Re: pgstat_read_statsfiles() and reset timestamp

Fujii Masao <masao(dot)fujii(at)oss(dot)nttdata(dot)com> writes:
> pgstat_read_statsfiles() sets each stat_reset_timestamp to
> the current timestamp, at the beginning of the function,
> just in case we fail to load the existing statsfile. This code is
> originally introduced by commit 4c468b37a2.

> But commit ad1b5c842b changed pgstat_read_statsfiles() so that
> the stats including stat_reset_timestamp are zeroed in that case,
> so now there seems no need to set each stat_reset_timestamp.

Huh? The zeroing happens before those fields are filled.

> Attached is the patch that removes such unnecessary sets of
> stat_reset_timestamp from pgstat_read_statsfiles().

-1, minus a lot actually. What this will do is that if there's
no stats file, the reset timestamps will all read as whatever
our epoch timestamp is (2000-01-01, I think). This is not a
corner case, either --- it's the expected path at first start.
We want current time to be used in that case.

If there are any code paths in pgstat_read_statsfiles that
re-zero these structs later, they need to be fixed to restore
the reset timestamps to these values, as well.

regards, tom lane