On Tue, Apr 30, 2019 at 7:14 PM Thomas Munro <thomas.mu...@gmail.com> wrote:
> I think it'd be nice to have a set of macros that can create wait
> points in the C code that isolation tests can control, in a special
> build. Perhaps there could be shm hash table of named wait points in
> shared memory; if DEBUG_WAIT_POINT("foo") finds that "foo" is not
> present, it continues, but if it finds an entry it waits for it to go
> away. Then isolation tests could add/remove names and signal a
> condition variable to release waiters.
>
> I contemplated that while working on SKIP LOCKED, which had a bunch of
> weird edge cases that I tested by inserting throw-away wait-point code
> like this:
>
>
> https://www.postgresql.org/message-id/CADLWmXXss83oiYD0pn_SfQfg%2ByNEpPbPvgDb8w6Fh--jScSybA%40mail.gmail.com
>
> Yes, I agree it would be nice to have a framework like this.
Greenplum actually has a fault injection framework that, I believe, works
similarly to what you are describing -- i.e. sets a variable in shared
memory.
There is an extension, gp_inject_fault, which allows you to set the faults.
Andreas Scherbaum wrote a blog post about how to use it [1].
The Greenplum implementation is not documented particularly well in the
code,
but, it is something that folks working on Greenplum have talked about
modifying
and proposing to Postgres.
[1]
http://engineering.pivotal.io/post/testing_greenplum_database_using_fault_injection/
--
Melanie Plageman
