This is really interesting, it's (basing it on SQLite) exactly what I was expecting CloudFlare to do for their first DB.
Its perfect for content type sites that want search and querying.
Anyone from CF here, is it using Litestream (https://litestream.io) for its replication or have you built your own replication system?
I assume this first version is somewhat limited on write performance having a single "main" instance and SQLite laking concurrent writes? It seems to me that using SQLite sessions[0] would be a good way to build an eventually consistent replication system for SQLite, would be perfect for an edge first sql database, maybe D2?
One thing I hope to see in the future is a better product filtering experience. When I worked on a jquery product filter I realized the DOM bloat was the main problem.
I wonder if D1 can help devs build instant product filtering pages that don’t require the reload like microcenter or Newegg does.
At any sufficient scale, it is difficult to do filtering on the client. Yes, it can be done, but with 10,000+ potential records, you don’t want to ship that to the client for each query. (Note: I’m thinking Newegg scale for “hdmi cable” here. There are certainly situations where you can ship the entire database to the client for filtering.)
It’s not DOM bloat… it’s too many records. If you’re building a DOM node for each record, that’s bloat, but you still have the problem even if the results are stored in a JSON object and dynamically queried on the client side.
So, for each new filter or new query you need to hit the server anyway. If that’s an asynchronous query that returns a json blob or a full refresh, IMHO, it doesn’t really matter that much. Either way, you’re rebuilding a large portion of the DOM with the new results. The only thing that skews things in favor of an async call is if the rest of the page is so heavyweight that reloading the page takes a significant amount of time. This is probably what you’re taking about.
Having a SQLite db close to your worker node really isn’t going to affect this problem all that much.
It's probably better - especially for more advanced search engines - to have an elasticsearch instance or whichever is the more recent example handle product search and filtering like that.
SQLite was originally great for desktop applications.
Problem is, there's still a huge market for these apps but everything has moved to the web (no one is making desktop apps anymore). So having a full-blown RDMS is overkill for these kind of app, and now SQLite is starting to fill these web app needs.
@sqlite - if you are reading this, any word on merging WAL2 and BEGIN CONCURRENT into main? There clearly is a new class of needs to do so in this world that has completely moved over to web app development (which introduces concurrency problems never experienced on desktop). Any thoughts of focusing more on these web related needs for SQLite (or maybe even fork your own code base to have a more enhanced SQLite version targeted at web needs)?
It's high-quality software too. It's well-commented and exceptionally well tested.[^1][^2]
> As of version 3.33.0 (2020-08-14), the SQLite library consists of approximately 143.4 KSLOC of C code. ... By comparison, the project has 640 times as much test code and test scripts - 91911.0 KSLOC.
I don't usually place much stock in those sort of counts, but 640x is notable.
It makes sense considering the wide variety of use-cases, from embedded devices to edge computing and everything in between.
I think it’s long overdue. While SQLite certainly has its limitations, it’s a winner in many categories. Even for sites with mild traffic using ordinary SQLite in PHP like a decade ago, it was always nice to use for its simplicity and the performance was totally acceptable. In comparison, the memory usage of typical relational database servers was high enough to make it hard to fit on a single lowend VPS with the same data and traffic. (I found myself tuning MySQL, but I never needed to tune SQLite.)
The main thing for tuning SQLite will be how to open it, e.g. in write-ahead mode, to turn on foreign keys (this needs to be enabled manually), and whether it should wait to get a database lock on slower hardware before giving up. There's also some gotchas like if you mark an ID column as primary key, it'll use the rowid as key - which can be reused if a row is removed. So you need to explicitly set primary key AND autoincrement, else you're going to have a bad time. (https://www.sqlite.org/autoinc.html)
Right! SQLite is great, but those two are great as well. It seems like the energy should be around "hey, you should consider a local, maybe even in-memory, database for some things!" more so than specifically "SQLite is great" (though it is).
Transactions, locks, queues, etc. No different than multiple app instances changing the same row in other databases.
Any state mutation is ultimately ordered in time and how that that ordering is accomplished depends on the abstractions you're using: in your app, network layer, database, etc.
Why would you use SQLlite once you start dealing with network, just use MySQl or PG.
It's just re-inventing the wheel badly, I need to read the details but basically you're using a tool SQLLite that was not designed to be used outside of a single app use case.
For Cloudflare, they're offering it because it's simple and lightweight, and they already have their Durable Objects product which serves as the transaction ordering mechanism and takes care of writes.
If you're doing it yourself then sure it's probably not the best fit but that's up to you to decide.
SQLLite was not meant to be used by multiple process so you have to build the missing parts yourself, 100% those have more limitations and issues than regular RDBMS that were built for it.
Is think one way to think about this is to have one database being tied to one replica (replicas could handle more than one database). Where (importantly) the idea would be one database for each user. You horizontally scale for the number of users, but each user is only using one end node.
It’s interesting because you have to consider how to scale your database as well as your application. The fact that you don’t have one central database opens up more possibilities. But it doesn’t work for all instances (such as a shared read-write data source for all users). For example, this approach wouldn’t work for something like Twitter (at least the original architecture).
With the mileage (and attention) those new products are getting out of using SQLite, I think Richard Hipp deserves a lot more acknowledgement for creating such an amazing piece of software.
New products getting a lot of mileage out of sqlite is old-hat at this point. It one of those rare evergreen techs - pay attention for a while and this latest round of attention will die down for 6-12 months then someone else will start another round of "look how cool sqlite is".
At least that's been my observation since I started coming around here.
SQLite is great but it's way overhyped and abused on HN. People are very eager to turn SQLite into a durable, distributed database and it's really not meant for that, and by going down that road instead of using something like MySQL or Postgres you're missing out on lots of important functionality and tooling.
I only say this because I have made this mistake at my previous startup. We built these really cool distributed databases on top of a similar storage engine (RocksDB) plus Kafka, but it ended up being more trouble than it was worth. We should have just used a battle-tested relational database instead.
Using SQLite for these applications is really fun, and it seems like a good idea on paper. But in practice I just don't think it's worth it. YMMV though.
I accept that you learned a lot about the limits of combining RockDB with Kafka, especially in the exact way you combined them.
This might have limited utility if the goal were to combine RocksDB with something else. And even less for SQLite and something else.
The big push of interest in SQLite serverside isn't driven by people who have never set up pgbounce, but rather by developers who have both read the SQLite docs very carefully and have used the library extensively, and know what it's good for.
I'm not sure why you concluded that SQLite is the problem when you built a "really cool distributed database" with Kafka. Distributed databases are complicated, Kafka's complicated.
If you're saying that a replicated Postgres setup would be simpler than what you're built, I agree; but SQLite+Litestream probably would be too.
Litestream is too much work if you're not using S3: replication over sftp. Even fossil has nicer no nonsense replication done over http/s. It's way easier to set up mysql with replication than manage unix accounts and public keys.
So you didn't use SQLite then? Because RocksDB + Kafka is not similar at all.
Also databases all use the same fundamental primitives and it's up to you to choose the level of abstraction you need. For example, FoundationDB is a durable distributed database that uses SQLite underneath as the storage layer but exposes an ordered key/value API, but then allows you to build your own relational DB on top.
If you just needed distributed SQL because a single instance wasn't enough then there are already plenty of choices like CockroachDB/Yugabyte/TiDB/Memsql/etc that can serve the purpose instead of building your own.
> So you didn't use SQLite then? Because RocksDB + Kafka is not similar at all.
To me, I could make the connection in the sense that just like sqlite, rocksdb is an embedded store, while Kafka can be used to build a replicated log (log device).
> If you just needed distributed SQL because a single instance wasn't enough then there are already plenty of choices...
Well, that was GP's point, too? In addition, they mention that existing DBMS like Postgres have way more breath and depth than a replicated sqlite can ever hope to have (which isn't really a controversial assertion at all, tbh).
It's actually quite similar. Both are embedded storage engines that are designed for a single node.
Actually, the case for RocksDB for backing a distributed data store is probably much stronger than SQLite given that it supports multiple concurrent writers.
SQLite lacks many important characteristics that one would expect a distributed data store to have. Row level locking is one obvious feature that's super important in a highly concurrent context (as mentioned, RocksDB has this). Want to backup your production DB? You're going to need to block all writes until the backup completes.
Additionally, features like profiling and replication are nonexistent or immature with SQLite. Rqlite and Litestream are super new relative to tools like Postgres and MySQL and you can't find a lot of people that know how to run them.
Also, you can't autoscale your app since your processes are now highly stateful. Sure, this is a problem with MySQL/Postgres too, but I can pay AWS or Google Cloud for a managed version that will abstract this problem away from me.
Most of these problems are solvable with enough net new software on top of SQLite. But... why? I think the only reason you'd subject yourself to such an architecture is because you want to learn (great!) or you're gunning for that next promotion and need to show off your system design skills :P
But that's orthogonal to my point. As a user of FoundationDB, you're not programming directly against SQLite, so you aren't going to run into these issues as much since FoundationDB exposes different semantics and coordinates concurrency across many SQLite instances in parallel.
I think it's best to think of SQLite as a replacement for your filesystem, rather than a replacement for your relational DBMS.
Any current or planned support for existing ORMs, such as Prisma or TypeOrm?
Also, I wonder how hard it will be to migrate existing PostgreSQL databases and SQL statements. Of course, I understand if Cloudflare is focused on greenfield applications.
We are definitely interested in ORMs. Want to make it easy to use. I hope someone creates the next Rails using Workers. And having other models on top of our SQL offerings will be important. Get in contact and let us know what you'd like.
Will not any existing ORM that supports SQLite support D1? I looked in the post for details on how it extends SQLite (is the query language different or extended, semantics very different, etc.) but didn't notice anything.
> I hope someone creates the next Rails using Workers
I too am eagerly waiting for a good serverless nodejs framework that is "batteries included". I've deployed on Lambda using the "Serverless Framework" but once your app grows to a certain size everything starts to fall apart and you lose some of the magic. Unfortunately, most of the things that advertise themselves as serverless/lambda/worker nodejs frameworks are monoliths and/or an existing monolith framework that "supports" lambda (with a billion asterisks after that). There is absolutely nothing wrong with monolith frameworks, I love them, but just not for lambda, I want to deploy a single endpoint as a single function (or as a cron, or queue listener, etc), not all of my code for every function (you hit size limits quick with this method).
I want express/nestjs/etc-type routes that I define with code or annotations that result in /only/ that function (endpoint) being bundled up and deployed. I ended up rolling my own "framework" on top of Serverless Framework (uses serverless.ts config file that scans my directories for a special file that defines the routes defined in that directory) but Serverless Framework is pretty shaky ground. Their documentation is a mess, Serverless Components appears dead, and they seem to be busy with their own "cloud" so I don't know how much longer I can keep building on top of them.
When it works it's like magic but there are a ton of walls you run headfirst into: Cloud formation entity limits, package size limits, typescript/bundling support, clear disregard for medium/large projects ("Just use multiple services", this leads to a terrible dev experience), and long deploy times.
I wish CF Workers had been out when I first started building my current project, I might have gone in that direction instead, I still might.
I'm not sure, I've not done the full research into CF Workers since I'm on AWS Lambda right now and don't have the capacity to evaluate alternatives. I just like a lot of the CF products and their general ethos/vibe so I'm interested in it. Who knows, it might have a whole new set of issues (most likely) but I don't know if those issues are worse or better than what I'm dealing with now.
I'm building an open source firebase alternative using sqlite. I'll be reaching out, I was thinking to build the distribution & durability part myself, but I would rather use D1!
Deno can compile to wasm, so it can plug in through that vertical. But it's just TS on the frontend.
I'm mainly a python programmer, but Deno's been the most alluring development in the JS ecosystem since typescript for me. Might be helpful to you all to capture some steam from source.
Before you consider using an ORM, try using regular SQL and some tooling first; your future self will thank you. Just write the code, it's only volume and it's not so bad.
I took this advice on my last project and ended up re-writing the whole thing to use Prisma later. I launched and had a successful event with raw sql but it quickly became unwieldy. Prisma gives me type safety throughout my app (written in Typescript) and would have prevented a number of bugs/pain points as my app grew. And I'm only 1 developer, this gets worse if you have multiple people working on it. I still write raw sql for reporting/aggregation (Prisma's features here only work for basic examples in my experience) and I'm not "scared of raw sql" but I can move much faster when I have the guardrails of types.
Before you consider using an ORM, try using regular SQL and some tooling first; your future self will thank you. Just write the code, it's only volume and it's not so bad. What is bad is learning a 3rd language on top of SQL and JS/TS that you somehow have to manually map to SQL.
Prisma won't work with D1 out of the box. The primary limitations are:
- SQLite is traditionally embedded in an application, so Prisma interacts with it by mounting a file. Workers does not have a local filesystem, and D1 is exposed over the network through an API accessible from a Worker. Prisma will have to create a specific connector for D1.
- Workers have a script size limit which is currently 1MB. My understanding is that Cloudflare will be increasing this in the future. We also have specific work to decrease the size of Prisma. Both of those will have to happen before Prisma could be used with D1.
Note that Prisma already support querying Postgres, MySQL, SQL Server and MongoDB from Cloudflare Workers through the Prisma Data Proxy, which will see a GA release next month.
We are also very excited about D1 as a way to bring a subset of data closer to users in order to deliver faster experiences. We hope this will be a way to bring the benefit of edge computing to larger organisations who cannot simply rearchitect everything to run on Workers.
> We are also very excited about D1 as a way to bring a subset of data closer to users in order to deliver faster experiences. We hope this will be a way to bring the benefit of edge computing to larger organisations who cannot simply rearchitect everything to run on Workers.
First, super excited by having Cloudflare offer a RDMS (can SQLite be called that?)
This enables entirely new classes of applications where everything can now be hosted by Cloudflare.
Questions:
a. To help with concurrent writes, will Cloudflare be using WAL2 and BEGIN CONCURRENT branches of SQLite?
b. How is Cloudflare replicating the data cross region? Will it be Litestream.io behind the scenes?
c. Will our Worker code need to be written differently to ensure only a single-writer is writing to SQLite database?
d. How does data persistency and database file size get factored in? I have to imagine their is a limit to how much storage can be used, whether or not that storage is local to the Worker machine, and if its persistent.
Backblaze B2 customers will be able to download data stored in B2 to Cloudflare for zero transfer fees. This happens automatically once Cloudflare is configured to distribute your B2 files.
I really don't care about the cost of storage. In my case it's the bandwidth costs that were killing me.
If it was available at the time, I would use R2 if only for simplicity.
If I was using Cloudflare Workers it would be another reason to use R2: I assume that it's easier to use and faster to use than any other storage system, since it's on the same network and written by the same people.
Also, exposing Backblaze via Cloudflare has it's issues. I ran into Cloudflare caching 404 responses from Backblaze and Backblaze being slow to make write visible.
So I would write into Backblaze and tried to access that key via Cloudflare proxy. While the write was acknowledged to my client it wasn't yet visible via http endpoint so Cloudflare would cache 404 response. I would have to clear the cache to fix and then I've added 5 min delay "just in case" to work around this.
Yes, like it was yesterday (or today). It was a strange time where the term was often used for things that had nothing to do with the original meaning of the term.
I'm right there with you. I wonder if this is an SQLite compatible API on top of their own solution, or if it's using actual SQLite under the hood with custom replication.
If the latter, and anyone from CloudFlare is here, is there any chance to have SpatiaLite enabled?
Well, it's sqlite... so presumably you will get most of the capabilities sqlite has.
RETURNING is covered.
Stored procedures are indirectly there by running your own code "next to the database", as mentioned in the post. Which is arguably much nicer than having to use some database specific language, given that you can run WASM on workers.
There is a layer on top of Sqlite here, so I imagine it's something less than all the capabilities sqlite has, at least initially. Plus the upsides and downsides from their approach to have a master and read replicas.
> Stored procedures are indirectly there by running your own code "next to the database",
"indirectly" is a keyword here, because running code when data is modified potentially won't replace triggers since they'll probably execute outside the running transaction.
Cloudflare and fly.io both promise hassle free read replicas and backup. They will both offer only a single node capable of writes, because that’s how SQLite rolls.
This is a pretty good fit for a read heavy load that requires SQL and very low latency.
Biggest benefit over hosted PostgreSQL is that you get SELECT queries that are measured in microseconds, because SQLite avoids needing network overhead per query.
Yes for writes, but it shouldn't for reads: it looks like it works by replicating the full database down to each edge location where the code is running.
The important drawback is async replication and therefore the lack of full consistency. On the other hand, this is the big advantage of hosted Postgres and the like.
Those offerings are great for use-cases that don't need that kind of consistency, which are many.
No and no. I think this is great for Edge computing, where there is currently no solution. So, it's better than nothing.
It all depends on the use-case, of course. A traditional hosted Postgres or MySQL database or cluster is certainly the go-to solution for all who need advanced features or full consistency, which only synchronous replication could provide.
It isn't as much as folks who need Postgres features are moving to SQLite just because it is cool, but it is folks who don't want those Postgres features moving to SQLite, because the latter has just enough features they only ever really need.
SQLite made sense as an embedded database on day a desktop or phone because there’s only a single person generally writing to it. The perfect use case.
I don’t understand how it will be usable at all in a website with multiple users. Is the idea to make your site to every user gets their own database? How do you stop SQL injection?
Once you solve all of these problems aren’t you better off just using Postgres?
> I don’t understand how it will be usable at all in a website with multiple users
With WAL mode enabled the database is locked during writes only, and concurrent writes are queued but you can still perform reads concurrently. If you keep your write transactions small and consider that a lot of apps aren't writing a lot, it can give perfectly good performance for a lot of usecases.
> Is the idea to make your site to every user gets their own database?
You can do... I know of B2B apps that give each billable customer their own database.
> How do you stop SQL injection?
In the exact same way you do in all other flavours of SQL - with parameterized queries.
> Once you solve all of these problems aren’t you better off just using Postgres?
Not necessarily. Postgres gives you a different set of problems and limitations to consider and work around.
Most of the problems with Postgres are a result of it not being embedded. If you’re using SQLite in an non embedded fashion I don’t see how you don’t inherit the same problems.
I think they're replicated asynchronously, so reading directly from the replica may return old data. That's why they've added the ability to deploy special workers that "live" closer to the primary:
> Embedded compute
> But we're going further. With D1, it will be possible to define a chunk of your Worker code that runs directly next to the database, giving you total control and maximum performance — each request first hits your Worker near your users, but depending on the operation, can hand off to another Worker deployed alongside a replica or your primary D1 instance to complete its work.
First, I'm very excited. Sure, SQLite has some limitations compared to Postgres, esp. regarding the type system and concurrency. But we get ACID compliance and SQL.
But it is really hard getting some useful information from this article. I can't even tell if it is not there or just buried in all this marketing hot air.
So, what is it really? Is there one Write-Master that is asynchronously replicated to all other locations? Will writes be forwarded to this master and then replicated back?
I'm very curious about how it performs in real life. Especially considering the locking behavior (SQLite has always the isolation level 'serializable' iirc). The more you put in a transaction or the longer you have to wait for another process to finish their writes, the more likely you have to deal with stale data.
But overall I'm very excited. Also by the fly.io announcement, of course. Lots of innovation and competition. Good times for customers.
>So, what is it really? Is there one Write-Master that is asynchronously replicated to all other locations? Will writes be forwarded to this master and then replicated back?
Not a lot of detail, but that is mentioned:
"But we're going further. With D1, it will be possible to define a chunk of your Worker code that runs directly next to the database, giving you total control and maximum performance—each request first hits your Worker near your users, but depending on the operation, can hand off to another Worker deployed alongside a replica or your primary D1 instance to complete its work."
sqlite is a great embedded database and thanks to use by browsers and on mobile the most used database in the world by orders of magnitude.
But it also comes with lots of limitations.
* there is no type safety, unless you run with the new strict mode, which comes with some significant drawbacks (eg limited to the handful of primitive types)
* very narrow set of column types and overall functionality in general
* the big one for me: limited migration support, requiring quite a lot of ceremony for common tasks (eg rewriting a whole table and swapping it out)
These approaches (like fly.io s) with read replication also (apparently?) seem to throw away read after write consistency. Which might be fine for certain use cases and even desirable for resilience, but can impact application design quite a lot.
With sqlite you have do to a lot more in your own code because the database gives you fewer tools. Which is usually fine because most usage is "single writer, single or a few local readers". Moving that to a distributed setting with multiple deployed versions of code is not without difficulty.
This seems to be mitigated/solved here though by the ability to run worker code "next to the database".
I'm somewhat surprised they went this route. It probably makes sense given the constraints of Cloudflares architecture and the complexity of running a more advanced globally distributed database.
On the upside: hopefully this usage in domains that are somewhat unusual can lead to funding for more upstream sqlite features.
Interesting that D1 is built on top of Durable Objects. Does this mean that it would be practical for a single worker to access multiple D1 databases, so it could use, for example, a separate database for each tenant in a B2B SaaS application? Edit: And could each database be in a different primary region?
I’m not familiar with Durable Objects. When D1 does replication to read replicas, if it’s not doing synchronous replication, then it’s not strongly consistent, is that correct?
That is interesting. I wish CF would give us some more information as I've assumed that there must be a lack of strong consistency which would be a major drawback.
Edit: But that would mean that durable objects can't be replicated asynchronously? That would mean a big latency hit. Then what's the difference to a central DB in one datacenter?
I'd like to see some up front D1 & R2 benchmarks (read/write/iops). I can't judge invocation cost value until I can judge my use case performance. Here's hoping its -gt NVMe raid 10 under the hood of D1 as some big SQLite reads suffer under slow storage.
"D1 will create read-only clones of your data, close to where your users are, and constantly keep them up-to-date with changes."
Sounds like there will be no synchronous replication and instead there will be a background process to "constantly keep [read-only clones] up-to-date". This means that a stale read from an older read replica can occur even after a write transaction has successfully committed on the "primary" used for writes.
So, while the consistency is not "thrown away", it's no longer a strong consistency? Anyway, Kyle from Jepsen will figure it out soon, I guess :)
Yeah, so you can always opt-in to strong consistency by transferring execution to the primary (see the "Embedded Compute" section of the blog). Then it's pretty much exactly the same as a DO.
Just clarifying - D1 without read replicas is strongly consistent. If you add read replicas, those can have replication lag and will not be strongly consistent.
Thanks for the clarification, that is what I would expect.
Does SQLite support some kind of monotonic transaction id that can be used as a cache coherency key? Say a client writes a new record to the database which returns `{"result": "ok", "transaction_id": 123}`, then to ensure that subsequent read requests are coherent they provide a header that checks that the read replica has transaction_id >= 123 and either waits for replication before serving or fails the request. (Perhaps a good use for the embedded worker?)
Yes you could do it manually, but it would be nice if the solution didn't require carefully managing update queries so the journal addition isn't missed and increasing write amplification to manually update a journal table when that information probably already exists somewhere in the WAL implementation.
I haven't tried duckdb but I have been googling about it. I think I saw a discussion where it was mentioned that duckdb isn't a replacement for SQLite. It is an OLAP database [0] which makes its ingestion time slower than SQLite, I think. So it is meant for analytics but not as fullfledge replacement for SQLite.
Why? Yes sqlite doesn't have all the features postgres has. Postgres doesn't have all the features the sqlite has either. What's wrong with having different tools with different sets tradeoffs. It's a different shape of Lego and that's fine - some things call for a 1/3height 2x2 and others call for a full height 1x8.
* the big one for me: very limited migration support, requiring quite a lot of ceremony for common tasks (eg rewriting a whole table and swapping it out)
I don't know where this idea of having to swap a whole table in SQLite came from, but it simply isn't true. Over the last 13 years I have upgraded production HashBackup databases at customer sites a total of 35 times without rewriting and swapping out tables by using the ALTER statement, just like other databases:
For the most recent upgrade, I upgraded to strict tables, which I could also have done without a rebuild/swap. I chose to do a rebuild/swap this one time because I wanted to reorder some columns. Why? Because columns stored with default or null values don't have row space allocated if the column is at the end of the row.
For a long time sqlite did not have DROP COLUMN and RENAME COLUMN support, which are both pretty essential.
I'm embarrassed to admit that I didn't realize RENAME COLUMN was actually added in 3.25, almost four years ago.
DROP COLUMN was only just added last year in 3.35.
I'm surprised a database schema lasted 9/12 years without ever renaming or dropping a column.
This changes things! But even now, ALTER TABLE is not transactional. So especially with many concurrent readers there can definitely be situations where you'd still want to rewrite.
I'm not sure what you mean by "not transactional". SQLite implements transaction support at the "page" level, and builds all other database operations on top of it, which means anything that touches the bytes of the database file is transaction-safe. You can verify this for yourself:
sqlite> CREATE TABLE foo(a,b,c);
sqlite> INSERT INTO foo VALUES (1,2,3);
sqlite> BEGIN;
sqlite> ALTER TABLE foo DROP COLUMN b;
sqlite> SELECT * FROM foo;
1|3
sqlite> ROLLBACK;
sqlite> SELECT * FROM foo;
1|2|3
It's of course still subject to SQLite's normal restrictions on locking, which means a long-running ALTER statement will block concurrent writers (and probably also concurrent readers if you're not running in WAL mode).
> I'm surprised a database schema lasted 9/12 years without ever renaming or dropping a column.
I did have a couple of columns that were no longer needed and would have dropped them, but instead I just set them to null and ignored them. Nulls only take 1 byte of space in a row. I dropped them when DROP COLUMN was added.
It would really help if SQLite3 had a `MERGE`, or, failing that, `FULL OUTER JOIN`. In fact, I want it to have `FULL OUTER JOIN` even if it gains a `MERGE`.
`FULL OUTER JOIN` is the secret to diff'ing table sources. `MERGE` is just a diff operation + insert/update/delete statements to make the target table more like the source one (or even completely like the source one).
`FULL OUTER JOIN` is essential to implementing `MERGE`. Granted, one could implement `MERGE` without implementing `FULL OUTER JOIN` as a public feature, but that seems silly.
Sadly, the SQLite3 dev team specifically says they will not implement `FULL OUTER JOIN`[0].
Implementing `MERGE`-like updates without `FULL OUTER JOIN` is possible (using two `LEFT OUTER JOIN`s), but it's an O(N log N) operation instead of O(N).
The lack of `FULL OUTER JOIN` is a serious flaw in SQLite3. IMO.
RIGHT and FULL JOIN are on the trunk branch of SQLite and will (very likely) appear in the next release. Please grab a copy of the latest pre-release snapshot of SQLite (https://sqlite.org/download.html) and try out the new RIGHT/FULL JOIN support. Report any problems on the forum, or directly to me at drh at sqlite dot org.
* Copying the database around its a file copy in sqlite. Each database is it's own single file. (there's also WAL stuff that you get control of)
* No extra service to deploy, manage, and/or optimize. I don't fully agree with the following, but I had a colleague who used to say "If you don't have multiple app servers writing to the db, postgres is a waste of effort".
* embedded means way lower data latency - if the dataset is in the fs cache even lower, no waiting on network transactions.
I've frequently chosen it over PG in cases where I needed basic relational data operations. In one case we ingested a large dataset (a few gbs of measurements) once an hour. Then we did some initial analytics on those measurements and threw the results in the same db file. After that step was done, the data was read only for several other systems and we just copied the db file to each of the systems that needed the data on-demand. A couple of the systems did additional analytics and effectively imported the db file to a different db (one was PG another was a graph db - neo4j). A couple of the systems just used the db file directly. It worked our really well.
less things to go wrong. This gives you benefits all the way up the dev stack. Changing an integration test from needing its own db server installed to just a couple of files on disk is a big difference in complexity. You can probably run that test with just a local disk almost infinitely entirely deterministically, conversely as soon as you go onto the network all bets are off.
Granted, if you already have the tooling its less of a big deal but if you don't then you need the tooling, e.g. your build and test machines need to have access to some sql installation somewhere and that the process of maintaining that can be a flaky one.
Wow, this looks potentially very interesting. Since this is sort of fresh in my mind from the recent Fly post about it:
* How exactly is the read replication implemented? Is it using litestream behind the scenes to stream the WAL somewhere? How do the readers keep up? Last I saw you just had to poll it, but that could be computationally expensive depending on the size of the data (since I thought you had to download the whole DB), and could potentially introduce a bit of latency in propagation. Any idea what the metrics are for latency in propagation?
* How are writes handled? Does it do the Fly thing about sending all requests to one worker?
I don't quite know what a "worker" is but I'm assuming it's kind of like a Lambda? If you have it replicated around the world, is that one worker all running the same code, and Cloudflare somehow manages the SQL replicating and write forwarding? Or would those all be separate workers?
"With D1, it will be possible to define a chunk of your Worker code that runs directly next to the database...each request first hits your Worker near your users, but depending on the operation, can hand off to another Worker deployed alongside a replica or your primary D1 instance to complete its work."
That's interesting to me. It opens the door for Cloudflare to offer something more like a "normal" serverless offering. One that can run containers, or least natively run Python/Golang/Java/etc, like AWS Lambda does. And with this ecosystem described above that can conditionally route between the lighter edge Workers and the heavier central serverless functions. To me, that's the tipping point where they start to threaten larger portions of AWS.
The demo is also a bit buggy: orders are duplicated as many times as there are products, but clicking on the various lines of the same order leads to the same record, where the user can only see the first product...
I also think the demo would have more impact if it wasn't read-only (although I understand that this could lead to broken pages if visitors mess up with the data).
Very cool! Glad to see all the love for SQLite recently.
One thing I've noticed that many commenters miss about read-replicated SQLite is assuming that the only valid model is having one, giant, centralized database with all the data. Lets be honest with ourselves, the vast majority of applications hold personal or B2B data and don't need centralized transactions, and at scale will use multi-tenant primary keys or manual sharding anyways. For private data, a single SQLite database per user / business will easily satisfy the write load of all but the most gigantic corporations. With this model you have unbounded compute scaling for new users because they very likely don't need online transactions across multiple databases at once.
Some questions:
Will D1 be able to deliver this design of having many thousands of separate databases for a single application? Will this be problematic from a cost perspective?
> since we're building on the redundant storage of Durable Objects, your database can physically move locations as needed
Will D1 be able to easily migrate the "primary" at will? CockroachDB described this as "follow the sun" primary.
I guess the first answer is: similar to Durable Object limits (unlimited databases / 50 GB total) since they alluded to those abilities more so than a simple file stored on R2 (only for backups).
The unique thing about Cloudflare's product offerings is how global-first they are; traditional cloud providers (AWS to DigitalOcean) have a very region-oriented domain model, with select christened services allowed or architected to be global (ex: AWS Cloudfront, IAM, Route53, that's about it there). That's their disaster/failure model; but all it really does is force cross-regional architecture onto the customer. Most customers don't bother.
In comparison, everything at CF is global. And its not just "global" from an AWS perspective of "we've got 14 regions and your stuff runs in all of them"; its global from 300+ points-of-presence, within 50ms of like 98% of all humans. CDN for compute, databases, etc.
CF has a way to go in DevEx on many of their products. For example; Workers, being based on V8 Isolates, is a pain to use even compared to e.g. Lambda. It's a battle of figuring out what's possible and what isn't within the runtime. But I'm sure it'll be improved!
240 comments
[ 3.0 ms ] story [ 358 ms ] threadIts perfect for content type sites that want search and querying.
Anyone from CF here, is it using Litestream (https://litestream.io) for its replication or have you built your own replication system?
I assume this first version is somewhat limited on write performance having a single "main" instance and SQLite laking concurrent writes? It seems to me that using SQLite sessions[0] would be a good way to build an eventually consistent replication system for SQLite, would be perfect for an edge first sql database, maybe D2?
0: https://www.sqlite.org/sessionintro.html
2. We are going all in on databases and D2 sounds like a cool name for something...
Maybe they will announce a Hashicorp competitor in their next reveal. Who knows.
One thing I hope to see in the future is a better product filtering experience. When I worked on a jquery product filter I realized the DOM bloat was the main problem.
I wonder if D1 can help devs build instant product filtering pages that don’t require the reload like microcenter or Newegg does.
IE https://www.newegg.com/p/pl?d=hdmi+cable&N=-1&SortType=8
It’s not DOM bloat… it’s too many records. If you’re building a DOM node for each record, that’s bloat, but you still have the problem even if the results are stored in a JSON object and dynamically queried on the client side.
So, for each new filter or new query you need to hit the server anyway. If that’s an asynchronous query that returns a json blob or a full refresh, IMHO, it doesn’t really matter that much. Either way, you’re rebuilding a large portion of the DOM with the new results. The only thing that skews things in favor of an async call is if the rest of the page is so heavyweight that reloading the page takes a significant amount of time. This is probably what you’re taking about.
Having a SQLite db close to your worker node really isn’t going to affect this problem all that much.
https://tailscale.com/blog/database-for-2022
https://fly.io/blog/all-in-on-sqlite-litestream
https://blog.cloudflare.com/introducing-d1
Problem is, there's still a huge market for these apps but everything has moved to the web (no one is making desktop apps anymore). So having a full-blown RDMS is overkill for these kind of app, and now SQLite is starting to fill these web app needs.
@sqlite - if you are reading this, any word on merging WAL2 and BEGIN CONCURRENT into main? There clearly is a new class of needs to do so in this world that has completely moved over to web app development (which introduces concurrency problems never experienced on desktop). Any thoughts of focusing more on these web related needs for SQLite (or maybe even fork your own code base to have a more enhanced SQLite version targeted at web needs)?
https://en.wikipedia.org/wiki/POPFile https://getpopfile.org/browser/trunk/engine/POPFile/Database...
> As of version 3.33.0 (2020-08-14), the SQLite library consists of approximately 143.4 KSLOC of C code. ... By comparison, the project has 640 times as much test code and test scripts - 91911.0 KSLOC.
I don't usually place much stock in those sort of counts, but 640x is notable.
It makes sense considering the wide variety of use-cases, from embedded devices to edge computing and everything in between.
[1]: https://www.sqlite.org/testing.html [2]: https://sqlite.org/src/dir?ci=trunk
The fact that it's just a file on disk limits the usage.
Any state mutation is ultimately ordered in time and how that that ordering is accomplished depends on the abstractions you're using: in your app, network layer, database, etc.
It's just re-inventing the wheel badly, I need to read the details but basically you're using a tool SQLLite that was not designed to be used outside of a single app use case.
For Cloudflare, they're offering it because it's simple and lightweight, and they already have their Durable Objects product which serves as the transaction ordering mechanism and takes care of writes.
If you're doing it yourself then sure it's probably not the best fit but that's up to you to decide.
It’s interesting because you have to consider how to scale your database as well as your application. The fact that you don’t have one central database opens up more possibilities. But it doesn’t work for all instances (such as a shared read-write data source for all users). For example, this approach wouldn’t work for something like Twitter (at least the original architecture).
At least that's been my observation since I started coming around here.
I only say this because I have made this mistake at my previous startup. We built these really cool distributed databases on top of a similar storage engine (RocksDB) plus Kafka, but it ended up being more trouble than it was worth. We should have just used a battle-tested relational database instead.
Using SQLite for these applications is really fun, and it seems like a good idea on paper. But in practice I just don't think it's worth it. YMMV though.
This might have limited utility if the goal were to combine RocksDB with something else. And even less for SQLite and something else.
The big push of interest in SQLite serverside isn't driven by people who have never set up pgbounce, but rather by developers who have both read the SQLite docs very carefully and have used the library extensively, and know what it's good for.
If you're saying that a replicated Postgres setup would be simpler than what you're built, I agree; but SQLite+Litestream probably would be too.
Also databases all use the same fundamental primitives and it's up to you to choose the level of abstraction you need. For example, FoundationDB is a durable distributed database that uses SQLite underneath as the storage layer but exposes an ordered key/value API, but then allows you to build your own relational DB on top.
If you just needed distributed SQL because a single instance wasn't enough then there are already plenty of choices like CockroachDB/Yugabyte/TiDB/Memsql/etc that can serve the purpose instead of building your own.
To me, I could make the connection in the sense that just like sqlite, rocksdb is an embedded store, while Kafka can be used to build a replicated log (log device).
> If you just needed distributed SQL because a single instance wasn't enough then there are already plenty of choices...
Well, that was GP's point, too? In addition, they mention that existing DBMS like Postgres have way more breath and depth than a replicated sqlite can ever hope to have (which isn't really a controversial assertion at all, tbh).
Actually, the case for RocksDB for backing a distributed data store is probably much stronger than SQLite given that it supports multiple concurrent writers.
SQLite lacks many important characteristics that one would expect a distributed data store to have. Row level locking is one obvious feature that's super important in a highly concurrent context (as mentioned, RocksDB has this). Want to backup your production DB? You're going to need to block all writes until the backup completes.
Additionally, features like profiling and replication are nonexistent or immature with SQLite. Rqlite and Litestream are super new relative to tools like Postgres and MySQL and you can't find a lot of people that know how to run them.
Also, you can't autoscale your app since your processes are now highly stateful. Sure, this is a problem with MySQL/Postgres too, but I can pay AWS or Google Cloud for a managed version that will abstract this problem away from me.
Most of these problems are solvable with enough net new software on top of SQLite. But... why? I think the only reason you'd subject yourself to such an architecture is because you want to learn (great!) or you're gunning for that next promotion and need to show off your system design skills :P
But that's orthogonal to my point. As a user of FoundationDB, you're not programming directly against SQLite, so you aren't going to run into these issues as much since FoundationDB exposes different semantics and coordinates concurrency across many SQLite instances in parallel.
I think it's best to think of SQLite as a replacement for your filesystem, rather than a replacement for your relational DBMS.
I've been looking for a turn key solution that is better than me running a single node Postgres instance "bare metal" or in a container.
postgres-operator seems cool but... k8s, pretty heavy I guess.
Also, I wonder how hard it will be to migrate existing PostgreSQL databases and SQL statements. Of course, I understand if Cloudflare is focused on greenfield applications.
I too am eagerly waiting for a good serverless nodejs framework that is "batteries included". I've deployed on Lambda using the "Serverless Framework" but once your app grows to a certain size everything starts to fall apart and you lose some of the magic. Unfortunately, most of the things that advertise themselves as serverless/lambda/worker nodejs frameworks are monoliths and/or an existing monolith framework that "supports" lambda (with a billion asterisks after that). There is absolutely nothing wrong with monolith frameworks, I love them, but just not for lambda, I want to deploy a single endpoint as a single function (or as a cron, or queue listener, etc), not all of my code for every function (you hit size limits quick with this method).
I want express/nestjs/etc-type routes that I define with code or annotations that result in /only/ that function (endpoint) being bundled up and deployed. I ended up rolling my own "framework" on top of Serverless Framework (uses serverless.ts config file that scans my directories for a special file that defines the routes defined in that directory) but Serverless Framework is pretty shaky ground. Their documentation is a mess, Serverless Components appears dead, and they seem to be busy with their own "cloud" so I don't know how much longer I can keep building on top of them.
When it works it's like magic but there are a ton of walls you run headfirst into: Cloud formation entity limits, package size limits, typescript/bundling support, clear disregard for medium/large projects ("Just use multiple services", this leads to a terrible dev experience), and long deploy times.
I wish CF Workers had been out when I first started building my current project, I might have gone in that direction instead, I still might.
I'm building a serverless firebase alternative that uses SQLite. If CF gives me access I will totally support D1 & workers.
Check it out: javascriptdb.com
I guess it would count as a client focused ORM :)
I'll be reaching out from jp@javascriptdb.com
Great addition, congrats!
Deno can compile to wasm, so it can plug in through that vertical. But it's just TS on the frontend.
I'm mainly a python programmer, but Deno's been the most alluring development in the JS ecosystem since typescript for me. Might be helpful to you all to capture some steam from source.
[1]: https://deno.land/
Source: someone who avoided learning SQL for 20 years.
- SQLite is traditionally embedded in an application, so Prisma interacts with it by mounting a file. Workers does not have a local filesystem, and D1 is exposed over the network through an API accessible from a Worker. Prisma will have to create a specific connector for D1. - Workers have a script size limit which is currently 1MB. My understanding is that Cloudflare will be increasing this in the future. We also have specific work to decrease the size of Prisma. Both of those will have to happen before Prisma could be used with D1.
Note that Prisma already support querying Postgres, MySQL, SQL Server and MongoDB from Cloudflare Workers through the Prisma Data Proxy, which will see a GA release next month.
We are also very excited about D1 as a way to bring a subset of data closer to users in order to deliver faster experiences. We hope this will be a way to bring the benefit of edge computing to larger organisations who cannot simply rearchitect everything to run on Workers.
I am also excited about this :)
This enables entirely new classes of applications where everything can now be hosted by Cloudflare.
Questions:
a. To help with concurrent writes, will Cloudflare be using WAL2 and BEGIN CONCURRENT branches of SQLite?
b. How is Cloudflare replicating the data cross region? Will it be Litestream.io behind the scenes?
c. Will our Worker code need to be written differently to ensure only a single-writer is writing to SQLite database?
d. How does data persistency and database file size get factored in? I have to imagine their is a limit to how much storage can be used, whether or not that storage is local to the Worker machine, and if its persistent.
Am I missing something? Is there no bandwidth cost at all?
Backblaze B2 customers will be able to download data stored in B2 to Cloudflare for zero transfer fees. This happens automatically once Cloudflare is configured to distribute your B2 files.
I really don't care about the cost of storage. In my case it's the bandwidth costs that were killing me.
If it was available at the time, I would use R2 if only for simplicity.
If I was using Cloudflare Workers it would be another reason to use R2: I assume that it's easier to use and faster to use than any other storage system, since it's on the same network and written by the same people.
Also, exposing Backblaze via Cloudflare has it's issues. I ran into Cloudflare caching 404 responses from Backblaze and Backblaze being slow to make write visible.
So I would write into Backblaze and tried to access that key via Cloudflare proxy. While the write was acknowledged to my client it wasn't yet visible via http endpoint so Cloudflare would cache 404 response. I would have to clear the cache to fix and then I've added 5 min delay "just in case" to work around this.
With R2, user downloads are free, aren't they?
"Our vision for R2 includes multi-region storage that automatically replicates objects to the locations they’re frequently requested from."
edit: maybe one day! this looks cool regardless
If the latter, and anyone from CloudFlare is here, is there any chance to have SpatiaLite enabled?
https://www.gaia-gis.it/fossil/libspatialite/index
I see cloudflare people are on this post, any chance to compar D1 vs postgres in terms of DB features?
Insert ... Returning
Stored procedures and triggers
Etc etc
Would be really helpful to get a comparison like cockroachDB did here https://www.cockroachlabs.com/docs/stable/postgresql-compati...
Or even better, a general sql compatibility matrix like this https://www.cockroachlabs.com/docs/stable/sql-feature-suppor...
Kudos to the cloudflare team!
RETURNING is covered.
Stored procedures are indirectly there by running your own code "next to the database", as mentioned in the post. Which is arguably much nicer than having to use some database specific language, given that you can run WASM on workers.
"indirectly" is a keyword here, because running code when data is modified potentially won't replace triggers since they'll probably execute outside the running transaction.
Long story short, don't expect anything fancy. Support for alter table is limited, and concurrency can be an issue.
Also, any plans to support PATCH x-update-range so SQLite can be used entirely in the browser via SQLite.js?
Can someone enlighten me with the types of use cases this would be better for vs say Postgres?
Cloudflare and fly.io both promise hassle free read replicas and backup. They will both offer only a single node capable of writes, because that’s how SQLite rolls.
This is a pretty good fit for a read heavy load that requires SQL and very low latency.
https://www.sqlite.org/np1queryprob.html
Those offerings are great for use-cases that don't need that kind of consistency, which are many.
It all depends on the use-case, of course. A traditional hosted Postgres or MySQL database or cluster is certainly the go-to solution for all who need advanced features or full consistency, which only synchronous replication could provide.
I don’t understand how it will be usable at all in a website with multiple users. Is the idea to make your site to every user gets their own database? How do you stop SQL injection?
Once you solve all of these problems aren’t you better off just using Postgres?
With WAL mode enabled the database is locked during writes only, and concurrent writes are queued but you can still perform reads concurrently. If you keep your write transactions small and consider that a lot of apps aren't writing a lot, it can give perfectly good performance for a lot of usecases.
> Is the idea to make your site to every user gets their own database?
You can do... I know of B2B apps that give each billable customer their own database.
> How do you stop SQL injection?
In the exact same way you do in all other flavours of SQL - with parameterized queries.
> Once you solve all of these problems aren’t you better off just using Postgres?
Not necessarily. Postgres gives you a different set of problems and limitations to consider and work around.
From the blog post it says read-only replicas are created close to users and kept up to date with the latest data.
- How should I think about this in terms of CAP? If there's a write and I query a replica what happens?
- How are writes handled? Do they go to a single location or are they handled by various locations?
I'm excited to try this. It's so cool to see databases being distributed "on CDNs" for lack of a better term.
> Embedded compute
> But we're going further. With D1, it will be possible to define a chunk of your Worker code that runs directly next to the database, giving you total control and maximum performance — each request first hits your Worker near your users, but depending on the operation, can hand off to another Worker deployed alongside a replica or your primary D1 instance to complete its work.
But it is really hard getting some useful information from this article. I can't even tell if it is not there or just buried in all this marketing hot air.
So, what is it really? Is there one Write-Master that is asynchronously replicated to all other locations? Will writes be forwarded to this master and then replicated back?
I'm very curious about how it performs in real life. Especially considering the locking behavior (SQLite has always the isolation level 'serializable' iirc). The more you put in a transaction or the longer you have to wait for another process to finish their writes, the more likely you have to deal with stale data.
But overall I'm very excited. Also by the fly.io announcement, of course. Lots of innovation and competition. Good times for customers.
Not a lot of detail, but that is mentioned:
"But we're going further. With D1, it will be possible to define a chunk of your Worker code that runs directly next to the database, giving you total control and maximum performance—each request first hits your Worker near your users, but depending on the operation, can hand off to another Worker deployed alongside a replica or your primary D1 instance to complete its work."
and well R2 and D2 would make for a great naming scheme.
sqlite is a great embedded database and thanks to use by browsers and on mobile the most used database in the world by orders of magnitude.
But it also comes with lots of limitations.
* there is no type safety, unless you run with the new strict mode, which comes with some significant drawbacks (eg limited to the handful of primitive types)
* very narrow set of column types and overall functionality in general
* the big one for me: limited migration support, requiring quite a lot of ceremony for common tasks (eg rewriting a whole table and swapping it out)
These approaches (like fly.io s) with read replication also (apparently?) seem to throw away read after write consistency. Which might be fine for certain use cases and even desirable for resilience, but can impact application design quite a lot.
With sqlite you have do to a lot more in your own code because the database gives you fewer tools. Which is usually fine because most usage is "single writer, single or a few local readers". Moving that to a distributed setting with multiple deployed versions of code is not without difficulty.
This seems to be mitigated/solved here though by the ability to run worker code "next to the database".
I'm somewhat surprised they went this route. It probably makes sense given the constraints of Cloudflares architecture and the complexity of running a more advanced globally distributed database.
On the upside: hopefully this usage in domains that are somewhat unusual can lead to funding for more upstream sqlite features.
Edit: But that would mean that durable objects can't be replicated asynchronously? That would mean a big latency hit. Then what's the difference to a central DB in one datacenter?
It's more of a "quickstart" than a peek under the hood.
Sounds like there will be no synchronous replication and instead there will be a background process to "constantly keep [read-only clones] up-to-date". This means that a stale read from an older read replica can occur even after a write transaction has successfully committed on the "primary" used for writes.
So, while the consistency is not "thrown away", it's no longer a strong consistency? Anyway, Kyle from Jepsen will figure it out soon, I guess :)
Disclaimer: I work at Cloudflare :)
Does SQLite support some kind of monotonic transaction id that can be used as a cache coherency key? Say a client writes a new record to the database which returns `{"result": "ok", "transaction_id": 123}`, then to ensure that subsequent read requests are coherent they provide a header that checks that the read replica has transaction_id >= 123 and either waits for replication before serving or fails the request. (Perhaps a good use for the embedded worker?)
I know of a very important system at AWS that did this with MySQL :D
https://duckdb.org/
[0]: https://en.wikipedia.org/wiki/Online_analytical_processing
Duckdb on HN: https://news.ycombinator.com/item?id=23287278
https://realm.io/
all the hip service providers seem to be all over it which would indicate pretty good modernity to me at least.
I don't know where this idea of having to swap a whole table in SQLite came from, but it simply isn't true. Over the last 13 years I have upgraded production HashBackup databases at customer sites a total of 35 times without rewriting and swapping out tables by using the ALTER statement, just like other databases:
https://www.sqlite.org/lang_altertable.html
For the most recent upgrade, I upgraded to strict tables, which I could also have done without a rebuild/swap. I chose to do a rebuild/swap this one time because I wanted to reorder some columns. Why? Because columns stored with default or null values don't have row space allocated if the column is at the end of the row.
I'm embarrassed to admit that I didn't realize RENAME COLUMN was actually added in 3.25, almost four years ago.
DROP COLUMN was only just added last year in 3.35.
I'm surprised a database schema lasted 9/12 years without ever renaming or dropping a column.
This changes things! But even now, ALTER TABLE is not transactional. So especially with many concurrent readers there can definitely be situations where you'd still want to rewrite.
I did have a couple of columns that were no longer needed and would have dropped them, but instead I just set them to null and ignored them. Nulls only take 1 byte of space in a row. I dropped them when DROP COLUMN was added.
`FULL OUTER JOIN` is the secret to diff'ing table sources. `MERGE` is just a diff operation + insert/update/delete statements to make the target table more like the source one (or even completely like the source one).
`FULL OUTER JOIN` is essential to implementing `MERGE`. Granted, one could implement `MERGE` without implementing `FULL OUTER JOIN` as a public feature, but that seems silly.
Sadly, the SQLite3 dev team specifically says they will not implement `FULL OUTER JOIN`[0].
Implementing `MERGE`-like updates without `FULL OUTER JOIN` is possible (using two `LEFT OUTER JOIN`s), but it's an O(N log N) operation instead of O(N).
The lack of `FULL OUTER JOIN` is a serious flaw in SQLite3. IMO.
Finally!
Thank you so much for this Mr. Hipp!
EDIT: Don't forget to edit the `omitted.html` page when you ship it!
Prisma Migrate can automatically generate these steps, removing most of the pain. I'm sure other migration tools can do this as well.
In 2-3 years from now, these services will be so mature and strong they will be crushing the cloud market.
They're turning dreams into reality, one after another.
> We will ensure that D1 costs less and performs better than comparable centralized solutions.
* Copying the database around its a file copy in sqlite. Each database is it's own single file. (there's also WAL stuff that you get control of)
* No extra service to deploy, manage, and/or optimize. I don't fully agree with the following, but I had a colleague who used to say "If you don't have multiple app servers writing to the db, postgres is a waste of effort".
* embedded means way lower data latency - if the dataset is in the fs cache even lower, no waiting on network transactions.
I've frequently chosen it over PG in cases where I needed basic relational data operations. In one case we ingested a large dataset (a few gbs of measurements) once an hour. Then we did some initial analytics on those measurements and threw the results in the same db file. After that step was done, the data was read only for several other systems and we just copied the db file to each of the systems that needed the data on-demand. A couple of the systems did additional analytics and effectively imported the db file to a different db (one was PG another was a graph db - neo4j). A couple of the systems just used the db file directly. It worked our really well.
Granted, if you already have the tooling its less of a big deal but if you don't then you need the tooling, e.g. your build and test machines need to have access to some sql installation somewhere and that the process of maintaining that can be a flaky one.
* How exactly is the read replication implemented? Is it using litestream behind the scenes to stream the WAL somewhere? How do the readers keep up? Last I saw you just had to poll it, but that could be computationally expensive depending on the size of the data (since I thought you had to download the whole DB), and could potentially introduce a bit of latency in propagation. Any idea what the metrics are for latency in propagation?
* How are writes handled? Does it do the Fly thing about sending all requests to one worker?
I don't quite know what a "worker" is but I'm assuming it's kind of like a Lambda? If you have it replicated around the world, is that one worker all running the same code, and Cloudflare somehow manages the SQL replicating and write forwarding? Or would those all be separate workers?
That's interesting to me. It opens the door for Cloudflare to offer something more like a "normal" serverless offering. One that can run containers, or least natively run Python/Golang/Java/etc, like AWS Lambda does. And with this ecosystem described above that can conditionally route between the lighter edge Workers and the heavier central serverless functions. To me, that's the tipping point where they start to threaten larger portions of AWS.
The demo is also a bit buggy: orders are duplicated as many times as there are products, but clicking on the various lines of the same order leads to the same record, where the user can only see the first product...
I also think the demo would have more impact if it wasn't read-only (although I understand that this could lead to broken pages if visitors mess up with the data).
Anyway, kudos to the CloudFlare team!
One thing I've noticed that many commenters miss about read-replicated SQLite is assuming that the only valid model is having one, giant, centralized database with all the data. Lets be honest with ourselves, the vast majority of applications hold personal or B2B data and don't need centralized transactions, and at scale will use multi-tenant primary keys or manual sharding anyways. For private data, a single SQLite database per user / business will easily satisfy the write load of all but the most gigantic corporations. With this model you have unbounded compute scaling for new users because they very likely don't need online transactions across multiple databases at once.
Some questions:
Will D1 be able to deliver this design of having many thousands of separate databases for a single application? Will this be problematic from a cost perspective?
> since we're building on the redundant storage of Durable Objects, your database can physically move locations as needed
Will D1 be able to easily migrate the "primary" at will? CockroachDB described this as "follow the sun" primary.
In comparison, everything at CF is global. And its not just "global" from an AWS perspective of "we've got 14 regions and your stuff runs in all of them"; its global from 300+ points-of-presence, within 50ms of like 98% of all humans. CDN for compute, databases, etc.
CF has a way to go in DevEx on many of their products. For example; Workers, being based on V8 Isolates, is a pain to use even compared to e.g. Lambda. It's a battle of figuring out what's possible and what isn't within the runtime. But I'm sure it'll be improved!