tightly coupled to spin out.
The advantage of micro services is the separation in fact of things that are separate in logic. The complexity of systems grows super-linearly, so it's easier to reason about and test several smaller systems with clear (narrow) interfaces between them than one big. It's easier to isolate faults. It's harder to accidentally introduce bugs in a different part of the system when the system doesn't have a different part. If done right, scaling can be made easier. But these are hard architectural questions, there's no clear-cut rule for when you should spin off a new service and when you should keep things together.
Someone else mentioned separating the shopping app from the payment system for an ecommerce business, which even has security benefits. I think that's an excellent example.
Edit: Another clear benefit is that you can choose different languages, libraries, frameworks and paradigms for different parts of the code. You can write your boring CRUD backend admin app in Ruby on Rails, your high-performance calculation engine in Rust and your user-facing app in Node.js (so the front- and backend an share Javascript validation code).
> You introduce a whole new set of failure modes due to going over the network.
A thousand times yes. Distributed systems are hard.
> Debugging is more difficult since you now can no longer step through your program in a debugger but rather have an opaque network request that you can't step into.
Yes. Folks underestimate how difficult this can be.
In theory it should be possible to have tooling to fix this, but I've not seen it in practice.
> You can no longer use editor/IDE features like go to definition.
Not a problem with a good editor.
> Version control becomes harder if the different services are in different repositories.
No organisation should have more than one regular-use repo (special-use repos, of course, are special). Multiple repos are a smell.
> No organisation should have more than one regular-use repo (special-use repos, of course, are special). Multiple repos are a smell.
I would modify this slightly. Larger organizations with independent teams may want to run on per-team repos. Conway's law is an observation about code structure but it sometimes also makes good practice for code organization. And of course, sometimes the smell is "this company is organized pathologically".
Another problem is that large monolithic repositories can be difficult to manage with currently available software. Git is no panacea and Perforce isn't either.
I guess Facebook, Twitter, and Google are doing things "flat out wrong", then. Yes, that's a weak argument (argument from authority) but it is true that monolithic repositories have major advantages even for organizations with multiple products. Common libraries and infrastructure are much easier to work with in monolithic repositories.
My personal take on it, at this point, is that much of our knowledge of how to manage projects (things like individual project repos, semantic versioning, et cetera) is centered on the open-source world of a million mostly-independent programmers. Things change when you work in larger organizations with multiple projects. You even start to revisit basic ideas like semantic versioning in favor of other techniques like using CI across your entire codebase.
Those are huge organizations with commensurately large developer resources, and they simply work at a different scale than most people on HN. "It works for Google" is not an argument for anything.
Monorepos come with their own challenges. For example, if any of your code is open source (which means it must be hosted separately, e.g. on Github), you have to sync the open-source version with your private monorepo version.
Monorepo are large. Having to pull and rebase against unrelated changes on every sync puts an onerous burden on devs. When you're remote and on the road, bandwidth can block your ability to even pull.
And if you're going to do it like Google, you'll vendor everything -- absolutely everything (Go packages, Java libraries, NPM modules, C++ libraries) -- which requires a whole tool chain to be built to handle syncing with upstream, as well as a rigid workflow to prevent your private, vendored fork from drifting away from upstream.
There are benefits to both approaches. There is no "one right way".
It seems we agree, we are both claiming that "there is no one right way".
I love Git, and I used submodules for years in personal projects. It started with a few support libraries shared between projects, or common scripts for deployment, but it quickly ballooned into a mess. I'm in the process of moving related personal projects to a monolithic repository, and in the process I'm giving up the ability to tag versions of individual projects or provide simple GitHub links to share my code.
Based on these experiences, I honestly think that the only major problem with monolithic repositories is that the software isn't good at handling it, and this problem could be solved with better software. If the problem is solved at some point in the future, I don't think the answer will look much like any of the existing VCSs.
Based on experiences in industry, my observation is that the choice of monolithic repository versus separate repository is highly specific to the organization.
> No organisation should have more than one regular-use repo (special-use repos, of course, are special). Multiple repos are a smell.
Totally agree with everything else, but gotta completely disagree on this last point. Monorepos are a huge smell. If there's multiple parts of a repo that are deployed independently, they should be isolated from each other.
Why? Because you're fighting human nature, otherwise. It's totally reasonable to think that once you excise some code from a repo that it's no longer there, but when you have multiple projects all in one repo, different services will be on different versions of that repo, and your change may have changed semantics enough that interaction bugs across systems may occur.
You may think that you caught all of the services using the code you refactored in that shared library, but perhaps an intermediate dependency switched from using that shared library to not using it, and the service using that intermediate library hasn't been upgraded, yet?
When separately-deployable components are in separate repositories, and libraries are actual versioned libraries in separate repositories these relationships are explicit instead of implicit. Explicit can be `grep`ed, implicit cannot, so with the multi-repo approach you can write tools to verify that all services currently in production are no longer using an older, insecure shared library, or find out exactly which services are talking to which services by the IDLs they list as dependencies.
While with the monorepo approach you can get "fun" things like service A inspecting the source code of service B to determine if cache should be rebuilt (because who would forget to deploy service A and service B at the same time, anyways...), as an example I have personally experienced.
My personal belief is that the monorepo approach was a solution back when DVCSs were all terrible and most people were still on centralized VCSs like Subversion that couldn't deal with branches and cross-repo dependencies well, and that's just what you had to do, while Git and Mercurial, along with the nice language-level package managers, make this a non-issue.
Finally, there's an institutional bias to not rock the boat (which I totally agree with) and change things that are already working fine, along with a "nobody got fired buying IBM" kind of thing with Google and Facebook being two prominent companies using monorepos (which they can get away with by having over a thousand engineers each to manage the infrastructure and build/rebuild their own VCSs to deal with the problems inherent to monorepos that most companies don't have the resources and/or skills to replicate).
EDIT: Oh, I forgot, I'm not advocating a service-oriented architecture as the only way to do things, I'm just advocating that whatever your architecture, you should isolate the deployables from each other and make all dependencies between them explicit, so you can more easily write tooling to automatically catch bad deploy states, and more easily train new hires on what talks to/uses what, since it's explicitly (and required to be) documented.
If that still means a monorepo for your company's single service and a couple of tiny repos for small libraries you open source, that's fine. If it means 1000 repos for each microservice you deploy multiple times a day, that's also fine (good luck!).
> It's totally reasonable to think that once you excise some code from a repo that it's no longer there, but when you have multiple projects all in one repo, different services will be on different versions of that repo, and your change may have changed semantics enough that interaction bugs across systems may occur.
But having multiple repos doesn't prevent the equivalent situation from happening (and, I think, actually makes it much likelier): no matter what, you have to have the right processes in place to catch that sort of issue.
> You may think that you caught all of the services using the code you refactored in that shared library, but perhaps an intermediate dependency switched from using that shared library to not using it, and the service using that intermediate library hasn't been upgraded, yet?
That's the sort of problem which happens with multiple repos, but not (as often) with a single repo.
> Explicit can be `grep`ed, implicit cannot, so with the multi-repo approach you can write tools to verify that all services currently in production are no longer using an older, insecure shared library, or find out exactly which services are talking to which services by the IDLs they list as dependencies.
A monorepo is explicit, too, even more explicit than multiple repos: WYSIWYG. And you can always see if your services are using the same API by compiling them (with a statically-typed language, anyway).
The beautiful thing about a monorepo is it forces one to confront incompatibilities when they happen, not at some unknown point down the road, when no-one know what changed and why.
You don't debug distributed systems by tracing into remote calls and jumping into remote code. You debug it by comparing requests and responses (you use discrete operations, right) with the specified requests and responses, and then opening the code that has a problem¹.
It calls for completely different tooling, not for a "better debugger".
1 - Or the specs, because yes, now that your system is distributed you also have to debug the specs. Why somebody would decide on doing that for no reason at all? Yet lots of people do.
I'd also add that microservices have increased value if you begin with such an architecture in the first place. It's much more difficult to "gracefully" rip an existing monolith into modular pieces than to build modularly from the start.
I don't like correcting with "well, actually", however, I have to say that the author of the book "Building Microservices" in his first few chapters (in particular: Chapter 3: Premature Decomposition) warns against using microservices with new apps, especially if you are new to the domain. He claims that they are actually easier to use when you have to refactor a large monolith, and that normally you shouldn't start with microservices unless you know what you are doing - therefore my criticism towards the article which starts with a pre optimization (split one service in 12), which seems to be a common, yet arguable practice.
This has not been my experience. I've seen a few projects where microservices had been added from the start because it's the thing to do and, in all cases, it didn't work well. It's extremely difficult to split in microservices if you do not have a clear big picture of your projects functions and coupling. And, in most cases, in new projects, you don't have that big picture.
Microservices also make it much harder to refactor the code which you often need to do in the early stage of a project.
> A small team starting a new project should not waste a single second considering microservices unless there's something that is so completely obviously decoupled in a way that not splitting it into a microservice will lead to extra work.
That's a good point. I think this thought extrapolates to other parts of software engineering as well. Sometimes writing very modular and decoupled software from the beginning is very hard for a small team, and we can't see well if this is the best approach since it's also hard to grasp the big picture.
I'm currently facing this issue. I'm trying to write very modular and reusable applications, but now I'm paralyzed trying to picture the best patterns to use, where should I use a facade, a decorator, etc. I think I'll adopt this strategy for myself--only focus on modularizing from the beginning if it'd lead to extra work otherwise.
saw done properly once, with good separation of services etc, but it still required a truckload of commitment to make it work, because one thing is interface changes when the code just don't work when you merge, another is figuring out the publish and restart order of each service when you have to add an operation so you don't have to knock out the whole system at every upgrade.
I don't really like that model applied to everything, but eh now you are kind of forced in a hybrid approach - say, your macro vertical plus whatever payment gateway service, intercom or equivalent customer interaction services, metrics services, retargeting services, there are a lot of heterogeneous pieces going into your average startup.
but back on topic, what Docker really needs now is a whack on the head of whoever thought swarms/overlays and a proper, sane way to handle discovery and fail-over - instead we got a key-value service deployment to handle, which cannot be in docker and highly available unless you like infinite recursion.
>Any positive experiences with micro-services here?
It makes sense for some thing. We run a webshop, but have a separate service that handles everything regarding payments. It has worked out really well, because it allows us to fiddle around with pretty much everything else and not worry about breaking the payment part.
It helps that it's system where we can have just one test deployment and everyone just uses that during testing of other systems.
I've also work at a company where we had to run 12 different systems in their own VMs to have a full development environment. That sucked beyond belief.
The idea of micro-service are is enticing, but if you need to spin up and configure more than a couple to do your work, it starts hurting productivity.
> have a separate service that handles everything regarding payments. It has worked out really well, because it allows us to fiddle around with pretty much everything else and not worry about breaking the payment part.
Is the payments service a single service that manages the whole transaction, or have you go for multiple services handling each part and, if so, how did you manage failure with a distributed transaction?
Not sure if it's the case here. But what works really well for us is queues with at-least-once guarantee. (For payment services you might need an additional check to guarantee exactly one execution.) I think you can find this queue offered by most providers.
We had almost the same story with payments. Except for we've jumped to a payment-processing SaaS but got dissatisfied (all those SaaSes I saw don't work with PayPal EC without so-called "reference transactions" enabled) and decided that wasn't a good idea and we have to jump back to in-house implementation.
I didn't want to re-integrate the payments code back to the monolith - thought it would take me more time and make code messier. So I wrote a service (it's small but to heck with "micro" prefix) that resembled that SaaS' API (the parts we've used). It had surely evolved and isn't compatible anymore, but it doesn't matter as we're not going back anyway.
Works nicely and now I'm feel more relaxed - touching the monolith won't break payments.
On the other hand, I see how too many services may easily lead to fatigue. Automated management tooling (stuff like docker-compose) may remedy this, but also may bring their own headaches.
I don't think having a handful of services handling a specific, atomic section of the app really classifies as this 'micro-services' claim, it's just smart separation of concerns.
We have specific services that process different types of documents, or communicate and package data from different third parties, or process certain types of business rules, that multiple apps hook into, but it's literally like 20 services total for our department, some that are used in some apps and not others.
When I hear 'micro-services' I'm picturing something more akin to like node modules, where everything is broken up to the point where they do only one tiny thing and that's it. Like your payment service would be broken into 20 or 30 services.
But maybe I'm mistaken in my terms. I haven't done too much with containers professionally, so I'm not too hip with "the future".
> Any positive experiences with micro-services here?
Yep. We already had a feature flag system, a minimal monitoring system, and a robust alerting system in place. Microservices make our deployments much more granular. No longer do we have to roll back perfectly good changes because of bugs in unrelated parts of the codebase. Before, we had to have involved conversations about deployments, and there were many things we just didn't do because the change was too big.
We can now incrementally upgrade library versions, upgrade language versions, and even change languages now, which is a huge win from the cleaning up technical debt perspective.
How granular are your services? I've heard a lot of talk about microservices without much talk about how micro they are. As someone who's happy with the approach, would you mind giving a little context in terms of what sort of degree you've split the system down?
They are of varying sizes. We have maybe 3-4 per developer. And they vary in size. The smallest are maybe 5-6 python classes.
To be honest, we still have a monolithic application at the heart of our system that we've been slow to decompose, though we're working on it. We deploy it on a regular cadence and use feature flags heavily to make it play nice with everything else.
"We rolled out this update with 220 changes. There's a breaking bug. Where is it? We need to find out in the next 5 minutes, revert, and deploy. Otherwise we have to revert the whole thing- we're losing money."
Git doesn't really help with that. More granular deployments do, and if microservices help with more granular deployments, go for it.
Only if you have a test that catches the bug and it still needs time to run. You'll also need time to write a fix, validate it and deploy. Plus any extra time your organization needs between code and deployment
It helps with the find part. The revert and deploy not so much, especially say if it's the middle commit if 200 and you'd still like to deploy all the commits before and after.
My experience is that most devs don't even know how to use SVN correctly. I just had a conversation with someone waiting for me to finish something before they could branch. The idea that I could merge my change into their branch afterwards didn't occur to them.
I'm building a podcast discovery app and I find myself being de-facto pulled towards modularity. It's because my feed checker is in Elixir, my site is WordPress-based, and I communicate between them using the WP API, and I'm using Google Cloud SQL, and Elasticsearch on its own virtual machine...
The thing is though, the Elixir feed checker has its own database table that tracks whether it's seen an episode in a feed. And when there's a new episode it sends an API call to WP to insert the new post. The problem is that sometimes the API calls fail! Now what? I'll need to build logging, re-try etc. So I'm thinking of making the feed checker 'stateless' and only using WP with a lot of query caching as the holder of 'state' information about whether an episode has been seen before.
To sum up my experience so far, there's something nice about being able to use the right tech for each task, and separating resources for each service, but the complexity--keeping track of whether a task completed properly--definitely increases.
Care sharing your progress with the podcast discovery app? I'm a cofounder of Podigee which is a podcast hosting service. Maybe we can exchange know-how, find some synergies or even join forces on certain topics. Feel free to drop me a line at mati@podigee.com
I think your problem might be not expanding wordpress. PHP will gladly do the above through wordpress plugins and cron jobs. I built something similar and ended up going that way. The system is still running to this day. Sure, its not hip but gets the job done with miniml fuss and makes money.
You are right that PHP can do the feed checking part but I wanted to use something with easy async/concurrency out of the box (and wanted to learn Elixir instead of using Node.)
One hard tech limit is that with 50k podcasts, 4million+ episodes, search definitely doesn't work well. Not just WP, but SQL itself. Hence Elasticsearch. I also plan to work on recommendations, etc. so will need probably to be exporting SQL data into other systems anyway for making the "people who liked this also liked this" kinda things.
Also I kinda lied about using the WP API--that's how I built the system initially (and will switch to it moving forward), but to import the first few million posts from the content of the feeds, I just used wp_insert_post against the DB of new entries that Elixir fetched (I posted the code I used here: http://wordpress.stackexchange.com/a/233786/30906).
I also plan to write the whole front-end in React (including server side rendering) so will have to figure out how to get that done. Would probably use the WP-API with a Node.js app in front of it, will look into hypernova from AirBNB. So probably more usage of WP API accessed by another service...
I hope you are not doing all of this alone. I'd try and keep things as simple as possible within a monolith and then improve as needs increase. Good luck :)
I generally write what I consider monolithic Django apps. I would add in Haystack (a search module for Django) and configure it to use Elasticsearch to overcome the problems you describe.
It doesn't sound like microservices are needed, just adding in the appropriate tech for the job.
That's an incisive question. My impression, which may be mistaken, is that a cronjob would be used to move data (pages compiled from templates, chart images, etc.) into the PHP host on a "batch" basis. To me, that implied the existence of other systems that handle the data in their own way, but I guess in this thread the salient difference between micro and mono is that the former connects components via a web stack. Are there more agile interfaces available for cronjobs? If instead we're only considering transformations of data already resident on the host (as what, flat files?), I don't imagine that cronjobs are the best solution available.
I do have a positive micro-service experience, and although we are still in that process of breaking down our monolith SOA based app, we have seen the benefits already.
The more dramatic effect was on a particular set of endpoints that have a relative high traffic (it peaks at 1000 req/s) that was killing the app, making upset our relational database (with frequent deadlocks) and driving our Elasticsearch cluster crazy.
We did more than just split the endpoints into microservices. We also designed the new system to be more resilient. We changed our persistence strategy to make it more sensible to our traffic using a distributed key-value database and designed documents accordingly.
The result was very dramatic, like entering into a loud club and suddenly everything goes silent. No more outages, very consistent response times, the instances scaled with traffic increase very smoothly and in overall a more robust system.
The moral of this experience (at least for me) is that breaking a monolith app into pieces has to have a purpose and implies more than just move the code to several services keeping the same strategy (that's actually slower, time consuming and harder to monitor)
Which one of the four major improvements do you attribute the success to though? Could you have done the work on making it more resilient, persistence, sensible, redesign the docs without breaking into micro-services and still have seen the positive results?
I don't think the level success comes from one dimension, but I don't think either that we could have achieved the resiliency without breaking it in micro-services (or just services that happened to be small if you will).
One key factor was decoupling the high volume updates from the users requests so one didn't affect the other one.
Do you think the result could also be a dramatic improvement if you kept old system and do those other things except splitting into microservices?
I can't get my head around how people introduce changes to their system if they have to update 12 different microservices at once? It must be horrible.
Often you hear stories how people are converting monolithic app to microservices - but this is easy. Rewriting code is easy and it's fair to say it always yields better code (with or without splitting into microservices - it doesn't matter).
What I'd like to hear is something about companies doing active development in microservice world. How do they handle things like schema changes in postgres where 7 microservices are backed by the same db? What are the benefits compared to monolithic app in those cases?
It seems to me that microservices can easily violate DRY because they "materialise" communication interfaces and changes need to be propagated at every api "barrier", no?
>Do you think the result could also be a dramatic improvement if you kept old system and do those other things except splitting into microservices?
As I said in another thread, the separation in different components was key for resiliency. That allowed independence between the higher volume update and the business critical user facing component.
>I can't get my head around how people introduce changes to their system if they have to update 12 different microservices at once? It must be horrible.
The thing is, if you design the microservices properly it is very rare to introduce a change in so many deployments at once. Most of the time is just 1 or 2 services at a time.
>What I'd like to hear is something about companies doing active development in microservice world. How do they handle things like schema changes in postgres where 7 microservices are backed by the same db? What are the benefits compared to monolithic app in those cases?
We don't introduce new features in our monolith service anymore. So, from that perspective we do all active development in microservices.
>"How do they handle things like schema changes in postgres where 7 microservices are backed by the same db?
The trick is, you want to avoid sharing relational data between microservices. I don't know if it is just us, but we have been able to split our data model so far and in most cases we don't even need a relational database anymore, so having a schemaless key/value store makes seems easy too.
> What are the benefits compared to monolithic app in those cases?"
There are several advantages, but the critical one for me is being able to have a resilient platform that can still operates even if a subsystem is down. With our monolithic app is an all or nothing thing. Another advantage is splitting the risk of new releases.
>It seems to me that microservices can easily violate DRY because they "materialise" communication interfaces and changes need to be propagated at every api "barrier", no?
Not necessarily. YMMV but you can have separation of concerns and avoid sharing data models. When you do have shared dependencies (like logging strategy or data connections) you can always have modules/libraries.
Multiple microservices are supposed to have different data backends, so that they are completely independent. Splitting your data up this way isn't all roses, but ideally the services are isolated so an update to one doesn't affect the others.
"Message passing as the fundamental operation of the OS is just an excercise in computer science masturbation. It may feel good, but you don't actually get anything DONE."
-- Linus Torvalds
He said that about micro-kernels but it may apply to micro-services as well.
Breaking things apart too much will only cause you grief.
Microservices is a technology. IMO, like any other tech, it should be used when there are clear benefits expected in the near future, not as a blanket "always microservices" policies.
Although I personally had to deal with some monolithic monsters that I wished were split into smaller services.
The problem with microservices is that your state is spread over multiple systems. You completely lose the concept of transactional integrity, so you will have to work around that from the start.
The advantage though is that APIs (system boundaries) are usually better defined.
Perhaps one should use the best of both worlds, and run microservices on a common database, and somehow allow to pass transactions between services (so multiple services can act within the same transaction).
One of the big advantages of microservices is scaling/migrating the databases behind each service independently. If you need transactions across multiple services then one could argue that either your API endpoint is doing too much, or your services are doing too little. It's not perfect, and certainly not always convenient, but it's a balance. Microservices with a common DB is asking for trouble. The monolith is a better option in that case IMO.
Say you sell a widget. You want to update both your cash account and your inventory, and never one without the other. Which is easier to understand and more reliable: doing them atomically, or making sure you have designed in 2^n intermediate states and all the code required to complete work that should happen but hasn't yet?
The problem with microservices is that your state is spread over multiple systems.
Then again, sometimes it's advantageous to identify parts of your system where aspects of state can be safely decoupled. And in which having them reside in disparate systems (and yes, sometimes be inconsistent or differently available) might actually be a better overall fit.
You completely lose the concept of transactional integrity, so you will have to work around that from the start.
Then again, sometimes your state changes not only don't need to be transactional; it can be disadvantageous to think of them that way.
> Then again, sometimes your state changes not only don't need to be transactional; it can be disadvantageous to think of them that way.
I'm curious; in what kinds of situation would this apply?
> Depends, depends, depends.
Flexibility is usually an important requirement. Often you cannot freeze your architecture and be done with it. I think a transactional approach could better fit with this.
I'm curious; in what kinds of situation would this apply?
Any situation where the business value of having your state be 100% consistent does not outweigh the performance or implementation cost of making it so.
> Any positive experiences with micro-services here?
In my experience, any monolith that can be broken up into a queue based system will benefit enormously. This cleans up the pipelines, and adds monitoring and scaling points (the queues). Queues removes run-time dependencies to the other services. It requires that these services are _actually_ independent, of course.
I do, however, avoid RPC based micro-services like the plague. RPC adds run-time dependencies to services. If possible, I limit RPC to other (micro) services to launch/startup/initialization/bootstrap, not run-time. In many cases, though, the RPC can be avoided entirely.
Probably more. I'd say, like, at least 30-40 years.
I mean, the infamous "UNIX way" of "do one thing, do it well" (something we nearly lost with popularity of "do everything in a manner incompatible with how others do it" approach in too many modern systems), when complex behavior was frequently achieved through the modularity of smaller programs communicating through well-defined interfaces.
Heck, microkernels are all about this, and their ideas haven't grew out of nowhere. And HURD (even though it was never finished) is quarter a century old already.
Oh yeah, there's certainly other ways of doing it. My own experience is just that message queuing seems to be the default loosely-coupled RPC mechanism for larger orgs (from before the term 'micro services' was popular).
was it a complete rewrite? I don't think thats the right way to transition. Why didn't you try to separate the features one-by-one from the monolith? That would give more immediate feedback and real problems to work on instead of the possibility to get stuck on the holy architecture debate.
No it wasn't a complete rewrite. We started by separating out the most mission critical components. Maybe that's where we went wrong; the most mission critical components were quite large and unwieldy to split out all at once. There was also the overhead of keeping the newly separated out component and the monolithic app in sync.
I only use a microservice if its something that can operate by itself. Things like a file data store, reports generation, etc. But all business logic goes in the monolith.
positive experience with microservices: identify discrete functions that can considered "stateless" (i.e. no side effects, deterministic output for given input) and factor those out into stand-alone microservices.
a good example of this that I've used in production at my current $dayjob: dynamic PDF generation. user makes request from our website, request data is used to fill out a pdf template context which is then sent over to our PDFgen microservice which does its thing and streams a response back to the user.
Does it connect to the database to fill in some values in the template? Does it keep connection pool of let's say 5 connection always open (as libraries like to do)? Does it have authentication? Is it public or private API? Who is managing security? Is it running behind it's own nginx or other proxy? Does it have DoS protection (PDF generation can be CPU intense)? What about the schema for request? How do you manage changes to the schema? They need to be deployed together with changes in other services, right? What about changes to database schema - you need to remember to update that service as well and redeploy it at the right time as well - just after successful db migrations - which live in another project.
All of that and much more needs to be replicated for each microservice, right?
Why not just have a module in your monolithic app that does it. The logic will still be separate. In most languages/frameworks you can spawn pdf generation task. Any changes are easier to introduce as well. There's no artificially materialised interface. Updates are naturally introduced. All auth logic is there already, you don't need to worry about deploying yet another service, same with logging etc.
> Does it connect to the database to fill in some values in the template?
the template has values that are related to database models. the main app (still mostly monolithic) fills out the template context. the context itself is what's passed to the microservice. the microservice does not connect to a database at all.
> Does it keep connection pool of let's say 5 connection always open (as libraries like to do)?
no. the service probably handles a few hundred requests per day, it is not in constant use. communication is over HTTPS. it opens a new connection on each request. this does impact throughput, but its a low throughput use case, and pdf rendering itself is much slower and that time totally dominates the overhead of opening and closing connections anyway.
> Does it have authentication?
yes, it auths with a bearer token that is borne only by our own internal server. this is backend technology so we don't have to auth an arbitrary user. we know in advance which users are authorized.
> Is it public or private API?
private
> Who is managing security?
we are, with a lot of assistance from the built-in security model of AWS.
> Is it running behind it's own nginx or other proxy?
the main app is behind nginx. the microservice is running in a docker container that exposes itself over a dedicated port. there's no proxy for the microservice, again, because of the low throughput/low load on the service. no need to have a load balancer for this so the most obvious benefit of a proxy wasn't applicable.
> Does it have DoS protection (PDF generation can be CPU intense)?
yes, it's an internal service and our entire infrastructure is deployed behind a gatekeeper server and firewall. the service is inaccessible by outside requests. the internal requests are queue'd up and processed 1 at a time.
> What about the schema for request?
request payload validation handled on both ends. the user input is validated by the main app to form a valid template context. the pdf generator validates the template context before attempting to generate one also. its possible to have a valid schema that has data that can't be handled correctly though. errors are just returned as a 500 response though. happens infrequently.
> They need to be deployed together with changes in other services, right?
nope. the microservice is fully stand alone.
> What about changes to database schema - you need to remember to update that service as well and redeploy it at the right time as well - just after successful db migrations - which live in another project.
the microservice doesn't interact with a database at all. schema changes in the main app database could potentially influence the pdf template context generation, but there are unit tests for that, so if it does happen we'll get visibility in a test failure and update the template context generation code as needed. none of this impacts the microservice itself though. it is fully stand alone. that's the point.
> All of that and much more needs to be replicated for each microservice, right?
in principle yes, and these are good guidelines for determining what is or is not suitable to be a microservice. if it would need to auth an arbitrary user, or have direct database access, or be exposed to public requests, it might not be a good candidate for a microservice. things that can stand alone and have limited functional dependencies are much better candidates.
> Why not just have a module in your monolithic app that does it.
because the monolithic app is Python/django and the PDF generation tool is Java. one of the main advantages of microservices architecture is much greater flexibility in technology selection. A previous solution used Python subprocesses to call out to PDF generation software. It's actually easier and cleaner for us to use a microservice instead.
Yeah, we do micro services, the "real" kind. Not the "SOA with a new name kind", but the "some services are literally a few douzan lines of code and we have 100x the amount of services as we do devs" kind.
The thing is, you need a massive investment in infrastructure to make it happen. But once you do, its great. You can create and deploy a new service in a few seconds. You can rewrite any individual service to be latest and greatest in an afternoon. Different teams don't have to agree on coding standards (so you don't argue about it).
But, the infrastructure cost is really high, a big chunk of what you save in development you pay in devops, and its harder to be "eventually consistant" (eg: an upgrade of your stack across the board can take 10x longer, because there's no big push that HAS to happen for a tiny piece to get the benefits).
Monolithic apps have their advantages too, and many forget it: less devops cost, easier to refactor (especially in statically typed languages: a right click -> rename will propagate through the entire app) and while its harder to upgrade the stack, once its done, your entire stack is up to date, not just parts of it being all over. Code reuse is significantly easier, too.
Yeah, add in things like MORE THAN ONE PRODUCTION ENVIRONMENT and LETTING YOUR CUSTOMER HOST AN INSTANCE OF YOUR MICROSERVICES and you have guaranteed your own suffering.
It is a matter of tooling. One data center or ten, it does not matter much with proper tooling. We deploy to seven data centers with a click of a button, with rollback, staggered deployment etc. Centralized logging using ELK gives us great visibility in to each DC, without worrying about individual microservice instances.
Dockerized apps make it simple to run on dev environment, which is what we do. Of course , one cannot run everything on a laptop, we have a dev cluster
beyond just the docker environment, you only need to be able to run the service you're working on locally. Anything you don't run local should hit some shared dev/QA infrastructure (which share a db with local). Whatever you use to develop should be able to detect what you have running locally and prefer those when available.
Anything you're not running locally just hits the shared infra.
Seems like a terrible idea if "different teams" actually means "random assortment of developers for this specific project." If it actually means "different teams," e.g. you rarely if ever would move from one team to another, I don't see the issue if one team uses tabs and one uses spaces, or you have different naming conventions or whatever.
Well, you use the standard you like, and the other team can use the standard they like. Then you write a micro service to convert from one standard to the other.
Deployment has to be easy. Create a new service from scratch, including monitoring, logging, instrumentation, authentication/security, etc, and deploying it to QA/Production with tests has to take minutes from the moment you decide "Hey, I need a service to do this" until it's in prod.
Because individuals may be jumping through dozens of services a day, moving, refactoring, deploying, reverting (when something goes wrong), etc. It has to be friction-free, else you're just wasting your time.
eg: a CLI to create the initial boilerplate, a system that automatically builds a deployable on commit, and something to deploy said deployable nearly instantly (if tests passed). The services are small, so build/tests should be very quick (if you push above 1-5 minutes for an average service, it's too slow to be productive).
Anyone should be able to run your service locally by just cloning the repo and running a command standard across all services. Else having to learn something every time you need to change something will slow you down.
That infrastructure is expensive to build and have it all working together.
Building Microservices needs discipline, an eye to find reusable components and extracting them and as you said, investment in infrastructure.
Monoliths invariably tend to become spaghetti over time, and completely impossible to any non trivial refactoring. With microservices, interfaces between modules are stable and spaghetti is localized.
So that sounds pretty much like a function call in a monolithic app. How do you store state? I assume you need to between all those microservices.
>The thing is, you need a massive investment in infrastructure to make it happen.
I thought that one of the selling points of microservice architectures was the minimal infrastructure. I am really struggling to see an advantage in this way of doing things. You are just pushing the complexity to a dev ops layer rather than the application layer - even further form the data.
I wonder what language I would pick for that. I usually use Scala, but it seems a bit silly when the footprint of the platform would massively outweigh the actual service. I don't like Go. I like Python, but I prefer static typing. Rust seems a bit too low level (although I'd like to try it for embedded). I don't see any point in learning Ruby when i already know Python well.
Maybe Swift? Scala Native in a year or two? I've done a little Erlang before, so maybe Elixir?
Yes. A lot of success. And with only one person on the backend full time.
That said, in places where it doesn't make sense we didn't try to force it. Our main game API is somewhat monolithic, but behind it we have almost 10 other services. Here's a quick breakdown:
- Turn based API service (largest, "monolithic")
- Real-time API service (about 50% the size of turn-based)
- config service (serves configuration settings to clients for game balancing)
- ad waterfall service (dynamic waterfall, no actual ads)
- push notification service
- analytics collection service (mostly a fast collector that dumps into Big Query)
- Open graph service (for rich sharing)
- push maintenance service (executes token management based on GCM/APNS feedback)
- help desk form service (simple front-end to help desk)
- service update service (monitors CI for new binaries, updates services on the fly - made easy by Go binary deployment from CI to S3)
- service ping service (monitors all service health, responds to ELB pings)
- Facebook web front-end service (just serves WebGL version of our game binary for play on Facebook)
- NATS.io for all IPC between services
...and a few more in the works. Some of these might push the line of "micro" in that they almost all do more than a single function's worth of work, but that level of granularity isn't practical.
But don't get too caught up on the "micro" part. Split services where domain lines naturally form, and don't constrain service size by arbitrary definitions. You know, right tool for the job and whatnot.
> Also, the team spent too much time debating standards etc.
Ah yes, the 'let's have decentralised microservices with centralised standards!' anti-pattern. It results in lots of full-fledged, heavyweight, slow-to-update services, which also have all the problems of a distributed system. It's the worst of both worlds.
> It was mostly because the monitoring and alerting were now split into many different pieces
Well, there's your problem - you need a monitoring microservice and an alerting microservice! Well, those may be too coarse by themselves, but once you break them down into 5 or 6 microservices each, you'll be ready for production.
We did have a lead with a vision, and part of the vision was standards for each service (for example, file structure in Go). I can see the rationale behind it; a new dev can onboard very quickly on to a new service. But in hindsight, maybe it wasn't thought out enough.
> Any positive experiences with micro-services here?
I'm currently working on a large refactoring effort along these lines. The end goal is to create a modular, potentially distributed system that can be deployed in a variety of configurations, updated piecemeal for different customers, and integrated by our customers with the third-party or in-house code of their choice using defined APIs. We aren't typical of the other examples, though, in that we do literally ship our software to our customers and they run it on their own clusters.
Oh yes. We're splitting up a large monolith into a bunch of different services. Completely amazing, though there's a ton of tools (like Netflix's Hysterix, etc) that make it much, much easier to do.
I wouldn't, however, just "do microservices" from day one on a young app. But usually that young app has no idea what the true business value is, i.e., you have no idea what down time of certain parts of your services really means to the business. That's the #1 pain point we're solving: having mission critical things up 100%, and then rapidly iterating on new, less stable feature designs in separate services.
You should, however, keep an eye on how "splittable" everything is, i.e., does everything need to be in the same DB schema? Most languages have package concepts, which typically align (somehow) with "service" concepts. Do you know their dependencies? That sort of thing. Then, the later process of "refactor -> split out service" is pretty straightforward and easy to plan.
I don't know if the community will consider this example to be microservices as currently defined, but many years ago I wrote client-server systems using a PC-based fat client application, and transactions running on a CICS server. I think this was pretty similar to what people currently think of as a microservices architecture, although we didn't have to worry about running/monitoring multiple servers (all the transactions/services ran on a single mainframe server), and the transaction monitor managed things like start-up and shutdown pretty simply. This approach worked really well for us, and we built several robust, scalable applications using this approach. To be clear, we numbered our users in hundreds, not thousands or millions. I can well understand how scaling this approach across many servers could be very challenging.
You split off specific parts of your app into microservices because you want to scale those parts independently from the rest. It's not just a blind decomposition of a monolith for the sake of decomposition.
Out of topic but, man, is coreos not ready for prime time. I've been using it following the stable channel and ended up having to turn off automated updates because it would break docker (docker would just hang with the last 4 updates)... Not a very convincing test of coreos :)
I pine for the days of yore when the Unix Philosophy was strong and pure, and every program did one thing well, and only one.
Like the way the shell would fork off an "expr" sub-process to parse a mathematical expression to add two numbers, then write the result to a pipe via stdout, then terminate the process, clean up all its resources, and switch context back to the shell, which then read the serialized sum back in from the other end of the pipe, and went about its business, regardless of the fact that the CPU running the shell already had its own built-in "add" instruction in hardware.
This is the first coherent refutation of the "do one thing well" ethos I have ever read. Thanks for putting into words what I haven't been able to express myself.
Shell/batch scripting can often be useful in the devops world, where you have no guarantee that any additional tools (python, ruby, perl, powershell, whatever) wil be available.
Shell scripts are guaranteed to be runnable on all machines.
Unfortunately the shell "language" sucks, but still...
Shell languages are great at doing interactive programming. Few non-shell languages can match the convenience, flexibility, and expressiveness in that domain. Of course, in any other domain (scripts) they are awful.
I'm curious because these simple things that are delightfully easy in bash often turn out to be surprisingly tedious in other languages.
Of course, some things are tedious in bash too. But a basic principle of shell scripting is that you call other programs to do the stuff you don't want to do in shell.
I agree it is tedious, but to be honest, reading and writing to stdin/out isn't something that would commonly need to be done in a robust system. If the world were perfect you would use library functions.
I definitely think there is scope for a language that works well as an interactive shell, and as a general purpose language. They have somewhat conflicting constraints but I'm sure we can do better than Bash. Have you seen how [ is implemented?
Now you have to clean up 'a'. And decide whether /tmp or /var/tmp or a dir on some other filesystem has enough space to hold all of 'a' until 'bar' is finished. Is it a security problem that other processes could snoop the contents of 'a' or even tamper with it?
When you're not writing shell, just use the tools the language gives you.
For the matter, I think a shell script is cleaner than a python script for devops; but I don't think the composability of unix tools is that much of an advantage compared to the amount of python libraries out there.
When I use shell it's often exactly because I want to construct pipelines of processes and FIFOs and do all the other things that shell does very well and has done well for decades.
I'm likely to be using Python programs and other programs in those shell scripts. The beauty of shell is that it makes it so easy to compose programs written in different languages.
Everything in a computer is a stream of bytes... My shell scripts often use tools like jq and jshon to deal with JSON structures, etc. File hierarchies can also be very pleasant data structures.
The kinds of scripts I write would be awkward to have as compiled JVM programs, I think. Shell is just way more ergonomic for me for many tasks.
Data can be meaningfully separated from control and structure in many cases, and failure to do that is a major (perhaps the major) source of security bugs.
You could then also criticize for example HTTP or even TCP for making you turn everything into "bytes".
Shell doesn't enforce any particular interpretation of data. Pipelines simply connect one program's output to another's input. Interpretation is up to the programs.
>You could then also criticize for example HTTP or even TCP for making you turn everything into "bytes".
If these were the only standard protocols that existed, and people were trying to tell me this was great because it's easy to compose different network applications, that criticism would be completely valid.
>Interpretation is up to the programs.
But because there are no standards beyond "stream of bytes", the chance that two independently written programs working with non-stream-like data can communicate directly is extremely low.
Lots of programs can communicate with JSON, XML, standard formats like that. If some legacy program outputs a non-standardized kind of output, that's a problem to be solved, not an inherent failure of shell scripting. There is no overarching successful solution to the problem of different programs using different data representations, but I don't blame this on shell; I work happily with shell scripts as do many many others. The same problem shows up the minute you want to use a Ruby module from Python, and rewriting everything in every language is not an economically viable solution.
> Lots of programs can communicate with JSON, XML, standard formats like that. If some legacy program outputs a non-standardized kind of output, that's a problem to be solved, not an inherent failure of shell scripting.
But the shell language itself is one of these legacy non-standardized formats. Arcane escaping rules, multiple incompatible implementations, surprising ways things get interpreted as code (e.g. the recent bash CGI bug),...
I worry when I can't tell if a comment like this is based on fact or just trying to be funny.
Because I've seen my share of nasty "legacy" automation but, surprisingly, I still think a good set of well thought-out shell scripts written by someone that understands what's being automated still beat modern tools, even when the person doing the automation is the same.
I don't quite know why this is, but there's something timeless about shell scripts. I've also seen shell script automation survive for a long time unattended and with zero issues. Not so with some of the modern tools that are supposed to be all unicorns and rainbows.
My take is that the shell script does not have any unspoken assumption or magic that is performed by the tool.
It all has to be in the script building up strictly from well-understood and long stable basic bricks (and the few places where you don't it's even worse with devops tools.)
Any issue, any question can be answered by reading the damn shell script and you're never dependent on a cookbook/recipe/playbook/component that you got off of some github repo that you need 5% of to do X.
An un-researched opinion from someone who replaced chef/ansible/puppet with his own shell-script config-management system:
I don't have to rewrite my shell-scripts every 6 months when a new version comes out. New updates usually only happen when security issues arise.
Shell-scripts tend to be simple. There's not a lot of magic hand-holding going on, which means not a lot of complexity to break things.
It keeps you from getting too abstract. Your writing pretty close and specific to what you want it to do, not "how it should be".
They are typically standalone. It's really easy to have 1 script that solves one problem, and another script that solves another. You don't need a giant code-infrastructure to keep things going.
I think config-mgmt tools can be extremely useful if your running a widely-ranged environment. But, you probably shouldn't be running a widely-ranged environment. If you keep things simple, and run as homogeneous as possible, you probably don't need all the added complexity.
> I think config-mgmt tools can be extremely useful if your running a widely-ranged environment. But, you probably shouldn't be running a widely-ranged environment. If you keep things simple, and run as homogeneous as possible, you probably don't need all the added complexity.
This applies to small environments. If the environment is large the situation almost reverses.
Deploying automation throughout a large homogenous environment is where config-management tools really shine. They make it easy to ensure homogeneity is maintained (even if that just means ensuring all machines have the same set of shell scripts) and allow grouping for staggered updates.
If the environment is widely-ranged and large, the utopia starts to break down. Their configuration explodes in complexity and (if you're not careful) you end up with mostly the same amount of work as if they were managed as small independent environments. With the added risk that there is now a single place from where you can break everything at once.
And this happens... Usually from wrong assumptions of what's common between all machines in the environment. In homogeneous environments almost everything is common, but in widely-ranged environments you sometimes add some configuration that wasn't there before and you think applies to the whole set and all hell breaks loose. If you're lucky this will happen suddenly, if you're not, breakage will spread slowly and you'll spend quite a lot of time scratching your head on why.
I don't think large/small is a good deciding factor. You can be large and homogeneous, or small and diverse. I think similar/dissimilar is a better decider for config-mgmt vs shell-scripts.
I'd argue that config-mgmt usually does a better job if your setup is large and complex. No need to write a script that checks if it needs to install a .deb, .rpm, or whatever, if your config-mgmt tools have already done that work.
Also, if you build your shell-scripts right, they can ensure that your system is kept the same.
Kubernetes/Docker will become increasingly accessible to developers and it will loosen the reliance on lock-in PaaS like Heroku - This is the future; I'm betting everything on it.
With tools like Rancher http://rancher.com, you can already see things moving in that direction. Next step is rancher-as-a-service.
When it comes to developers, I think open systems will always prevail in the end (it's just more flexible).
You still have to write code, define your database schema and declare config files to specify how containers should be orchestrated but soon open source developers will start creating frameworks/boilerplates which capture some of these requirements and which can automatically run and scale on Kubernetes/Swarm/Mesos and this will greatly speed up application development and deployment.
Right now, we think of frameworks as being components (part of) larger software systems - But in the future, frameworks will provide the foundations for entire software systems - They will be be responsible for declaring their own network topologies and resource requirements and they will be capable of scaling automatically to any number of machines.
Developers will extend and customize the framework with their own logic but the framework itself will handle all the difficult stuff related to its own operations.
I can always count on at least one comment per tech article with exactly this text.
Having read the article back then (and reread it now) it seems like it's still relevant. Maybe we'll have to add the year qualifier after a while when AWS lambda becomes "the way".
Agreed, was reading through this and before I checked the date - all I could think was "I know I've read this before - did someone just republish this as their own?"
1) Small teams (~1-5 people) trying to seem "big" by working at Google's scale.
2) Heroku's prices. We are currently (successfully so far) migrating a small Django project from bare Amazon EC2 instances to ECS with Docker. Even using 3 EC2 micro instances (1 vCPU, 1 GB RAM) for the Docker cluster we would spend ~8 USD/month/instance. With Heroku the minimum would be 25 USD/month/dyno. That's a 3x increase in expenses.
It's very possible to take advantage of technologies like containers without getting too caught in the hype.
Indeed, the goal is to solve your business problem with technology, not use Docker for everything that you can find in your infra. Many people are mixing up the two. Docker can be replaced with anything that is hyped at this level.
Exactly. Docker provides a set of features that are nice to standardized development environments and deployments across projects. Anything else that accomplishes that works as well.
wait. you're comparing $25 with $75. it is 3x but it's still accounting noise by any standard imaginable unless you're running a charity server for an open source project.
What about the standard of "I'm young and this is a side project I'm doing in a couple of hours at the weekends"? Of course once you have a real company with more than two customers $75 is nothing. But version 0.1 is often a tool that's only useful to you.
Heroku has free dynos for side projects, and hobby dynos ($7/dyno/month) for slightly-less-side projects. So that original $75/m quote isn't quite right for that situation.
> What about the standard of "I'm young and this is a side project I'm doing in a couple of hours at the weekends"?
Well, even someone who's young (for values of 'young' older than high-school age) is probably spending more than that every month on beer, food & entertainment each …
A few hours a week dedicated to re-building our deployment process (which was a pain since everything had to be provisioned manually for each new project). Not saying it was the best approach, it sure was an improvement and worth the (relatively little) time.
Once again keep in mind that for new projects the process is so streamlined it will take a fraction of the time to set them up.
To answer the concerns raised in the comments: we are a real company and it took 2 weeks (while working on other features and bugfixes) to migrate to Docker. The plus is that now we have experience with the platform and we can streamline the process. Again: we are not using microservices or anything like that, simply Docker containers instead of EC2 instances, which makes life pretty damn easier (and cheaper).
And 25 and 75 are bogus numbers, what if we start running 10 instances?
What a nice clickbait title. A is better than B, even though A is an apple while B is an orange, and we use them for entirely different purposes. Some functionality provided by Heroku can be replaced with Docker, and some missing features of Heroku are in the Docker infra, that can be added to Heroku using software (like service discovery: https://blog.heroku.com/managing_your_microservices_on_herok...)
535 comments
[ 2.2 ms ] story [ 347 ms ] threadAs usual, if a joke needs to be explained (to a person) then it's not funny (to them). I found this amusing since it aligned with my experiences:
> -Since no-one understands Paxos, this guy Diego…
> Oh, you know him?
> -No, he works at CoreOS. Anyway, Diego built Raft for his PhD thesis cause Paxos was too hard. Wicked smart dude. And then he wrote etcd as an implementation, and Aphyr said it wasn’t shit.
> What’s Aphyr?
> -Aphyr is that guy who wrote, ‘Call Me Maybe.’ You know, the distributed systems and BDSM guy?
> What? Did you say BDSM?
There are several jokes here based on cultural references related to the aforementioned topics.
* First when you tell it,
* then when you explain it,
* then finally when they get it.
(x used to be English people when I was a kid.)
Even renders on console browsers.
http://discuss.joelonsoftware.com/?joel.3.219431.12
(Factory factory factory factory.)
The idea is you can deploy any app to any infrastructure of your choice (inside Docker containers). This means that you are not locked into Heroku and it gives you much more flexibility.
It's basically a hosted Rancher http://rancher.com/ service with a focus on a specific stack.
I think in the future, there will be a lot of services like Baasil.io (specializing in various stacks/frameworks) and managed by various open source communities.
Docker and Kubernetes WILL become more accessible to developers - I would bet my life on it.
I'm currently building a CLI tool to allow deploying in a single command - So you can get the simplicity of Heroku while not losing any flexibility/control over your architecture.
Also Baasil.io is essentially just a control panel/dashboard (Rancher-as-a-service), you can quit Baasil.io at any time and switch to your own hosted Rancher instance and you don't have to change any of your application code or change infrastructure providers.
The main benefit of Baasil.io is that it was built by an open source community using open source software and so we can offer the best possible support for apps built on top of our own open source stacks.
"Also Baasil.io is essentially just a control panel/dashboard (Rancher-as-a-service), you can quit Baasil.io at any time and switch to your own hosted Rancher instance and you don't have to change any of your application code or change infrastructure providers."
But so you can with AWS and Azure - especially with their Docker offerings. So I'm not sure what problem baasil.io actually solves? If anything it just adds to the list of dependencies and points of failure.
Users don't have to use Baasil.io, but if they do, they will get the best possible support - For example, a customer can give us access to their Rancher control panel and this would allow us to SSH into their machines to help resolve any problems in a hands-on way.
It's probably more accurate to describe it as "DevOps as a service - With a focus on realtime apps/services". The value proposition is probably closest to Cloud 66 http://www.cloud66.com/ except more focused on realtime apps.
Another similar service is Zeit.co https://zeit.co/ - Except Zeit.co only runs Node.js. Baasil.io can be extended with components written in any language.
Since they upped the cost of their small tier, I moved to Digital Ocean and installed Dokku, which gives me that Heroku-like deployment experience so managing my (admittedly very small) website isn't that much of a hassle.
And you automatically get things like auto-scaling, database auto-provisioning. easy debugging and more.
Disclaimer: I'm Boxfuse's founder and CEO
I find them much better than walls of text.
is there ANY way i can spin up a server, add the ssh keys to some configuration file somewhere and just "docker-magic push" and have my rails application running ?
or do "docker-magic bundle exec db:migrate" and have that command run on the server.
Or push a Procfile with worker definitions and have the PAAS automatically pick it up, add it to supervisord/systemd and run it ?
In a very specific case, Heroku is the best solution for my problem. Sounds like it is for you too.
http://deis.io/
http://nickjanetakis.com/courses/scaling-docker-on-aws
It covers using RDS, ElastiCache and also handles load balancing your app + much more.
Yep, it costs $20. Basically the cost of chinese food for 2, to ensure guaranteed victory in learning the essentials of AWS' platform while having a guided tour on how to deploy a fault tolerant web app with Amazon ECS from start to finish.
You can definitely learn everything for free, but the value in a course is that you're getting a cohesive learning path that was carefully planned and tested. You get a system that you can apply to your own projects and plenty of source code to reference.
You're paying the $20 so you can avoid spending 6 months trying to figure out everything on your own while stringing together a bunch of half-assed blog posts and tutorials.
You pay the small fee for certainty and it's well worth it because your time (and sanity) is not infinite.
I see that you recently left Amazon after making 200k/year there (a comment you made 5 days ago). I can see why you don't like people promoting Amazon products, I fully understand.
There is, however, still a hump to get over in installation -- you need to learn what BOSH is, install BOSH, then install Cloud Foundry with BOSH. In the long run, for a production deployment, this is what you want. But it certainly doesn't feel that way when you just want to kick some damn tires.
If you just want to tinker, you can try PCFDev[0]. It's a fully-functional Cloud Foundry installation in a single VM.
Disclosure: I work for Pivotal, we donate the majority of engineering on Cloud Foundry.
[0] https://pivotal.io/platform/pcf-tutorials/getting-started-wi...
Edit: yes, I know we ask you to signup during the PCFDev install. I hate it too. We have to for export compliance, it can't be avoided.
CF is already a relatively successful business with hundreds of million dollars in annual revenues across a pile of companies. Kubernetes and Docker are small in comparison as "businesses" but of course the momentum there is surging in terms of both pure open source adoption and contribution. It's likely going to be a big market with a lot of choice like Deis, or plain k8s, or RedHat OpenShift, or IBM bluemix, Pivotal , Docker Datacenter, or Mesos/Marathon, etc. It's a bit of a market war brewing and that competition will make for better solutions.
What is great about open source vs. past tech "gold rushes" is that these experiments and feedback loops exist across communities that are otherwise competing and overlapping. Mesos adopted the Docker image format independently of the Docker runtime ; Kubernetes introduced pods independently but also reused parts of Docker. Docker container networking and volumes are being used compatibly in the latest incubated CF releases. RedHat submitted a way to get CF style buildpacks working on K8S. Someone found a way to make CF run on Mesos; I could see a similar attempt on Kubernetes some day. It's a confusing and busy time but also an explosion of activity. And even if there is competition for dollars in the end among all these players that will lead to tension , the work is out in the open mainly.
Most of Cloud Foundry is built the way I like software to be built. Pair programming, TDD, small balanced teams, prioritising for user valu.
That style of development is actually baked into the Cloud Foundry Foundation rules. Companies who join the Foundation are expected to send engineers to ramp up on developing in this style. And voting rights are based on the number of full-time engineers you have assigned to the effort.
The reason I mention all this is that I trust the way we build Cloud Foundry. We still get production bugs and oversights and mistakes. It's around 4 million lines of code that turns into a distributed system of ~50 different interacting processes. We built a fully-featured, robust PaaS, using containers, in about 3 years, starting from scratch.
Nobody outside Google had built a container platform of this level before. Nobody but Heroku had built a fully-featured PaaS of this level before. We are, to my knowledge, the first system to do both of these things. Certainly the first opensource one.
The reason you never hear about Cloud Foundry is because we've already built all the components other folks are trying to roll up into full PaaSes. "It just works, already" is a boring story.
But again, quite seriously: I am obviously very biased.
Please excuse me while I take an hour to digest that.
Right now the companies who provide the most engineering -- Pivotal and IBM -- are laser-focused on capturing enterprise dollars.
Which from a business perspective makes perfect sense. Pivotal's commercial distribution of Cloud Foundry (PivotalCF) holds the record for fastest-growing sales of any opensource product. Ever.
But there's not much effort on promoting to devs and startups (IBM are starting to do this more with BlueMix, which is their CF distribution). But it's early days.
So basically you'll usually find that I and a sprinkling of my colleagues show up in threads like these out of the goodness of our hearts and fondness for our work.
Obviously I have a financial interest. Pivotal makes money from PivotalCF, I work for Pivotal I'm on an options plan as well. So YMMV.
But I think Cloud Foundry is just ... way ahead, in terms of actually getting work done.
Edit: and since I'm musing aloud about business-y things, I should emphasise again that nothing I say is in an official capacity, consult your lawyer, financial planner and astrologer, etc etc.
As such, I maintain SOAP should be gone for the good of the running system.
It's literally billed as "A fast and modern Python SOAP client". Python 2 and 3 compatible. Last commit was two weeks ago.
http://docs.python-zeep.org/en/master/
And going by the bugtracker, it's running into quite a few problems with almost-but-not-quite compliant servers/WSDL files, which is a real issue when you're trying to interface ass-old legacy APIs (we're talking "not upgraded since 2006"-old) made by $BigEnterprise. Maybe this time the project won't die before they work out all the little kinks.
Nooooooooooooooooo. Everytime someone says "service discovery" a kitten dies (Except for consul, that's the biz).
Read it, it's a lovely 5 minutes piece of writing.
At least you understand the author's intention. I would be worried if some non-technical people took the title literally...
Hitler uses Docker: https://www.youtube.com/watch?v=PivpCKEiQOQ
By the way, why all the downvotes to the parent?
People seem to underestimate just how powerful modern machines really are. And I don't get why people seem to think it's hard to deploy simple web applications. Just write a 4-line shell script that rsync's, runs whatever DB migrations you may have and restarts the thing.
I have 200k visitors per month generating 8m page views and about 50m hits on the servers (with CDNs taking another few hundred million hits).
These all peak during the UK weekdays and wind down at nights and weekends.
Divisions over time aren't going to work, but neither is translating visitors into requests, especially as it's only the page views that have a beyond trivial computation cost.
> [...] before switching to Docker (even though I do love Docker).
This seems like a very drastic solution to problems he does not yet have. I've been responsible for similar thoughts before, but shortly after realized just how damn stupid I was.
I definitely understand that modern machines are powerful, but I used to STIG RHEL 4 and RHEL 5 boxes for 4 years as my primary day job. I've done everything from create creating kickstart files to manually locking down a whole Linux instances to creating RPM files. At this point in my career I just don't care about the extra cycles I get by using <insert your infrastructure tech>.
If my product needed the extra performance, trust me I would switch.
[1] - https://github.com/joeblau/gitignore.io/wiki/Advanced-Comman...
exactly. I mean look, if you have a lifestyle business that's only going to support 5-10 people, it's totally a waste of time. if you have some hope of scaling this is the way to go. I get it, just use Heroku. It's easy and convenient. If you're planning on a billion dollar exit, this way is way better.
> I need to decide if i believe my own hype?
yeah. sorry.
That seems excessive"
A 100 times yes. We tried to split our monolithic Rails app into micro-services built in Go. 2 years and many fires later, we decided to abandon the project. It was mostly because the monitoring and alerting were now split into many different pieces. Also, the team spent too much time debating standards etc. I think micro-services can be valuable, but we definitely didn't do it right, and I think a lot of companies get it wrong. Any positive experiences with micro-services here?
A small team starting a new project should not waste a single second considering microservices unless there's something that is so completely obviously decoupled in a way that not splitting it into a microservice will lead to extra work. It's also way easier to split into microservices after the fact than when you're developing a new app and you don't have a clue how it will look like or what the overall structure of the app will be in a year (most common case for startups).
Multiple platforms is not a problem and generally a good thing as long as it's not excessive. You don't want to be in a case where you have the same number of different platforms as developers or anything like that. I'm guessing there is a rule of thumb here, but I'm not sure what it would be. Max 1 different platform per 5 developers? Something like that.
I do wish people would stop conflating "running in a different service" and "loose coupling". They are completely orthogonal.
I've worked on some horrendously tightly coupled microservices.
Unless you can coax dOSGi into working (which is tons of fun), then you can have services tightly coupled to other services running on entirely different machines causing frequent (and hilarious) cascades of bundle failures whenever the network hiccups.
OSGi is a trigger word for me now. I've worked on two large OSGi projects (previous job and current job) and it's always the same. Sh*t is always broken (and my lead still insists that OSGi is the one true way to modular bliss). And the OSGi fanboys always say "Your team is using it wrong!" Which very well might be true, but I no longer care. Apparently it's just too damn hard to get a team of code monkeys to respect service boundaries when OSGi makes it so damn easy to ignore them.
If I'm ever in a position of getting to design a new software architecture (hasn't happened in 10 years, but hey I can dream), I'll punch anyone who suggests "OSGi" to me right in the face.
With microservices, the production version of their service would conceivably be stable. It moves the contract from the repo to the state of production services.
With a monolithic repo done right, the other teams broke their build of their branch, and it's up to them to resolve it. You, meanwhile, are perfectly happy working on your branch. When their changes are mergeable into trunk, then they may merge them, not before — and likewise for you.
With multiple repos, they break your build, but don't know it. You don't know it either, until you update your copies of their repos — and now you have to figure out what they did, and why, and how to update your logic to handle their new control flow, and then you update again and get to do it again, until finally you ragequit and go live in a log cabin with neither electricity nor running water.
I don't see how this is a problem if you are pushing frequently and have a CI system. You know within minutes if the build is broken. If it broke, don't pull the project with the breaking changes.
My point is, I don't think one approach is inherently better than the other. Both require effort on the part of the teams to manage changes (or a CM team), and both require defined processes.
I agree with the overall sentiment of your comment, but the quoted part is where I've seen trouble brew. The tendency is to be conservative about pulling updates to dependencies, which can easily get you into a very awkward state when a critical update eventually sits on top of a bunch of updates you didn't take because they broke you. It is usually better to be forced to handle the breakage immediately, one way or another.
Yes, that's the contract that you need to have with other teams. And it's the contract that is automatically enforced with microservices.
As for advantages, microservices tend to keep code relatively simple and free from complex inheritance schemes. There's rarely a massive tangled-up engine full of special cases in the mix, as there often is in monolithic apps. This substantially decreases technical debt and learning curve, and can make it simple to understand the function an isolated microservice performs.
There is the obvious advantage that if you have disparate applications executing nearly-identical logic to read or write data to the same location, and the application platforms can't execute the same library code, you can centralize that logic into an HTTP API, which reduces maintenance burden and prevents potentially major bugs.
My opinion is that adopting microservices as a paradigm leads to a slow, difficult-to-debug application, primarily because people take the "micro" in microservices too seriously. One shouldn't be afraid to split functionality out into an ordinary service after it's been shown to be reasonable to do so.
If you expect to need to step into a function call when debugging, then it's too tightly coupled to spin out. You should be able to look at the arguments to the call and the response and determine if it's correct (and if not, now you have isolated a test case to take to the other service and continue debugging there).
If the interface will change so often that you expect it will be a problem that it's in a separate repository, if you expect that you will always need to deploy in tandem, then it's too tightly coupled to spin out.
The advantage of micro services is the separation in fact of things that are separate in logic. The complexity of systems grows super-linearly, so it's easier to reason about and test several smaller systems with clear (narrow) interfaces between them than one big. It's easier to isolate faults. It's harder to accidentally introduce bugs in a different part of the system when the system doesn't have a different part. If done right, scaling can be made easier. But these are hard architectural questions, there's no clear-cut rule for when you should spin off a new service and when you should keep things together.
Someone else mentioned separating the shopping app from the payment system for an ecommerce business, which even has security benefits. I think that's an excellent example.
Edit: Another clear benefit is that you can choose different languages, libraries, frameworks and paradigms for different parts of the code. You can write your boring CRUD backend admin app in Ruby on Rails, your high-performance calculation engine in Rust and your user-facing app in Node.js (so the front- and backend an share Javascript validation code).
A thousand times yes. Distributed systems are hard.
> Debugging is more difficult since you now can no longer step through your program in a debugger but rather have an opaque network request that you can't step into.
Yes. Folks underestimate how difficult this can be.
In theory it should be possible to have tooling to fix this, but I've not seen it in practice.
> You can no longer use editor/IDE features like go to definition.
Not a problem with a good editor.
> Version control becomes harder if the different services are in different repositories.
No organisation should have more than one regular-use repo (special-use repos, of course, are special). Multiple repos are a smell.
Mind elaborating on this?
I would modify this slightly. Larger organizations with independent teams may want to run on per-team repos. Conway's law is an observation about code structure but it sometimes also makes good practice for code organization. And of course, sometimes the smell is "this company is organized pathologically".
Another problem is that large monolithic repositories can be difficult to manage with currently available software. Git is no panacea and Perforce isn't either.
Flat out wrong for any organization with multiple products. Which, let's be honest, is most of them.
My personal take on it, at this point, is that much of our knowledge of how to manage projects (things like individual project repos, semantic versioning, et cetera) is centered on the open-source world of a million mostly-independent programmers. Things change when you work in larger organizations with multiple projects. You even start to revisit basic ideas like semantic versioning in favor of other techniques like using CI across your entire codebase.
Monorepos come with their own challenges. For example, if any of your code is open source (which means it must be hosted separately, e.g. on Github), you have to sync the open-source version with your private monorepo version.
Monorepo are large. Having to pull and rebase against unrelated changes on every sync puts an onerous burden on devs. When you're remote and on the road, bandwidth can block your ability to even pull.
And if you're going to do it like Google, you'll vendor everything -- absolutely everything (Go packages, Java libraries, NPM modules, C++ libraries) -- which requires a whole tool chain to be built to handle syncing with upstream, as well as a rigid workflow to prevent your private, vendored fork from drifting away from upstream.
There are benefits to both approaches. There is no "one right way".
I love Git, and I used submodules for years in personal projects. It started with a few support libraries shared between projects, or common scripts for deployment, but it quickly ballooned into a mess. I'm in the process of moving related personal projects to a monolithic repository, and in the process I'm giving up the ability to tag versions of individual projects or provide simple GitHub links to share my code.
Based on these experiences, I honestly think that the only major problem with monolithic repositories is that the software isn't good at handling it, and this problem could be solved with better software. If the problem is solved at some point in the future, I don't think the answer will look much like any of the existing VCSs.
Based on experiences in industry, my observation is that the choice of monolithic repository versus separate repository is highly specific to the organization.
Totally agree with everything else, but gotta completely disagree on this last point. Monorepos are a huge smell. If there's multiple parts of a repo that are deployed independently, they should be isolated from each other.
Why? Because you're fighting human nature, otherwise. It's totally reasonable to think that once you excise some code from a repo that it's no longer there, but when you have multiple projects all in one repo, different services will be on different versions of that repo, and your change may have changed semantics enough that interaction bugs across systems may occur.
You may think that you caught all of the services using the code you refactored in that shared library, but perhaps an intermediate dependency switched from using that shared library to not using it, and the service using that intermediate library hasn't been upgraded, yet?
When separately-deployable components are in separate repositories, and libraries are actual versioned libraries in separate repositories these relationships are explicit instead of implicit. Explicit can be `grep`ed, implicit cannot, so with the multi-repo approach you can write tools to verify that all services currently in production are no longer using an older, insecure shared library, or find out exactly which services are talking to which services by the IDLs they list as dependencies.
While with the monorepo approach you can get "fun" things like service A inspecting the source code of service B to determine if cache should be rebuilt (because who would forget to deploy service A and service B at the same time, anyways...), as an example I have personally experienced.
My personal belief is that the monorepo approach was a solution back when DVCSs were all terrible and most people were still on centralized VCSs like Subversion that couldn't deal with branches and cross-repo dependencies well, and that's just what you had to do, while Git and Mercurial, along with the nice language-level package managers, make this a non-issue.
Finally, there's an institutional bias to not rock the boat (which I totally agree with) and change things that are already working fine, along with a "nobody got fired buying IBM" kind of thing with Google and Facebook being two prominent companies using monorepos (which they can get away with by having over a thousand engineers each to manage the infrastructure and build/rebuild their own VCSs to deal with the problems inherent to monorepos that most companies don't have the resources and/or skills to replicate).
EDIT: Oh, I forgot, I'm not advocating a service-oriented architecture as the only way to do things, I'm just advocating that whatever your architecture, you should isolate the deployables from each other and make all dependencies between them explicit, so you can more easily write tooling to automatically catch bad deploy states, and more easily train new hires on what talks to/uses what, since it's explicitly (and required to be) documented.
If that still means a monorepo for your company's single service and a couple of tiny repos for small libraries you open source, that's fine. If it means 1000 repos for each microservice you deploy multiple times a day, that's also fine (good luck!).
Most likely it means something like 3-10 repos for most companies, which seems like the right range for Miller's Law) ( https://en.wikipedia.org/wiki/The_Magical_Number_Seven,_Plus... ) and therefore good for organizing code for human consumption.
But having multiple repos doesn't prevent the equivalent situation from happening (and, I think, actually makes it much likelier): no matter what, you have to have the right processes in place to catch that sort of issue.
> You may think that you caught all of the services using the code you refactored in that shared library, but perhaps an intermediate dependency switched from using that shared library to not using it, and the service using that intermediate library hasn't been upgraded, yet?
That's the sort of problem which happens with multiple repos, but not (as often) with a single repo.
> Explicit can be `grep`ed, implicit cannot, so with the multi-repo approach you can write tools to verify that all services currently in production are no longer using an older, insecure shared library, or find out exactly which services are talking to which services by the IDLs they list as dependencies.
A monorepo is explicit, too, even more explicit than multiple repos: WYSIWYG. And you can always see if your services are using the same API by compiling them (with a statically-typed language, anyway).
The beautiful thing about a monorepo is it forces one to confront incompatibilities when they happen, not at some unknown point down the road, when no-one know what changed and why.
What editor are you thinking of that can jump from HTTP client API calls to the corresponding handler on the server?
You don't debug distributed systems by tracing into remote calls and jumping into remote code. You debug it by comparing requests and responses (you use discrete operations, right) with the specified requests and responses, and then opening the code that has a problem¹.
It calls for completely different tooling, not for a "better debugger".
1 - Or the specs, because yes, now that your system is distributed you also have to debug the specs. Why somebody would decide on doing that for no reason at all? Yet lots of people do.
Microservices also make it much harder to refactor the code which you often need to do in the early stage of a project.
That's a good point. I think this thought extrapolates to other parts of software engineering as well. Sometimes writing very modular and decoupled software from the beginning is very hard for a small team, and we can't see well if this is the best approach since it's also hard to grasp the big picture.
I'm currently facing this issue. I'm trying to write very modular and reusable applications, but now I'm paralyzed trying to picture the best patterns to use, where should I use a facade, a decorator, etc. I think I'll adopt this strategy for myself--only focus on modularizing from the beginning if it'd lead to extra work otherwise.
I don't really like that model applied to everything, but eh now you are kind of forced in a hybrid approach - say, your macro vertical plus whatever payment gateway service, intercom or equivalent customer interaction services, metrics services, retargeting services, there are a lot of heterogeneous pieces going into your average startup.
but back on topic, what Docker really needs now is a whack on the head of whoever thought swarms/overlays and a proper, sane way to handle discovery and fail-over - instead we got a key-value service deployment to handle, which cannot be in docker and highly available unless you like infinite recursion.
It makes sense for some thing. We run a webshop, but have a separate service that handles everything regarding payments. It has worked out really well, because it allows us to fiddle around with pretty much everything else and not worry about breaking the payment part.
It helps that it's system where we can have just one test deployment and everyone just uses that during testing of other systems.
I've also work at a company where we had to run 12 different systems in their own VMs to have a full development environment. That sucked beyond belief.
The idea of micro-service are is enticing, but if you need to spin up and configure more than a couple to do your work, it starts hurting productivity.
Is the payments service a single service that manages the whole transaction, or have you go for multiple services handling each part and, if so, how did you manage failure with a distributed transaction?
We had almost the same story with payments. Except for we've jumped to a payment-processing SaaS but got dissatisfied (all those SaaSes I saw don't work with PayPal EC without so-called "reference transactions" enabled) and decided that wasn't a good idea and we have to jump back to in-house implementation.
I didn't want to re-integrate the payments code back to the monolith - thought it would take me more time and make code messier. So I wrote a service (it's small but to heck with "micro" prefix) that resembled that SaaS' API (the parts we've used). It had surely evolved and isn't compatible anymore, but it doesn't matter as we're not going back anyway.
Works nicely and now I'm feel more relaxed - touching the monolith won't break payments.
On the other hand, I see how too many services may easily lead to fatigue. Automated management tooling (stuff like docker-compose) may remedy this, but also may bring their own headaches.
We have specific services that process different types of documents, or communicate and package data from different third parties, or process certain types of business rules, that multiple apps hook into, but it's literally like 20 services total for our department, some that are used in some apps and not others.
When I hear 'micro-services' I'm picturing something more akin to like node modules, where everything is broken up to the point where they do only one tiny thing and that's it. Like your payment service would be broken into 20 or 30 services.
But maybe I'm mistaken in my terms. I haven't done too much with containers professionally, so I'm not too hip with "the future".
Yep. We already had a feature flag system, a minimal monitoring system, and a robust alerting system in place. Microservices make our deployments much more granular. No longer do we have to roll back perfectly good changes because of bugs in unrelated parts of the codebase. Before, we had to have involved conversations about deployments, and there were many things we just didn't do because the change was too big.
We can now incrementally upgrade library versions, upgrade language versions, and even change languages now, which is a huge win from the cleaning up technical debt perspective.
To be honest, we still have a monolithic application at the heart of our system that we've been slow to decompose, though we're working on it. We deploy it on a regular cadence and use feature flags heavily to make it play nice with everything else.
Git doesn't really help with that. More granular deployments do, and if microservices help with more granular deployments, go for it.
That's your problem right here
The thing is though, the Elixir feed checker has its own database table that tracks whether it's seen an episode in a feed. And when there's a new episode it sends an API call to WP to insert the new post. The problem is that sometimes the API calls fail! Now what? I'll need to build logging, re-try etc. So I'm thinking of making the feed checker 'stateless' and only using WP with a lot of query caching as the holder of 'state' information about whether an episode has been seen before.
To sum up my experience so far, there's something nice about being able to use the right tech for each task, and separating resources for each service, but the complexity--keeping track of whether a task completed properly--definitely increases.
One hard tech limit is that with 50k podcasts, 4million+ episodes, search definitely doesn't work well. Not just WP, but SQL itself. Hence Elasticsearch. I also plan to work on recommendations, etc. so will need probably to be exporting SQL data into other systems anyway for making the "people who liked this also liked this" kinda things.
Also I kinda lied about using the WP API--that's how I built the system initially (and will switch to it moving forward), but to import the first few million posts from the content of the feeds, I just used wp_insert_post against the DB of new entries that Elixir fetched (I posted the code I used here: http://wordpress.stackexchange.com/a/233786/30906).
I also plan to write the whole front-end in React (including server side rendering) so will have to figure out how to get that done. Would probably use the WP-API with a Node.js app in front of it, will look into hypernova from AirBNB. So probably more usage of WP API accessed by another service...
It doesn't sound like microservices are needed, just adding in the appropriate tech for the job.
Once these are doing anything other than rotating log files, can the system really be considered monolithic?
The more dramatic effect was on a particular set of endpoints that have a relative high traffic (it peaks at 1000 req/s) that was killing the app, making upset our relational database (with frequent deadlocks) and driving our Elasticsearch cluster crazy.
We did more than just split the endpoints into microservices. We also designed the new system to be more resilient. We changed our persistence strategy to make it more sensible to our traffic using a distributed key-value database and designed documents accordingly.
The result was very dramatic, like entering into a loud club and suddenly everything goes silent. No more outages, very consistent response times, the instances scaled with traffic increase very smoothly and in overall a more robust system.
The moral of this experience (at least for me) is that breaking a monolith app into pieces has to have a purpose and implies more than just move the code to several services keeping the same strategy (that's actually slower, time consuming and harder to monitor)
One key factor was decoupling the high volume updates from the users requests so one didn't affect the other one.
I can't get my head around how people introduce changes to their system if they have to update 12 different microservices at once? It must be horrible.
Often you hear stories how people are converting monolithic app to microservices - but this is easy. Rewriting code is easy and it's fair to say it always yields better code (with or without splitting into microservices - it doesn't matter).
What I'd like to hear is something about companies doing active development in microservice world. How do they handle things like schema changes in postgres where 7 microservices are backed by the same db? What are the benefits compared to monolithic app in those cases?
It seems to me that microservices can easily violate DRY because they "materialise" communication interfaces and changes need to be propagated at every api "barrier", no?
As I said in another thread, the separation in different components was key for resiliency. That allowed independence between the higher volume update and the business critical user facing component.
>I can't get my head around how people introduce changes to their system if they have to update 12 different microservices at once? It must be horrible.
The thing is, if you design the microservices properly it is very rare to introduce a change in so many deployments at once. Most of the time is just 1 or 2 services at a time.
>What I'd like to hear is something about companies doing active development in microservice world. How do they handle things like schema changes in postgres where 7 microservices are backed by the same db? What are the benefits compared to monolithic app in those cases?
We don't introduce new features in our monolith service anymore. So, from that perspective we do all active development in microservices.
>"How do they handle things like schema changes in postgres where 7 microservices are backed by the same db?
The trick is, you want to avoid sharing relational data between microservices. I don't know if it is just us, but we have been able to split our data model so far and in most cases we don't even need a relational database anymore, so having a schemaless key/value store makes seems easy too.
> What are the benefits compared to monolithic app in those cases?"
There are several advantages, but the critical one for me is being able to have a resilient platform that can still operates even if a subsystem is down. With our monolithic app is an all or nothing thing. Another advantage is splitting the risk of new releases.
>It seems to me that microservices can easily violate DRY because they "materialise" communication interfaces and changes need to be propagated at every api "barrier", no?
Not necessarily. YMMV but you can have separation of concerns and avoid sharing data models. When you do have shared dependencies (like logging strategy or data connections) you can always have modules/libraries.
He said that about micro-kernels but it may apply to micro-services as well.
Breaking things apart too much will only cause you grief.
Although I personally had to deal with some monolithic monsters that I wished were split into smaller services.
The advantage though is that APIs (system boundaries) are usually better defined.
Perhaps one should use the best of both worlds, and run microservices on a common database, and somehow allow to pass transactions between services (so multiple services can act within the same transaction).
A shared database is an anti-pattern in distributed systems.
Similarly, distributed transactions (ala. DTC) is an anti-pattern.
Distributed systems aren't hard. They're just different.
Then again, sometimes it's advantageous to identify parts of your system where aspects of state can be safely decoupled. And in which having them reside in disparate systems (and yes, sometimes be inconsistent or differently available) might actually be a better overall fit.
You completely lose the concept of transactional integrity, so you will have to work around that from the start.
Then again, sometimes your state changes not only don't need to be transactional; it can be disadvantageous to think of them that way.
Depends, depends, depends.
I'm curious; in what kinds of situation would this apply?
> Depends, depends, depends.
Flexibility is usually an important requirement. Often you cannot freeze your architecture and be done with it. I think a transactional approach could better fit with this.
Any situation where the business value of having your state be 100% consistent does not outweigh the performance or implementation cost of making it so.
In my experience, any monolith that can be broken up into a queue based system will benefit enormously. This cleans up the pipelines, and adds monitoring and scaling points (the queues). Queues removes run-time dependencies to the other services. It requires that these services are _actually_ independent, of course.
I do, however, avoid RPC based micro-services like the plague. RPC adds run-time dependencies to services. If possible, I limit RPC to other (micro) services to launch/startup/initialization/bootstrap, not run-time. In many cases, though, the RPC can be avoided entirely.
The non-web world has been doing this with message queueing for about 15 years. Maybe more.
I mean, the infamous "UNIX way" of "do one thing, do it well" (something we nearly lost with popularity of "do everything in a manner incompatible with how others do it" approach in too many modern systems), when complex behavior was frequently achieved through the modularity of smaller programs communicating through well-defined interfaces.
Heck, microkernels are all about this, and their ideas haven't grew out of nowhere. And HURD (even though it was never finished) is quarter a century old already.
a good example of this that I've used in production at my current $dayjob: dynamic PDF generation. user makes request from our website, request data is used to fill out a pdf template context which is then sent over to our PDFgen microservice which does its thing and streams a response back to the user.
All of that and much more needs to be replicated for each microservice, right?
Why not just have a module in your monolithic app that does it. The logic will still be separate. In most languages/frameworks you can spawn pdf generation task. Any changes are easier to introduce as well. There's no artificially materialised interface. Updates are naturally introduced. All auth logic is there already, you don't need to worry about deploying yet another service, same with logging etc.
the template has values that are related to database models. the main app (still mostly monolithic) fills out the template context. the context itself is what's passed to the microservice. the microservice does not connect to a database at all.
> Does it keep connection pool of let's say 5 connection always open (as libraries like to do)?
no. the service probably handles a few hundred requests per day, it is not in constant use. communication is over HTTPS. it opens a new connection on each request. this does impact throughput, but its a low throughput use case, and pdf rendering itself is much slower and that time totally dominates the overhead of opening and closing connections anyway.
> Does it have authentication?
yes, it auths with a bearer token that is borne only by our own internal server. this is backend technology so we don't have to auth an arbitrary user. we know in advance which users are authorized.
> Is it public or private API?
private
> Who is managing security?
we are, with a lot of assistance from the built-in security model of AWS.
> Is it running behind it's own nginx or other proxy?
the main app is behind nginx. the microservice is running in a docker container that exposes itself over a dedicated port. there's no proxy for the microservice, again, because of the low throughput/low load on the service. no need to have a load balancer for this so the most obvious benefit of a proxy wasn't applicable.
> Does it have DoS protection (PDF generation can be CPU intense)?
yes, it's an internal service and our entire infrastructure is deployed behind a gatekeeper server and firewall. the service is inaccessible by outside requests. the internal requests are queue'd up and processed 1 at a time.
> What about the schema for request?
request payload validation handled on both ends. the user input is validated by the main app to form a valid template context. the pdf generator validates the template context before attempting to generate one also. its possible to have a valid schema that has data that can't be handled correctly though. errors are just returned as a 500 response though. happens infrequently.
> They need to be deployed together with changes in other services, right?
nope. the microservice is fully stand alone.
> What about changes to database schema - you need to remember to update that service as well and redeploy it at the right time as well - just after successful db migrations - which live in another project.
the microservice doesn't interact with a database at all. schema changes in the main app database could potentially influence the pdf template context generation, but there are unit tests for that, so if it does happen we'll get visibility in a test failure and update the template context generation code as needed. none of this impacts the microservice itself though. it is fully stand alone. that's the point.
> All of that and much more needs to be replicated for each microservice, right?
in principle yes, and these are good guidelines for determining what is or is not suitable to be a microservice. if it would need to auth an arbitrary user, or have direct database access, or be exposed to public requests, it might not be a good candidate for a microservice. things that can stand alone and have limited functional dependencies are much better candidates.
> Why not just have a module in your monolithic app that does it.
because the monolithic app is Python/django and the PDF generation tool is Java. one of the main advantages of microservices architecture is much greater flexibility in technology selection. A previous solution used Python subprocesses to call out to PDF generation software. It's actually easier and cleaner for us to use a microservice instead.
The thing is, you need a massive investment in infrastructure to make it happen. But once you do, its great. You can create and deploy a new service in a few seconds. You can rewrite any individual service to be latest and greatest in an afternoon. Different teams don't have to agree on coding standards (so you don't argue about it).
But, the infrastructure cost is really high, a big chunk of what you save in development you pay in devops, and its harder to be "eventually consistant" (eg: an upgrade of your stack across the board can take 10x longer, because there's no big push that HAS to happen for a tiny piece to get the benefits).
Monolithic apps have their advantages too, and many forget it: less devops cost, easier to refactor (especially in statically typed languages: a right click -> rename will propagate through the entire app) and while its harder to upgrade the stack, once its done, your entire stack is up to date, not just parts of it being all over. Code reuse is significantly easier, too.
Anything you're not running locally just hits the shared infra.
Unsure if sarcastic.
Because individuals may be jumping through dozens of services a day, moving, refactoring, deploying, reverting (when something goes wrong), etc. It has to be friction-free, else you're just wasting your time.
eg: a CLI to create the initial boilerplate, a system that automatically builds a deployable on commit, and something to deploy said deployable nearly instantly (if tests passed). The services are small, so build/tests should be very quick (if you push above 1-5 minutes for an average service, it's too slow to be productive).
Anyone should be able to run your service locally by just cloning the repo and running a command standard across all services. Else having to learn something every time you need to change something will slow you down.
That infrastructure is expensive to build and have it all working together.
Monoliths invariably tend to become spaghetti over time, and completely impossible to any non trivial refactoring. With microservices, interfaces between modules are stable and spaghetti is localized.
>The thing is, you need a massive investment in infrastructure to make it happen.
I thought that one of the selling points of microservice architectures was the minimal infrastructure. I am really struggling to see an advantage in this way of doing things. You are just pushing the complexity to a dev ops layer rather than the application layer - even further form the data.
Maybe Swift? Scala Native in a year or two? I've done a little Erlang before, so maybe Elixir?
That said, in places where it doesn't make sense we didn't try to force it. Our main game API is somewhat monolithic, but behind it we have almost 10 other services. Here's a quick breakdown:
...and a few more in the works. Some of these might push the line of "micro" in that they almost all do more than a single function's worth of work, but that level of granularity isn't practical.But don't get too caught up on the "micro" part. Split services where domain lines naturally form, and don't constrain service size by arbitrary definitions. You know, right tool for the job and whatnot.
Ah yes, the 'let's have decentralised microservices with centralised standards!' anti-pattern. It results in lots of full-fledged, heavyweight, slow-to-update services, which also have all the problems of a distributed system. It's the worst of both worlds.
Well, there's your problem - you need a monitoring microservice and an alerting microservice! Well, those may be too coarse by themselves, but once you break them down into 5 or 6 microservices each, you'll be ready for production.
IMHO. You need a lead with a clear vision that drives the effort. Too many leads will create chaos.
I'm currently working on a large refactoring effort along these lines. The end goal is to create a modular, potentially distributed system that can be deployed in a variety of configurations, updated piecemeal for different customers, and integrated by our customers with the third-party or in-house code of their choice using defined APIs. We aren't typical of the other examples, though, in that we do literally ship our software to our customers and they run it on their own clusters.
I wouldn't, however, just "do microservices" from day one on a young app. But usually that young app has no idea what the true business value is, i.e., you have no idea what down time of certain parts of your services really means to the business. That's the #1 pain point we're solving: having mission critical things up 100%, and then rapidly iterating on new, less stable feature designs in separate services.
You should, however, keep an eye on how "splittable" everything is, i.e., does everything need to be in the same DB schema? Most languages have package concepts, which typically align (somehow) with "service" concepts. Do you know their dependencies? That sort of thing. Then, the later process of "refactor -> split out service" is pretty straightforward and easy to plan.
Like the way the shell would fork off an "expr" sub-process to parse a mathematical expression to add two numbers, then write the result to a pipe via stdout, then terminate the process, clean up all its resources, and switch context back to the shell, which then read the serialized sum back in from the other end of the pipe, and went about its business, regardless of the fact that the CPU running the shell already had its own built-in "add" instruction in hardware.
Unless you are over fifty years old, you never experienced this.
Rob Pike said it best: "Those days are dead and gone and the eulogy was delivered by Perl."
Perl being a thing in 1995...
[1] https://github.com/baphomet-berlin/jQuery-basic-arithmetic-p...
Shell scripts are guaranteed to be runnable on all machines.
Unfortunately the shell "language" sucks, but still...
Oh, and be very careful of the commands your script invokes!
shell scripts have no guarantee of portability (often less than Python, which has a rich standard library available on all platforms).
Why do you say that? genuinely curious
I'm curious because these simple things that are delightfully easy in bash often turn out to be surprisingly tedious in other languages.
Of course, some things are tedious in bash too. But a basic principle of shell scripting is that you call other programs to do the stuff you don't want to do in shell.
I agree it is tedious, but to be honest, reading and writing to stdin/out isn't something that would commonly need to be done in a robust system. If the world were perfect you would use library functions.
I definitely think there is scope for a language that works well as an interactive shell, and as a general purpose language. They have somewhat conflicting constraints but I'm sure we can do better than Bash. Have you seen how [ is implemented?
Yeah, a nicer shell-like language would be cool. I've been thinking about it for a while.
Bash is quirky but it gets a lot of stuff right and once you understand it it can be extremely ergonomic and productive.
And not depending on language run times other than shell can be really glorious in some situations, too...
When you're not writing shell, just use the tools the language gives you.
For the matter, I think a shell script is cleaner than a python script for devops; but I don't think the composability of unix tools is that much of an advantage compared to the amount of python libraries out there.
I'm likely to be using Python programs and other programs in those shell scripts. The beauty of shell is that it makes it so easy to compose programs written in different languages.
I think the advantages of using a single language for everything outweigh the disadvantages - see e.g. http://www.teamten.com/lawrence/writings/java-for-everything... (though actually my single language is Scala)
The kinds of scripts I write would be awkward to have as compiled JVM programs, I think. Shell is just way more ergonomic for me for many tasks.
Data can be meaningfully separated from control and structure in many cases, and failure to do that is a major (perhaps the major) source of security bugs.
Shell doesn't enforce any particular interpretation of data. Pipelines simply connect one program's output to another's input. Interpretation is up to the programs.
If these were the only standard protocols that existed, and people were trying to tell me this was great because it's easy to compose different network applications, that criticism would be completely valid.
>Interpretation is up to the programs.
But because there are no standards beyond "stream of bytes", the chance that two independently written programs working with non-stream-like data can communicate directly is extremely low.
But the shell language itself is one of these legacy non-standardized formats. Arcane escaping rules, multiple incompatible implementations, surprising ways things get interpreted as code (e.g. the recent bash CGI bug),...
https://amoffat.github.io/sh
It's pretty elegant.
import sh
sh.bar(sh.foo())
Because I've seen my share of nasty "legacy" automation but, surprisingly, I still think a good set of well thought-out shell scripts written by someone that understands what's being automated still beat modern tools, even when the person doing the automation is the same.
I don't quite know why this is, but there's something timeless about shell scripts. I've also seen shell script automation survive for a long time unattended and with zero issues. Not so with some of the modern tools that are supposed to be all unicorns and rainbows.
It all has to be in the script building up strictly from well-understood and long stable basic bricks (and the few places where you don't it's even worse with devops tools.)
Any issue, any question can be answered by reading the damn shell script and you're never dependent on a cookbook/recipe/playbook/component that you got off of some github repo that you need 5% of to do X.
I don't have to rewrite my shell-scripts every 6 months when a new version comes out. New updates usually only happen when security issues arise.
Shell-scripts tend to be simple. There's not a lot of magic hand-holding going on, which means not a lot of complexity to break things.
It keeps you from getting too abstract. Your writing pretty close and specific to what you want it to do, not "how it should be".
They are typically standalone. It's really easy to have 1 script that solves one problem, and another script that solves another. You don't need a giant code-infrastructure to keep things going.
I think config-mgmt tools can be extremely useful if your running a widely-ranged environment. But, you probably shouldn't be running a widely-ranged environment. If you keep things simple, and run as homogeneous as possible, you probably don't need all the added complexity.
This applies to small environments. If the environment is large the situation almost reverses.
Deploying automation throughout a large homogenous environment is where config-management tools really shine. They make it easy to ensure homogeneity is maintained (even if that just means ensuring all machines have the same set of shell scripts) and allow grouping for staggered updates.
If the environment is widely-ranged and large, the utopia starts to break down. Their configuration explodes in complexity and (if you're not careful) you end up with mostly the same amount of work as if they were managed as small independent environments. With the added risk that there is now a single place from where you can break everything at once.
And this happens... Usually from wrong assumptions of what's common between all machines in the environment. In homogeneous environments almost everything is common, but in widely-ranged environments you sometimes add some configuration that wasn't there before and you think applies to the whole set and all hell breaks loose. If you're lucky this will happen suddenly, if you're not, breakage will spread slowly and you'll spend quite a lot of time scratching your head on why.
I don't think large/small is a good deciding factor. You can be large and homogeneous, or small and diverse. I think similar/dissimilar is a better decider for config-mgmt vs shell-scripts.
I'd argue that config-mgmt usually does a better job if your setup is large and complex. No need to write a script that checks if it needs to install a .deb, .rpm, or whatever, if your config-mgmt tools have already done that work.
Also, if you build your shell-scripts right, they can ensure that your system is kept the same.
With tools like Rancher http://rancher.com, you can already see things moving in that direction. Next step is rancher-as-a-service.
When it comes to developers, I think open systems will always prevail in the end (it's just more flexible).
I got really upset about this rancher tool because it doesn't design my database schema.
Shit, future was so close.
Right now, we think of frameworks as being components (part of) larger software systems - But in the future, frameworks will provide the foundations for entire software systems - They will be be responsible for declaring their own network topologies and resource requirements and they will be capable of scaling automatically to any number of machines.
Developers will extend and customize the framework with their own logic but the framework itself will handle all the difficult stuff related to its own operations.
Having read the article back then (and reread it now) it seems like it's still relevant. Maybe we'll have to add the year qualifier after a while when AWS lambda becomes "the way".
Is there an advantage to using docker when it takes 3 hours to rebuild our relatively small database?
1) Small teams (~1-5 people) trying to seem "big" by working at Google's scale.
2) Heroku's prices. We are currently (successfully so far) migrating a small Django project from bare Amazon EC2 instances to ECS with Docker. Even using 3 EC2 micro instances (1 vCPU, 1 GB RAM) for the Docker cluster we would spend ~8 USD/month/instance. With Heroku the minimum would be 25 USD/month/dyno. That's a 3x increase in expenses.
It's very possible to take advantage of technologies like containers without getting too caught in the hype.
or you just host it on an old laptop hidden in your closet.
Well, even someone who's young (for values of 'young' older than high-school age) is probably spending more than that every month on beer, food & entertainment each …
Once again keep in mind that for new projects the process is so streamlined it will take a fraction of the time to set them up.
And 25 and 75 are bogus numbers, what if we start running 10 instances?