great post, Basically put up or shut up! (code that is) ... I hope they get resumes from the post of people who can help.
- Re-writing their whole system incrementally will be expensive; I think they might need google money and brains to complete the whole thing... or a lot more than 15mm.
Rewriting any system does not take much money, it takes a lot of time and a pair of good devoted hackers. If you stuck with bad ones, even money won't help. I believe, twitter's hackers are good enough to deal with the problem and I don't think their system is so complex. It's too young to get covered with all the dust and rust :-).
Nicely written. The authors of nearly all the myriad "how to fix twitter" posts I've read have trivialized scaling a site that has different properties from most online services. Scaling anything to twitter-like traffic levels is hard, and you won't know exactly why it's hard until you do it. I'm deeply skeptical of any armchair architect who says otherwise.
Not sure why you got modded down because it's essentially true and I took their post to be alluding to as much. My intuition tells me that though they are huge, they aren't remotely in the same area as AIM, MSN, GMail, Yahoo Mail, etc. Twitter hangs right in the middle of being an instant messaging and email service.
The post says little that's specific. What it does say - our problem is different, it's hard, we're making progress - might all be true, but these are also the kinds of things people say when a software project is in trouble. It's hard to tell anything at all from reading the post.
Everyone should read this blog post and the 2 followups. It brings a lot of insight into the parts of Twitter that bring it down and it has nothing to do with the message pushing part of Twitter.
I have a Jabber bot running that is hooked into the entire public timeline and it continues to run even when the main website has gone down. I don't think you can directly compare Twitter's issues with issues of Meebo, AIM, GTalk, etc.
Probably never will? The average user may never use a Twitter like platform for banal updates or short, passionate discussion. However, simply sending mass messages to groups is powerful. For example, Pownce's calendar feature is very helpful for sending open invitation with a cal sync to large groups.
Media coverage? I haven't seen anything about Twitter on the New York Times or CNN or any other media outlet targeted at "the average person". Even mainstream-leaning "tech" sites like Wired or CNet haven't been covering the outages/scaling problems.
People who read sites like TechCrunch most definitely know what Twitter is, and coverage that talks about why they have scaling problems as opposed to making a soap opera about it is very useful for that demographic.
I admit their "Something's technically wrong here" error was getting old, and they needed to address the issues that have made that message so pervasive lately. So they did (if indirectly).
But people like to act like Twitter owes them something. Fact is, it's a free service.
In the spirit of building their business they clearly want -- in fact they need -- to build the best product they can; the fact that it's taking too long by our estimation doesn't matter. They're not going to leave money on the table so you can believe they're working hard to fix the problem and start working on their revenue model.
Kudos to the Twitter team for returning critical volleys from some in the tech community without losing their cool.
It's hard for me to imagine why they're choosing Scala over Erlang. It seems to me like one of their major requirements is to have task distribution.
Erlang makes distribution and concurrent execution semantically identical, Scala doesn't. And there doesn't seem to be a good reason to base on the JVM, they aren't already next deep in Java libraries.
I'd really like to know what they're thinking over there.
This task does seem to have Erlang written all over it, but of course we're not intimately familiar with all the details.
The mindset for selecting tools to build apps which need to scale seems be similar to the one we all had back in early dos days, when the company programmers were having to write all the internal software. We were either unaware of better solutions or we simply felt compelled to write all the stuff ourselves, from spreadsheets to a GL module. Now, with all the customizable apps around, why would anyone want to write a custom GL module?
By extension, why try to reinvent solutions for scaling problems when Erlang already does a bunch of the work for you?
There have been a lot of speculation and analysis about the challenges that this type of site will face but I think twitter's problems are much simpler. It was built by a group of people who had never scaled a site to that level before. They've been trying to cope with their success but it sounds like they don't have the right people to write scalable software and don't have enough operations people to keep their systems up.
I attended Railsconf 2007 and two of twitter's developers held an impromptu session to discuss the scaling challenges they faced. Twitter had just experienced several days of downtime and the room was filled with people interested in hearing all about it. It was pretty obvious from that one hour session that their issues are more than just technical.
At one point they showed some code that was used to balance some of their load among a group of servers. Unfortunately they wrote the function in such a way that it was limited to a maximum of 10 servers. Once they exceed the capacity of those 10 servers the whole site came crashing down. They seemed like nice guys and were excited about their work but it was obvious based on their comments and the crowd's reaction that they were out of their league. I got the impression that they are doing whatever it takes to keep the site up and running today, even if it brings the site down tomorrow.
25 comments
[ 3.3 ms ] story [ 57.8 ms ] thread- Re-writing their whole system incrementally will be expensive; I think they might need google money and brains to complete the whole thing... or a lot more than 15mm.
But we've had e-mail working fine for years, and Twitter is just e-mail.
I'll leave the armchair architecting to others.
http://www.hueniverse.com/hueniverse/2008/03/on-scaling-a-mi...
I have a Jabber bot running that is hooked into the entire public timeline and it continues to run even when the main website has gone down. I don't think you can directly compare Twitter's issues with issues of Meebo, AIM, GTalk, etc.
People who read sites like TechCrunch most definitely know what Twitter is, and coverage that talks about why they have scaling problems as opposed to making a soap opera about it is very useful for that demographic.
But people like to act like Twitter owes them something. Fact is, it's a free service.
In the spirit of building their business they clearly want -- in fact they need -- to build the best product they can; the fact that it's taking too long by our estimation doesn't matter. They're not going to leave money on the table so you can believe they're working hard to fix the problem and start working on their revenue model.
Kudos to the Twitter team for returning critical volleys from some in the tech community without losing their cool.
Erlang makes distribution and concurrent execution semantically identical, Scala doesn't. And there doesn't seem to be a good reason to base on the JVM, they aren't already next deep in Java libraries.
I'd really like to know what they're thinking over there.
The mindset for selecting tools to build apps which need to scale seems be similar to the one we all had back in early dos days, when the company programmers were having to write all the internal software. We were either unaware of better solutions or we simply felt compelled to write all the stuff ourselves, from spreadsheets to a GL module. Now, with all the customizable apps around, why would anyone want to write a custom GL module?
By extension, why try to reinvent solutions for scaling problems when Erlang already does a bunch of the work for you?
I attended Railsconf 2007 and two of twitter's developers held an impromptu session to discuss the scaling challenges they faced. Twitter had just experienced several days of downtime and the room was filled with people interested in hearing all about it. It was pretty obvious from that one hour session that their issues are more than just technical.
At one point they showed some code that was used to balance some of their load among a group of servers. Unfortunately they wrote the function in such a way that it was limited to a maximum of 10 servers. Once they exceed the capacity of those 10 servers the whole site came crashing down. They seemed like nice guys and were excited about their work but it was obvious based on their comments and the crowd's reaction that they were out of their league. I got the impression that they are doing whatever it takes to keep the site up and running today, even if it brings the site down tomorrow.
Slides from the session I referenced: http://www.slideshare.net/al3x/scaling-twitter-railsconf-200...