Re: isolation tests broken for other than 'read committed'

Excerpts from Kevin Grittner's message of vie jul 15 18:23:10 -0400 2011:
> It's been a few days since I've run through my usual builds and
> tests, and I just discovered that part of my routine was broken by
> this commit:
>
> http://git.postgresql.org/gitweb/?p=postgresql.git;a=commitdiff;h=846af54dd5a77dc02feeb5e34283608012cfb217

Sorry 'bout that.

> The isolation tests are broken when run against a database with
> default_transaction_isolation = 'repeatable read' or 'serializable'.
> (Which is ironic, really.)
>
> Adding the attached files to src/test/isolation/expected/ causes
> those stricter isolation levels to work in my tests so far, but I
> get random failures in 'read committed' due to apparent randomness
> in which process is chosen as the deadlock victim. I seem to
> remember Noah mentioning this and a suggested fix, but the problem
> in manifest in a current checkout of head.

Yeah, the patch I committed from Noah should fix the issues in read
committed. It hadn't crossed my mind that I need to manually set the
level to serializable in order for the tests to be meaningful :-(
Shouldn't we be running the tests three times with the different useful
isolation levels?

> Of course, another approach to this would be to set transaction
> isolation level in the new tests. If we do that, we might want to
> create tests at all three levels, for completeness.

I think your approach of adding alternative expected outputs makes more
sense.

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