As discussed [1 ][2] currently, the checkpoint-redo LSN can not be
accurately detected while processing the WAL. Although we have a
checkpoint WAL record containing the exact redo LSN, other WAL records
may be inserted between the checkpoint-redo LSN and the actual
checkpoint record. If we want to stop processing wal exactly at the
checkpoint-redo location then we cannot do that because we would have
already processed some extra records that got added after the redo
LSN.
The patch inserts a special wal record named CHECKPOINT_REDO WAL,
which is located exactly at the checkpoint-redo location. We can
guarantee this record to be exactly at the checkpoint-redo point
because we already hold the exclusive WAL insertion lock while
identifying the checkpoint redo point and can insert this special
record exactly at the same time so that there are no concurrent WAL
insertions.
[1] https://www.postgresql.org/message-id/CA%2BTgmoYOYZfMCyOXFyC-P%2B-mdrZqm5pP2N7S-r0z3_402h9rsA%40mail.gmail.com
[2] https://www.postgresql.org/message-id/20230614194717.jyuw3okxup4cvtbt%40awork3.anarazel.de
--
Regards,
Dilip Kumar
EnterpriseDB: http://www.enterprisedb.com