Hi,
On 2021-10-30 11:46:22 -0700, Peter Geoghegan wrote:
> Attached is a draft patch that fixes the problem.
I think it probably is worth adding an error check someplace that verifies
that problems of this kind will be detected with, uh, less effort.
I think it'd also be good to add a test that specifically verifies that
parallel vacuum doesn't have a bug around "parallel worthy" and not "parallel
worthy" indexes. It's too easy a mistake to make, and because visible
corruption is delayed, it's likely that we won't detect such cases.
> Also attached is a second patch. This adds assertions to
> heap_index_delete_tuples() to catch cases where a heap TID in an index
> points to an LP_UNUSED item in the heap (which is what this bug looked
> like, mostly). It also checks for certain more or less equivalent
> inconsistencies: the case where a heap TID in an index points to a
> line pointer that's past the end of the heap page's line pointer
> array, and the case where a heap TID in an index points directly to a
> heap-only tuple.
ISTM that at least a basic version of this is worth doing as a check throwing
an ERROR, rather than an assertion. It's hard to believe this'd be a
significant portion of the cost of heap_index_delete_tuples(), and I think it
would help catch problems a lot earlier.
Greetings,
Andres Freund