Twiddla has been getting a ton of attention this week. We picked up the
Technical Achievement award at SXSW Interactive, and have been getting a bunch of good press ever since. 25,000 people have signed up for the service since the award was mentioned, with 7,500 of those signups happening in a single day. It's about to get good.
For me though, it's been even better. We're finally getting enough traffic to start thinking about scaling issues. You might remember an article that I wrote a few months back, where I told people
not to sweat Performance and Scaling issues too much, but rather to focus on Readability, Debugability, Maintainability, and Development Pace. The idea was that getting your product to market quickly and being able to move fast if necessary are more important than having the Perfect Dream System that takes forever to build. Of course, the implied point was that when and if that Big Day came, you'd be able to move fast enough to deal with Scalability and Performance concerns as they appeared.
On March 12th, 2008, I got to see first hand whether I was talking out my arse…
3/11/2008 7:00pm: 150 signups/hr, 50 hits/sec, 0-5% CPU
It's the day after the awards, and the first brief announcements are out. Traffic has been building steadily all day, but we've seen worse. The only crisis at the moment is that we don't yet have a
Press Kit, so we're seeing writeups with the old logo and screenshots from the old UI. D'oh!
3/11/2008 11:00pm: 350 signups/hr, 120 hits/sec, 1-9% CPU
Japan wakes up. The Asian press really liked us, so we saw a big spike in users from China and Japan the first few days.
The sandbox is pretty clogged, and with 30 people drawing simultaneously it's starting to tax people's browsers. Every once in a while, somebody navigates the sandbox over to a porn site, and people write our support line to complain. We're wiping the sandbox every 5 minutes, but it's still not acceptable. Gotta get a handle on that.
3/12/2008 9:00am: 300 signups/hr, 100 hits/sec, 1-6% CPU
The sandbox is completely overloaded. There are 100 people in there, which is too many people communicating at once for any medium to really handle. Imagine 100 people drawing on a real whiteboard at the same time, or 100 people talking over each other on a conference call. It just doesn't work. To bring a little order into the picture, I fire up the Visual Studio.NET and add a little switcher that will direct traffic to any one of 5 sandboxes, each one holding 8 users. Throw that live, and now there are 5 overloaded sandboxes.
3/12/2008 9:30am: 500 signups/hr, 300 hits/sec, 3-15% CPU
I bump up the sandbox count to 10. Then think better of it and bump it up to 20 before pushing. Then think better of THAT and add a new page to show users in case all 20 of those sandboxes fill up. Push that live.
3/12/2008 9:41am
Testing out the above changes, I am immediately redirected to a page saying "Sorry, all the Sandboxes are full." Let me restate that: From the time I pushed those changes live to the time I could test them out, 160 people had beaten me into the sandboxes. Wow.
3/12/2008 10:00am: 700 signups/hr, 500 hits/sec, 5-20% CPU
Looking through the error logs, I'm starting to see our first concurrency issues. These are the little one-in-a-million things that you'd never find in test, but that happen every ten minutes under load. They're mostly low-hanging fruit, so I spend the next hour patching and re-deploying until the error logs go silent.
3/12/2008 12:00pm: 600 signups/hr, 400 hits/sec, 5-17% CPU
I'd been doing all of this from my sister's house up in Ft. Worth, who I had supposedly been visiting for a couple days, but whose house I had been mostly using for an office (thanks Lisa for tolerating that, and I promise to get out and visit sometime when I'm not trying to launch a new website!) Now I had to hop in the car and drive back to Austin to fly home. Our trusty server will be on its own for the next 12 hours, taking the beating of its life. I won't even know if it goes down.
3/13/2008 4000 signups/day, 100 hits/sec, 3-10% CPU
Twiddla ArtBack in a stable place, and ready to deal with the flood of feedback emails we've been getting. This part is fun, since most people have nice things to say, and it becomes readily apparent what features everybody wants to see. Nothing has broken, so I actually have some time to put a few minor features live. The "Wite-out" button was added this day, I think, and I re-did the way we handle snapshots and image exporting.
3/14/2008 3000 signups/day, 100 hits/sec, 2-5% CPU
I implemented a fix for the last little concurrency bug that we'd been seeing. Then, while profiling that fix on the server, I noticed that TwiddleBot was flipping out. TwiddleBot is the little service that runs the Guided Tour feature, and is also responsible for clearing out the sandboxes from time to time. Turns out, he was also pounding the database 20 times a second, asking for instructions. Hmm… Chill, TwiddleBot. Pushed a fix for that, and suddenly CPU usage dropped to zero. Like, ZERO! Every 5 seconds, it would spike up to 1%. Cool. I think we're gonna be able to scale this thing…
One week later, ~1000 signups per day, 50 hits/sec, 0% CPU
In the end, we came through our first little scaling event rather well. We were actually a bit over-prepared. Our colocation facility (
Easystreet in Beaverton, Oregon) had a couple extra boxes waiting to go for us, and I had taken the time a week earlier to write up and test a little software load balancer to allocate whiteboard sessions to various boxes when needed. In the end, we didn't get to try any of that out. Hell, we never spiked the processor on our one server over 50%. I'd love to congratulate myself for the design choices I made all those months back when I wrote that article, but I think it's still too early in the game to conclude that we'll really scale when we ramp up to the next level.
Still, it's worth noting that everything in Twiddla was built using the simple, Readable, Debuggable backend that we've been using on our more pedestrian sites for years, and it held up just fine under traffic. When it turned out that parts of that backend needed refactoring to handle the kind of concurrency we saw last week, it was a simple 5 minute task to crack open the code, find what needed to change, and change it.
Readable, Debuggable, Maintainable. That's the plan. Thus far, that has enabled us to keep on top of any Performance and Scalability issues that have come along. With luck, things will continue to work that way!
by
Jason Kester
Discuss on hacker news