Today was day 7 of the “first app”:http://twitterrific.com/ to ship with Tweet Marker support. Overall, things have been working great. I thought I’d write a little bit about my hosting experience so far.

First, the good news: the service is up and fast. Twitterrific is sending a lot of traffic, even with the preference off by default, but the server app is architected in such a way that things are zipping along nicely. Performance is especially important on the iPhone, where a Twitter client might be saving state on quit and any lag would be pretty obvious. Non-trivial server work, such as hitting the Twitter API, is done in the background using Resque, and the queue size there is usually small too.

There have been two distinct problems, though. Each was a good opportunity to improve the system.

On the first day, the background queue stalled and the requests piled up, so much so that my Redis instance ran out of memory and clients started receiving errors. I upgraded Redis, restarted the app, and bumped up the number of processes to quickly catch up on the backlog. Although it hasn’t resulted in errors since, it has delayed some requests on 2 more occasions after launch. I also adjusted how I’m collecting statistics to further improve performance.

More recently, I’ve noticed some sporadic SSL errors when Tweet Marker tries to verify the user account via Twitter. I added more wait-and-retry logic, and I’m keeping an eye on this to kick the background script if necessary. There have been no errors in the last couple days.

I still think it was the right decision to host on Heroku. Although I am manually monitoring the server for general health (need to automate this badly), everything is less work than I would do with a colocated box or VPS. The biggest cost is that at the last minute I upgraded to Heroku’s dedicated PostgreSQL database. I’m still evaluating that, but I would rather launch with too much horsepower and scale it down later, than not enough.

Again, thanks for your support. It is a little nerve-racking because I know that “other people”:http://www.iconfactory.com/ are depending on Tweet Marker being up and performing well. This is a different experience for me than working at a “larger company”:http://www.vitalsource.com/ with a dedicated system administrator, and also different than server backends for my Riverfold products, where it is only my own customers who will be disappointed if I fail.