Bug #76586 Lock during simultaneous disconnect and connect to cluter
Submitted: 3 Apr 2015 11:33 Modified: 28 Jul 2015 14:40
Reporter: Mike Pentagra Email Updates:
Status: Verified Impact on me:
None 
Category:MySQL Cluster: NDB API Severity:S2 (Serious)
Version:7.3.7 OS:Linux
Assigned to: CPU Architecture:Any

[3 Apr 2015 11:33] Mike Pentagra
Description:
Looks like we faced a problem with simultaneous multithreaded creating and destroying connections to cluster.
Stack traces of all threads related to NDB API attached.
Process spent several hours in this state. So it's not a temporal timeout, it's a hang.

How to repeat:
Create and destroy Ndb instances in several threads.
[3 Apr 2015 11:34] Mike Pentagra
Stack traces of threads that contain NDB API

Attachment: stacks.txt (text/plain), 128.74 KiB.

[7 Jul 2015 11:33] Mikael Ronström
Looked at the stack trace and it looks very similar to a bug we just fixed where you
had to change do_poll(0) to do_poll(10) in the close_clnt-method. In this bug however
the management server was the problem and went to 100% CPU usage.

From looking at the trace I see 60 threads blocked on the mutex and one thread that actually
does proceed and doesn't seem stuck.

I think this bug fix is coming in 7.3.10.
[28 Jul 2015 14:43] Bogdan Kecman
Hi Mike,

Thanks for your report.
I verified this bug on 7.3.9.
I did not manage to verify on 7.3.10 so it seems it is solved in 7.3.10 but I can't confirm 100% as I don't have reliable way of reproducing the problem.

all best
Bogdan Kecman