emojitrack-nodestreamer

First stab at experimental standalone replacement for emojitrack SSE streaming web services, using node evented methods rather than MRI ruby with threads. While the Ruby version was written in spaghetti-land while I had no idea what I was doing, this version was written from scratch after understanding the problem domain well. Thus, it's significantly better structured, has some unit tests, and should be better code in general overall.

This should be API compatible with the MRI version. (In fact, we're load balancing 50% of clients to each right now to test in production.) For a full explanation of the responsibilities of this system and it's API, see the README for mroth/emojitrack-streamer-spec.

The first stab at this was actually slower (albeit most consistent), so now experimenting with cluster workers to try to take advantage of multiple CPU cores to match ruby speed...

Feature differences from Ruby version

No handling of downgraded connection-close if needed, requires a fully working routing environment (should be good only with new routing layer on Heroku and labs:websockets enabled)
No handling of interactive kiosk-mode streams, left out for now for time savings since they aren't in active use until we have another emojitracker physical installation.
No implementation of logging to graphite yet (again, waiting until we know we want to use this as the primary streamer in production.)

Discussion

The first stab at this is in server-sse.coffee.
There is a revised multi-process version in server-cluster.coffee.
It attempts to distribute the load of clients across numCpus-1 forked processes, while preserving a core for the master process to handle reading from Redis and doing the ScorePacker rollup messages.

It's unclear why this isn't more significantly faster than the naive MRI implementation. This is almost pure IO and event handling, which NodeJS should excel at. There is some low hanging fruit in terms of redundant String operations, but nothing that should account for much CPU usage (and fixing those things proved to show no real difference in benchmarking, so I left them out to try to find the real critical paths.)

Benchmarks coming soon.

alexxnica / emojitrack-nodestreamer Goto Github PK