... with a blinding flash ...
The VACUUM funnies I was complaining about before may or may not be real
bugs, but they are not what's biting Alfred. None of them can lead to
the observed crashes AFAICT.
What's biting Alfred is the code that moves a tuple update chain, lines
1541 ff in REL7_0_PATCHES. This sets up a pointer to a source tuple in
"tuple". Then it gets the destination page it plans to move the tuple
to, and applies vc_vacpage to that page if it hasn't been done already.
But when we're moving a tuple chain, *it is possible for the destination
page to be the same as the source page*. Since vc_vacpage applies
PageRepairFragmentation, all the live tuples on the page may get moved.
Afterwards, tuple.t_data is out of date and pointing at some random
chunk of some other tuple. The subsequent copy of the tuple copies
garbage, which explains Alfred's several crashes in constructing index
entries for the copied tuple (all of which bombed out from the
index-build calls at lines 1634 ff, ie, for tuples being moved as part
of a chain). Once in a while, the obsolete pointer will be pointing at
the real header of a different tuple --- perhaps even the place where we
are about to put the copy. This improbable case explains the one
observed Assert crash in which a copied tuple's HEAP_MOVED_IN bit
mysteriously got turned off. Reason: it was cleared through the
old-tuple pointer just after being set via the new-tuple one.
Proof that this is happening can be seen in the core dumps for Alfred's
index-construction-crash cases: tuple.t_data does not point at the same
place that the tuple.ip_posid'th page line item points at. This could
only happen if the page was reshuffled since the tuple pointer was set
up. The explanation for the Assert crash is a bit of a leap of faith,
but I feel confident that it's right.
The solution is to do everything we're going to do with the source
tuple, especially copying it and updating its state, *before* we apply
vc_vacpage to the destination page. Then we don't care if the source
gets moved during vc_vacpage.
I will prepare a patch along this line and send it to Alfred for
testing.
regards, tom lane