Fix race in autovacuum freeze score calculations.

We calculate the transaction ID and multixact ID component scores
by subtracting a table's relfrozenxid or relminmxid from a "recent"
ID that we read before scanning pg_class.  If something advances
relfrozenxid or relminmxid past the recent ID in the meantime
(e.g., a concurrent VACUUM), the subtraction wraps around and gives
the table a much higher score than it deserves.  The window is
small enough that this is probably rare in practice.  To fix, set
the affected component score to 0 in this case.

Oversight in commit d7965d65fc.

Author: Baji Shaik <[email protected]>
Reviewed-by: Bharath Rupireddy <[email protected]>
Discussion: 
https://postgr.es/m/CA%2Bfm-RPOMc-GHjpv4DpXO1eqMahCeZDpAgXWqVgnsB3s3arptA%40mail.gmail.com
Backpatch-through: 19

Branch
------
REL_19_STABLE

Details
-------
https://git.postgresql.org/pg/commitdiff/b1bc8f8ea5daf86cb5669d62f08c03d05caeb0de

Modified Files
--------------
src/backend/postmaster/autovacuum.c | 17 ++++++++++++++---
1 file changed, 14 insertions(+), 3 deletions(-)

Reply via email to