Masahiko Sawada wrote: > While reading the code, I realized that the requesting an autovacuum > work-item could fail in silence if work-item array is full. So the > users cannot realize that work-item is never performed. > AutoVacuumRequestWork() seems to behave so from the initial > implementation but is there any reason of such behavior? It seems to > me that it can be a problem even now that there is only one kind of > work-item. Attached patch for fixing it.
Hmm, yeah. I think it would be better to return false to caller, and have them report the failure. Then they're in a better position to place an accurate log message -- for instance indicating the relation name rather than just OID, and specify work item type rather than some weird integer. (I think the ereport() should definitely be *outside* the locked section, in any case.) I'm inclined to change the API of AutoVacuumRequestWork even in branch 10. Since this stuff is not nicely extensible (c.f. perform_work_item), it doesn't seem likely that any outside code is calling it yet. Once we have hooks to register worker functions and stuff, it'll become more of a problem. -- Álvaro Herrera https://www.2ndQuadrant.com/ PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services
