On Tue, Apr 21, 2020 at 02:44:45AM +0300, Alexander Korotkov wrote:
> Among all the joined tables, only "pg_index I" is expected to have
> multiple rows associated with single relation. But we do sum() for
> toast index "pg_index X" as well. As the result, we multiply
> statistics for toast index by the number of relation indexes. This is
> obviously wrong.
Oops.
> As a bugfix, I think this should be backpatched. But this patch
> requires catalog change. Were similar cases there before? If so,
> how did we resolve them?
A backpatch can happen in such cases, see for example b6e39ca9. In
this case, the resolution was done with a backpatch to
system_views.sql and the release notes include an additional note
saying that the fix applies itself only on already-initialized
clusters. For other clusters, it was necessary to apply a SQL query,
given also in the release notes, to fix the issue (just grep for
CVE-2017-7547 in release-9.6.sgml on the REL9_6_STABLE branch).
--
Michael