On 2017-02-28 16:57, Jon Nelson wrote:
> What does pg_xlogdump say about the differences in the files?
What a nice tool. I didn't realize that it exists for 9.6.
Unfortunately, we gave up on shared WAL archive, so I don't if I will
still have all both WALs. I have one conflicting WAL from one of the
servers and there is the *whole* content:
rmgr: Standby len (rec/tot): 24/ 50, tx: 0, lsn:
43/74000028, prev 43/73000140, desc: RUNNING_XACTS nextXid 11617888
latestCompletedXid 11617887 oldestRunningXid 11617888
rmgr: XLOG len (rec/tot): 80/ 106, tx: 0, lsn:
43/74000060, prev 43/74000028, desc: CHECKPOINT_ONLINE redo 43/74000028;
tli 13; prev tli 13; fpw true; xid 0:11617888; oid 25304; multi 1;
offset 0; oldest xid 1750 in DB 13322; oldest multi 1 in DB 1;
oldest/newest commit timestamp xid: 0/0; oldest running xid 11617888; online
rmgr: Standby len (rec/tot): 24/ 50, tx: 0, lsn:
43/740000D0, prev 43/74000060, desc: RUNNING_XACTS nextXid 11617888
latestCompletedXid 11617887 oldestRunningXid 11617888
rmgr: XLOG len (rec/tot): 8/ 34, tx: 0, lsn:
43/74000108, prev 43/740000D0, desc: BACKUP_END 43/74000028
rmgr: XLOG len (rec/tot): 0/ 24, tx: 0, lsn:
43/74000130, prev 43/74000108, desc: SWITCH
I don't have WAL from other server, but I hope I will be able to find it
tomorrow (if I haven't deleted it).
But I can share with you what I have observed by manually looking into
WALs with hex editor:
* If I remember correctly, first page was same (what we see above)
* For one of the logs (the one I have right now), all first 16 pages
except for the first had XLogPageHeaderData set (magic number=0xD093,
tli=13, pageaddr) but WITHOUT any payload. Everything except header was
zero. Remaining part of WAL WAS ALL ZEROED.
* Same WAL from other server was also similar. First page contained data
and remaining part of WAL was filled with XLogPageHeaderData WITHOUT
payload.
* The only different was that one WAL had XLogPageHeaderData without
payload and other zeros instead.
This was on system during initial setup so for most time we didn't have
any clients at all => not much to be logged in WAL. We were trying out
replication, failover/switchover scenarios with repmgr and creating and
restoring backup with barman.
Regards,
Sasa