WAL functions cannot be executed in standby servers

[fmbiete]

WAL functions cannot be executed in standby servers

ERROR:  recovery is in progress
HINT:  WAL control functions cannot be executed during recovery.

In PostgreSQL 16 replication slots persist in the standby servers.
This is also the case when using extensions like pg_failover_slots that transfer the slot information to the standby.

This condition will make those queries to return values only in the primary node avoiding the errors.

[elpavel]

What about using this logic when pg_is_in_recovery ?

 		{
 			semver.MustParseRange(">=9.4.0 <10.0.0"),
 			`
-			SELECT slot_name, database, active, pg_xlog_location_diff(pg_current_xlog_location(), restart_lsn)
+			SELECT slot_name, database, active,
+				(case pg_is_in_recovery() when 't' then pg_xlog_location_diff(pg_last_xlog_receive_location(), restart_lsn) else pg_xlog_location_diff(pg_current_xlog_location(), restart_lsn) end) as pg_xlog_location_diff
 			FROM pg_replication_slots
 			`,
 		},
 		{
 			semver.MustParseRange(">=10.0.0"),
 			`
-			SELECT slot_name, database, active, pg_wal_lsn_diff(pg_current_wal_lsn(), restart_lsn)
+			SELECT slot_name, database, active,
+				(case pg_is_in_recovery() when 't' then pg_wal_lsn_diff(pg_last_wal_receive_lsn(), restart_lsn) else pg_wal_lsn_diff(pg_current_wal_lsn(), restart_lsn) end) as pg_wal_lsn_diff
 			FROM pg_replication_slots
 			`,
 		},

[fmbiete]

What about using this logic when pg_is_in_recovery ?

That doesn't throw an error but there are some buts.

• pg_last_wal_receive_lsn() can returns NULL, so we would need to make the statement resilient to it

• I'm not sure of the correctness of that data, it shows 2 different things depending of the server status

  • in the case of being in recovery it would show the difference between the last lsn sent by the primary to this standby and the received lsn of the remote standby
  • in case of not being in recover it would show the difference between the current position and the received lsn of the remote standby

The value in the case of being in recovery could be quite off, compared to the normal value.

What do you think?

[elpavel]

I am not sure what is the best solution here, but anything would be better than the original failing query.

[mbanck-ntap]

In any case, the patch proposed here leads to no metrics being sent back from the standby at all, is that what we want?

[fmbiete]

Yes, that query should only be executed in the primary.