Re: wrong fds used for refilenodes after pg_upgrade relfilenode changes Reply-To:

On Sun, May 8, 2022 at 7:30 PM Thomas Munro <thomas.munro@gmail.com> wrote:
> Simple idea: how about logging the PID of processes that block
> progress for too long?  In the attached, I arbitrarily picked 5
> seconds as the wait time between LOG messages.  Also, DEBUG1 messages
> let you see the processing speed on eg build farm animals.  Thoughts?
>
> To test this, kill -STOP a random backend, and then try an ALTER
> DATABASE SET TABLESPACE in another backend.  Example output:
>
> DEBUG:  waiting for all backends to process ProcSignalBarrier generation 1
> LOG:  still waiting for pid 1651417 to accept ProcSignalBarrier
> STATEMENT:  alter database mydb set tablespace ts1;
> LOG:  still waiting for pid 1651417 to accept ProcSignalBarrier
> STATEMENT:  alter database mydb set tablespace ts1;
> LOG:  still waiting for pid 1651417 to accept ProcSignalBarrier
> STATEMENT:  alter database mydb set tablespace ts1;
> LOG:  still waiting for pid 1651417 to accept ProcSignalBarrier
> STATEMENT:  alter database mydb set tablespace ts1;

This is a very good idea.

> Another thought is that it might be nice to be able to test with a
> dummy PSB that doesn't actually do anything.  You could use it to see
> how fast your system processes it, while doing various other things,
> and to find/debug problems in other code that fails to handle
> interrupts correctly.  Here's an attempt at that.  I guess it could go
> into a src/test/modules/something instead of core, but on the other
> hand the PSB itself has to be in core anyway, so maybe not.  Thoughts?
>  No documentation yet, just seeing if people think this is worth
> having... better names/ideas welcome.

I did this at one point, but I wasn't convinced it was going to find
enough bugs to be worth committing. It's OK if you're convinced of
things that didn't convince me, though.

-- 
Robert Haas
EDB: http://www.enterprisedb.com