Andrew Dunstan <andrew@dunslane.net> writes:
> Just to confirm that Simon is not suffering this uniquely, I saw this
> the other day on Windows, I believe - meant to report it but it got away
> from me.
Oh, I bet I know what's going on --- are you guys launching the
postmaster in a console window and then typing ^C to shut it down?
Depending on your shell that might result in SIGINT being delivered
not only to the postmaster but to all its children. Backends, if
any, will take that as a query-cancel while the bgwriter will take
it as a checkpoint request. Which could easily result in the
too-frequent-checkpoints message.
The only real fix I can see for this is to use some other signal than
SIGINT to transmit checkpoint requests to the bgwriter. Maybe we could
piggyback all bgwriter requests onto SIGUSR2, comparable to the way that
signals up to the postmaster are now handled. Not sure how important
it is though.
regards, tom lane