On Sat, May 04, 2019 at 10:35:23AM +0200, Julien Rouhaud wrote: > While reading vacuumdb code, I just noticed that it can return 0 if an > error happen when -j is used, if errors happen on the last batch of > commands.
Yes, I agree that this is wrong. GetIdleSlot() is much more careful about that than vacuum_one_database(), so your patch looks good at quick glance. > This behavior exists since 9.5. Trivial patch attached. I'm not sure > that a TAP test is required here, so I didn't add one. I'll be happy > to do so though if needed. You could make that reliable by getting a lock on a table using a two-phase transaction, and your test case from upthread won't fly high as we have no facility in PostgresNode.pm to keep around a session's state using psql. FWIW, I am not convinced that it is a case worth bothering, so no tests is fine. -- Michael
signature.asc
Description: PGP signature
