This is the very first guest post on MeteorHacks by Brent Abrahams. He talks about how SmartCollections help him to run his app – Standbench – in production. He also talks about how Standbench started and how he scaled it step by step.
This article has been written prior to the release of Meteor 0.7.2. Author has mentioned some performance issues with Meteor Collection at the time article was written. But now it’s all gone and you can use Meteor’s Collection implementation with trust.
I teach mathematics at an international school. Last year, our administrators were looking for online curriculum management software for the organization. I didn’t like any of the existing solutions, and had just discovered Meteor.
It just happened that I had a nine-day holiday coming up. So I wrote an extremely ambitious list of features for such a system and started hacking. To my amazement, I was able to tick off almost all the items in just over a week, leaving me with a pretty impressive prototype on my hands.
Back at school, I pitched the prototype to the administration and they went for it. So I spent two months over the summer writing a role based access control system, an undo-redo stack for all user transactions, and a bunch of widgets that gave me, among other things, one-line-in-a-template editable text, editable lists, routed tabs, and reactive dialogs.
With everything refactored, secure and more feature-complete, we launched “Standbench” in mid-August. Plaudits were rolling in and everyone was very pleased. Until disaster struck…
During a practical teacher training session that created a lot of simultaneous database writes, the server started crashing … hard. We were hosting Standbench at school, and the server couldn’t cope with the massive CPU hit that Standbench was dishing out (having to diff multiple subscriptions across numerous collections, all unique to each user).
So we had a problem.
In an extremely informative interview published in Discover Meteor (Premium Edition only), Nick Martin (Meteor core dev), says “I wrote a benchmark app that very heavily used data, and I got about three concurrent users,” but he also says that apps without much data can serve thousands of concurrent users.
In the case of Standbench, we were crashing at about ten concurrent users, which obviously isn’t acceptable for a production scenario where we could very well have 200+.
The solution to the issue of CPU use is not easy. After scaling vertically as far as possible, you’re still stuck with the fact that a node process runs in a single thread on a single core. And one CPU core can only do so much on its own.
So … I optimized the codebase as much as I knew how, using some principles from this post and others gleaned from that extremely informative interview. And … we were still crashing at ten concurrent users.
Enter Arunoda Susiripala and MeteorHacks. Arunoda had just published his Smart Collections package so, with nothing to lose, I simply switched (almost) all instances of
new Meteor.Collection in my app to
new Meteor.SmartCollection. It took about 2 minutes and everything just worked.
Since then, we haven’t had a single server crash, although an informal load test with 20 concurrent users performing data intensive read actions still pushed the CPU to 100% (without the server failing).
I was not using the Smart Collections MongoDB Oplog integration, so that was the next thing I did, which I understand is a huge factor in making the server load lighter. For good measure, I also scaled horizontally, using HAProxy to load balance across 4 instances of the node process (we have 4 cores available on the server), which presumably gives us close to 4 times the number of concurrent users.
At this stage, it feels like a lot of work to get a complex app deployed to scale on one’s own servers, even when serving a relatively small user base. Deploying to the meteor.com infrastructure is sooo much easier, so I’m really looking forward to Galaxy (the Meteor Development Group’s upcoming PaaS for Meteor apps).
Six months into production, we’re still alive. But I maintain, in all sincerity, that if Arunoda had not published his Smart Collections when he did, Standbench would have been on the shelf by now – beautiful but useless.
I recently tried the oplog tailing that has now been integrated into the Meteor core, but it still lacks the comprehensive support that Smart Collections has for mongo operators so, not surprisingly, I noticed a marked increase in CPU usage. I’ve switched back to Smart Collections for the time being.
There are two issues with Smart Collections:
Because of these issues, I will eventually switch back to the Meteor core Collections.
Despite the challenges with scaling, at present, I don’t think I’ll be building a web app again without using Meteor. The development process was just too sweet. What would otherwise take months-to-years to develop can be put together in a matter of weeks.
In my mind, the cost of having to provision/configure a lot of server resources is worth paying if it makes the development/maintenance process easier. Obviously this trade-off isn’t going to work for every use case.