Avoid race condition for event_triggers regress test - Mailing list pgsql-hackers

From Mikhail Gribkov
Subject Avoid race condition for event_triggers regress test
Date
Msg-id CAMEv5_sS5G7K1PCV2oBx5+1NF1ZneJ6C5Z_xV_SWsZMukdFZiA@mail.gmail.com
Whole thread Raw
Responses Re: Avoid race condition for event_triggers regress test
List pgsql-hackers
Hi hackers,

After committing the on-login trigger (e83d1b0c40ccda8955f1245087f0697652c4df86) the event_trigger regress test became sensible to any other parallel tests, not only DDL. Thus it should be placed in a separate parallel schedule group.

The current problem is that a race condition may occur on some systems, when oidjoins test starts a moment later than normally and affects logins count for on-login trigger test. The problem is quite a rare one and I only faced it once. But rare or not - the problem is a problem and it should be addressed.

Such race condition can be simulated by adding "select pg_sleep(2);" and "\c" at the very beginning of oidjoins.sql and adding "select pg_sleep(5);" after creation of the login trigger in event_trigger.sql.
The resulting symptoms are quite recognizable: regression.diffs file will contain unexpected welcome message for oidjoins test and unexpectedly increased result of "SELECT COUNT(*) FROM user_logins;" for event_triggers test. (These are accompanied with the expected responses to the newly added commands of course)

To get rid of the unexpected results the oidjoins and event_triggers tests should be splitted into separate paralell schedule groups. This is exactly what the proposed (attached) patch is doing.

What do you think?
--
 best regards,
    Mikhail A. Gribkov

e-mail: youzhick@gmail.com

Attachment

pgsql-hackers by date:

Previous
From: Thomas Munro
Date:
Subject: Re: Guiding principle for dropping LLVM versions?
Next
From: Stephen Frost
Date:
Subject: Re: Parent/child context relation in pg_get_backend_memory_contexts()