deadlock due to apparent interaction with sleep

[jhoblitt]

As a disclaimer, I've never worked with celluloid/celluloid-io before today but I can't find any mention of sleep being considered unsafe inside of an actor. Everything was going well with my experiments until I started encountering strange deadlocks. I managed to reduced the problem down to this minimal test case which seems to show some sort of odd interaction between sleep/timers running in the same method as a redis method invocation, but only if the redis method is called before the sleep.

#!/usr/bin/env ruby

require 'celluloid/io'
require 'celluloid/redis'
require 'redis/connection/celluloid'

class RedisTest
  include Celluloid::IO

  def initialize
    @redis = ::Redis.new
  end

  def lookup_redis
    puts "redis lookup"
    @redis.lindex 'fooq', 0
    puts "redis returned\n\n"
  end

  def yawn
    puts "going to sleep"
    sleep 1
    puts "wokeup\n\n"
  end

  def yawn_and_lookup_redis
    puts "going to sleep"
    sleep 1
    puts "wokeup"
    puts "redis lookup"
    record = @redis.lindex 'fooq', 0
    puts "redis returned\n\n"
  end

  def lookup_redis_and_yawn
    puts "redis lookup"
    record = @redis.lindex 'fooq', 0
    puts "redis returned"
    puts "going to sleep"
    sleep 1
    puts "wokeup\n\n"
  end

end

rt = RedisTest.new
rt.lookup_redis
rt.yawn
rt.yawn_and_lookup_redis
rt.lookup_redis_and_yawn

The output I get is:

redis lookup
redis returned

going to sleep
wokeup

going to sleep
wokeup
redis lookup
redis returned

redis lookup
redis returned
going to sleep
<hangs>

If I comment out the line require 'redis/connection/celluloid', so that redis uses it's default connection driver, the script exits normally with the expected output.

[tarcieri]

Celluloid provides its own sleep that interacts with its internal timer system.

That said, this is the likely culprit:

celluloid/celluloid-io#56

[jhoblitt]

Well that's a bummer but a bit of relief. I've been having flash backs of debugging pthreads + signals for the last 4 hours.

[tarcieri]

It's something we'd like to get fixed really soon but requires some changes to the design of the Mailbox API which is why it's still not fixed.

@halorgium we should get this fixed really soon now! 🔜