Re: All-zero page in GIN index causes assertion failure

Heikki Linnakangas wrote:

> Looking closer, heap vacuum does a similar thing, but there are two
> mitigating factors that make it safe in practice, I think:
>
> 1. An empty heap page is all-zeroes except for the small page header in the
> beginning of the page. For a torn page to matter, the page would need to be
> torn in the header, but we rely elsewhere (control file) anyway that a
> 512-byte sector update is atomic, so that shouldn't happen. Note that this
> hinges on the fact that there is no special area on heap pages, so you
> cannot rely on this for index pages.
>
> 2. The redo of the first insert/update on a heap page will always
> re-initialize the page, even when full-page-writes are disabled. This is the
> XLOG_HEAP_INIT_PAGE optimization. So it's not just an optimization, it's
> required for correctness.

Hmm, I guess this is a case of an optimization hiding a bug :-( I mean,
if we didn't have that, we would have noticed this a lot sooner, I
imagine.

> BRIN update is not quite right, however. brin_getinsertbuffer() can
> initialize a page, but the caller might bail out without using the page and
> WAL-logging the change. If that happens, the next update that uses the same
> page will WAL-log the change but it will not use the XLOG_BRIN_INIT_PAGE
> option, and redo will not initialize the page. Redo will fail.

Hmm, interesting.

> BTW, shouldn't there be a step in BRIN vacuum that scans all the BRIN pages?
> If an empty page is missing from the FSM for any reason, there's nothing to
> add it there.

Probably, yeah, makes sense. If you recall, the original design I
proposed had a scan of the whole index (though for other reasons I am
thankful we got rid of that).

> This is all very subtle.

No kidding ...

--
Álvaro Herrera http://www.2ndQuadrant.com/
PostgreSQL Development, 24x7 Support, Remote DBA, Training & Services