random isolation test failures

Andrew Dunstan <andrew@dunslane.net> wrote:

We are seeing numerous occasional buildfarm failures of the
fk-deadlock2 isolation test

If this is harmless, we could provide an alternative results file
as a simple fix. If it's not harmless, it should be fixed.

I agree, but don't look at me. I'm not the one who added the tests,
nor are they related to serializable snapshot isolation. Tom
recently raised the same issue on this thread:

http://archives.postgresql.org/pgsql-hackers/2011-09/msg00991.php

Alvaro?

-Kevin

Andrew Dunstan <andrew@dunslane.net> writes:

We are seeing numerous occasional buildfarm failures of the fk-deadlock2
isolation test,

Yeah, I complained about this already, but Kevin disclaims all
responsibility for the fk isolation tests. It looks like Alvaro
and Noah Misch are the people to be harassing.

regards, tom lane

On Mon, Sep 26, 2011 at 01:10:27PM -0400, Tom Lane wrote:

Andrew Dunstan <andrew@dunslane.net> writes:

We are seeing numerous occasional buildfarm failures of the fk-deadlock2
isolation test,

Yeah, I complained about this already, but Kevin disclaims all
responsibility for the fk isolation tests. It looks like Alvaro
and Noah Misch are the people to be harassing.

Yep; I took advantage of Kevin's test harness for some unrelated tests.

These sporadic failures happen whenever the test case takes longer than
deadlock_timeout (currently 100ms for these tests) to setup the deadlock. I
outlined some mitigating strategies here:
http://archives.postgresql.org/message-id/20110727171438.GE18910@tornado.leadboat.com

I'd vote for #1: let's double the deadlock_timeout until the failures stop.
Other opinions?

Thanks,
nm

Excerpts from Noah Misch's message of lun sep 26 21:57:40 -0300 2011:

These sporadic failures happen whenever the test case takes longer than
deadlock_timeout (currently 100ms for these tests) to setup the deadlock. I
outlined some mitigating strategies here:
http://archives.postgresql.org/message-id/20110727171438.GE18910@tornado.leadboat.com

I'd vote for #1: let's double the deadlock_timeout until the failures stop.
Other opinions?

I just tweaked isolationtester so that it collects the error messages
and displays them all together at the end of the test. After seeing it
run, I didn't like it -- I think I prefer something more local, so that
in the only case where we call try_complete_step twice in the loop, we
report any errors in either. AFAICS this would make both expected cases
behave identically in test output. The only thing left to figure out is
where to store the error message between calls ... clearly Step is not
the right place for it. I'm on it now, anyway.

--
Álvaro Herrera <alvherre@commandprompt.com>
The PostgreSQL Company - Command Prompt, Inc.
PostgreSQL Replication, Consulting, Custom Development, 24x7 support

Alvaro Herrera <alvherre@commandprompt.com> writes:

I just tweaked isolationtester so that it collects the error messages
and displays them all together at the end of the test. After seeing it
run, I didn't like it -- I think I prefer something more local, so that
in the only case where we call try_complete_step twice in the loop, we
report any errors in either. AFAICS this would make both expected cases
behave identically in test output.

Hmm, is that really an appropriate fix? I'm worried that it might mask
event-ordering differences that actually are significant.

regards, tom lane

Excerpts from Tom Lane's message of mar sep 27 01:11:39 -0300 2011:

Alvaro Herrera <alvherre@commandprompt.com> writes:

I just tweaked isolationtester so that it collects the error messages
and displays them all together at the end of the test. After seeing it
run, I didn't like it -- I think I prefer something more local, so that
in the only case where we call try_complete_step twice in the loop, we
report any errors in either. AFAICS this would make both expected cases
behave identically in test output.

Hmm, is that really an appropriate fix? I'm worried that it might mask
event-ordering differences that actually are significant.

In the attached, it only affects the case where there is one blocking
command and another command that unblocks it; this is only exercised by
the much-beaten fk-deadlock cases. If either of the steps fails with a
deadlock error, it is reported identically, i.e. the error message is
emitted as

"error in s1u1 s2u1: ERROR: deadlock detected"

So the deadlock could have been detected in either s1u1 or s2u1; we
don't really care.

The way error messages are reported in all the other cases is not
changed, and these do not have a prefix; so if anything were to behave
differently, we would find out because a spurious prefix would appear.

--
Álvaro Herrera <alvherre@commandprompt.com>
The PostgreSQL Company - Command Prompt, Inc.
PostgreSQL Replication, Consulting, Custom Development, 24x7 support

Alvaro Herrera <alvherre@commandprompt.com> writes:

Excerpts from Tom Lane's message of mar sep 27 01:11:39 -0300 2011:

Hmm, is that really an appropriate fix? I'm worried that it might mask
event-ordering differences that actually are significant.

In the attached, it only affects the case where there is one blocking
command and another command that unblocks it; this is only exercised by
the much-beaten fk-deadlock cases. If either of the steps fails with a
deadlock error, it is reported identically, i.e. the error message is
emitted as
"error in s1u1 s2u1: ERROR: deadlock detected"
So the deadlock could have been detected in either s1u1 or s2u1; we
don't really care.

Hmm. For the case of "deadlock detected", we actually don't *want* to
care because the infrastructure is such that either process might report
it. So I agree that this is a good fix for that case. I'm just worried
whether it will obscure other situations where it's important to know
which command failed. But if you're convinced there aren't any, fine.

regards, tom lane