Is the count correct?

[ioquatix]

connection_pool/lib/connection_pool.rb

Line 122 in 428c06f

::Thread.current[@key] = @available.pop(options[:timeout] || @timeout)

It seems to me, we increment the count, and then perform an operation to get a connection, which may time out.

It seems the invariant of checkout can be violated, i.e. it increments the count, even though it fails to check out a connection.

This is actually, somewhat suprisingly, handled correctly in with because the timeout interrupt is deferred until within the begin block which has ensure checkout which correctly decrements the count, even in the case of an exception. But during that time, the actual count seems incorrect.

It would appear this makes it tricky/unsafe to use checkout and checkin in general, since the count invariant can be violated.

[mperham]

Yeah, I can see some shady edge cases. I wonder if we can write some tests which surface the bad behavior.

[ioquatix]

I tried to write some but ran out of time.

I think anything that calls checkout with a timeout can fail, i.e. it's very hard to write safe code that uses checkout and checkin directly.