Details
-
Improvement
-
Status: Closed
-
Major
-
Resolution: Later
-
3.2.M1
-
None
-
None
Description
Currently when all the DB connections in the pool are used up threads will wait for available connection. As connections become available they are given to essentially random threads, not the thread that was waiting first. This causes somewhat erratic behavior since some requests will come in and get a connection right away even though others have been waiting for a long time, eventually causing more threads to throw exceptions than may have been necessary if requests were processed in order.
Particularly in cases with large but brief spikes in activity - errors may be wholly avoided if requests are processed in order where otherwise they could produce many failures. I have observed this in my testing.
I have a patch I am testing out. It creates a queue to keep track of the waiting threads and then grant them connections in order.