|
Doc Text:
|
Cause: Under certain conditions, with a mix of concurrent search and update and outgoing replication operations, there will be deadlocks in the changelog db, leading to error messages like this:
NSMMReplicationPlugin - changelog program - _cl5WriteOperationTxn: failed to write entry with csn (XXXXXXX); db error - -30994 DB_LOCK_DEADLOCK: Locker killed to resolve a deadlock
This is caused by a deadlock between the changelog readers, writers, and main database writers.
Consequence: Update operations will fail with the above error message in the directory server errors log.
Fix: A new configuration parameter is introduced:
dn: cn=config,cn=ldbm database,cn=plugins,cn=config
nsslapd-db-deadlock-policy: 9
With the default policy 9 (DB_LOCK_YOUNGEST), the last locker gets killed when there is a deadlock. In the case that this is the changelog writer, the write will fail, and the entire update will fail.
Users who frequently see the above errors in the errors log are advised to change this setting to 6 (DB_LOCK_MINWRITE) will which instead kill the locker that has the fewest write locks (that is, the changelog reader). The changelog reader code has been changed to handle this deadlock condition and retry. The setting can be changed like this:
ldapmodify -x -D "cn=directory manager" -W <<EOF
dn: cn=config,cn=ldbm database,cn=plugins,cn=config
changetype: modify
replace: nsslapd-db-deadlock-policy
nsslapd-db-deadlock-policy: 6
EOF
You may ask why the default is not changed to 6. The answer is that the setting will apply to _all_ threads, so that changing this setting could cause regular search requests to fail, if the directory server is under a heavy update load. In our testing, we did not see this happen, but we cannot guarantee that changing this value to 6 will not impact regular search requests.
Result: After changing nsslapd-db-deadlock-policy to 6, updates will succeed and no longer cause errors like the above.
|