I’d like to start this post by wishing everybody a wonderful 2010 year, and a great 2nd decade of this millennium.
We will remember these early days are pretty hectic, here at Superfeedr. As a matter of facts our users keep feeding us with more and more feeds (+20% in a week!). It’s great, except that it makes little issues much bigger. And we’re currently fighting hard (like 16hours a day) to improve our scheduling algorithm.
The difficulty that we have is that we need to deal with a lot of pings, telling us that a given feed has been updated. Technically, at the moment when we receive a ping, we need to fetch the feed and push that to our subscribers, which means that we’re already late to deal with this feed.
So, at any time, we know how many feeds are late. Also, we need to fetch all of our feeds at least every 15 minutes, which roughly tells us how late we are on average for each feed.
Right now, we’re at about 10% of late feeds, which means that, on average, we’re 1.5 minutes late on pings. Yes, it’s much higher than acceptable and we’re working on that – but it reached 50% earlier today :( -
Contrary to many services out there, Superfeedr has to scale “internally”. Most of web apps and web services usually work like this :
- A user sends a query
- the application processes that query
- the application returns the response.
It’s not always easy to scale all the time, but at least it’s a fairly known science. Adding cache to skip the 2nd step is the most obvious thing and it can do miracles :)
Unfortunately for us, that is not how Superfeedr works. Of course, like I mentioned we deal with incoming pings, but we also have to deal with polling for a lot of feeds (even the pinged feeds, as a failover), which means that as opposed to almost every web service out there, we do not act upon user input. The good news is that we can probably have a bigger query → response time. The bad news is that caching1 is useless!
As a matter of fact, our biggest problem is to distribute all our data-processing smoothly and evenly across our whole infrastructure so that we always have a capacity utilization of 100%. And that’s what I’ll be spending the next hours and days on :)
1 And caching is evil for real-time stuff :)