Re: 8.4.0 data loss / HOT-related bug

Andrew Gierth <andrew@tao11.riddles.org.uk> writes:
> "Greg" == Greg Stark <gsstark@mit.edu> writes:
>  Greg> Either of two things are true.
>  Greg> Either transaction 6179 committed,
> [snip]

> This is all missing the point. The row should have been killed by
> transaction 4971, NOT 6179. By the time transaction 6179 tried to
> do anything with it, it was almost certainly already broken (or
> possibly 6179 broke it).

If there have been any crashes in this database, the problem might be
a variant of Jeff Ross' issue --- to wit, the row killed by
transaction 4971 actually is dead (and its associated toast items
have been removed), but that row is still found by seqscans because
of an incorrect PD_ALL_VISIBLE flag.

            regards, tom lane