On Mon, Aug 10, 2015 at 8:18 PM, Michael Paquier <michael.paquier@gmail.com>
wrote:
> But this fixed a problem on master where queries got stuck because of
> the WAL insert lock patch that has been introduced in 9.4 because of a
> race condition between two backends.
>
Interesting, thanks for pointing this out.
Do you have some remnants of the logs when this problem happened? Or
> even better, a stack trace to understand where the startup process got
> stucked? That's perhaps unrelated, but were you using commit_delay?
>
We do have logs from the primary and the replica, but I can't share them
directly. Is there anything specific I should look for? Unfortunately, I
did not think to take a stack trace from the startup process. And
commit_delay is not set.
I have been playing with pgbench, creating and dropping views with a
> couple of hundred connections on 9.4.1 but could not reproduce the
> issue.
>
Thanks for trying it out. Were you running queries against these views on
the replica when dropping/recreating them on the primary? I was guessing
maybe the lock interplay there had something to do with it, although I
admit my understanding of these mechanisms is pretty superficial.