Calling Redis from multiple fibers causes fibers to suspend indefinitely

felixyz commented 11 years ago

The bug is illustrated here: https://gist.github.com/Felixyz/5771982

And a repo containing the same code and branches with variations here: https://github.com/Felixyz/celluloid-redis-bug

Reading from a Redis server using the redis gem with the Celluloid driver from multiple tasks in the same actor causes fibers to suspend and never be resumed, typically after 2 resumes (but in production code it's usually taken longer for this to happen, presumably because other processing intervenes, making the intervals between thread/fiber switching longer).
Under JRuby, the same thing happens but consistently on all the fibers – not a single one is resumed. Update: This is not true. It's the behavior I saw earlier, but running the script repeatedly with JRuby yields a variety of results. At most, I've seen 6 out of 10 tasks resumed. The main point is that the bug appears similarly on MRI and JRuby (1.7.4).
All the commands are received by the Redis server when on MRI, but not on JRuby, as illustrated by the gist.
Using futures causes the same bug.
Reading from separate actors (10 tasks 10 actors, as opposed to 10 tasks 1 actor) does not cause the bug.
Sleeping or reading from a socket instead of calling the redis gem does not cause the bug. (This is illustrated in branches in the repo linked to above.)

felixyz commented 11 years ago

I've run the tests on OS X 10.8.2.

felixyz commented 11 years ago

I discussed this with @tarcieri on #celluloid and he pointed out that Redis can't handle interleaved requests on the same connection, so I guess this should just be closed.

Perhaps it might be interesting to add functionality to this gem to create an internal pool of connections? I guess it would have to be implicitly one pool per actor.

tarcieri commented 11 years ago

Another interesting approach would be a sort of "pipeline manager" that gives out Futures for requested commands, sends them to the server in a pipelined manner, and can match up these futures to Redis's responses. See:

http://redis.io/topics/pipelining

celluloid / celluloid-redis

Calling Redis from multiple fibers causes fibers to suspend indefinitely #9