Re: 9.1.9 -> 9.1.10 causing corruption

[Ben Chobot <bench@xxxxxxxxxxxxxxx>]

Anybody? I've tried this again on another streaming replication server, and again had pg_toast errors until I re-basebackup'd it. Does it make sense to try disabling streaming replication for the backup and just use wal replay from the archive? My slave will be a bit behind, but I'll have drained the clients from it anyway, so that's not a big deal for a temporary thing.

On Nov 1, 2013, at 1:44 PM, Ben Chobot wrote:

I've got a bunch of independent database clusters, each with a couple of streaming replication slaves. I'm starting to upgrade them to 9.1.10, but on 3 of the 3 I've tried so far, this has somehow resulted in data corruption. I'm hoping it was the upgrade itself that caused the corruption, instead of 9.1.10 just exposing some latent corruption, for two reasons:

1. after doing a basebackup from the master, the same queries that started causing problems on the new 9.1.10 node no longer cause problems

2. the second 9.1.9 slave of each cluster doesn't have issues that the newly-upgraded 9.1.10 node did have with the same query. 

Interestingly, this seems to only be affecting my replication slaves. We snapshot all our clusters at a filesystem level and restore those snapshots into standalone, non-replicating test environments, and those test environments were all upgraded to 9.1.10 without issue. 

All my servers are running a mostly-stock config of 9.1.9, which has been modified like so:

max_connections = 1000 # we don't really go this high, fwiw

shared_buffers = 6000MB # 60GB machines

wal_level = hot_standby

hot_standby = on

effective_cache_size = 34278MB

log_min_duration_statement = 50

log_duration = on

 

They store their data on XFS, with the options noatime,allocsize=1m.

As for the errors, on two nodes, a primary key lookup resulted in multiple rows. Interestingly, this was on the same index for both nodes, which I would find unlikely if it was just random corruption or hardware-related, given that the systems have the same schema but independent data and usage. Using pg_reorg on that table on the master node fixed the issue on the node I tried it on, but exposed another problem:

ERROR: unexpected chunk number 3 (expected 1) for toast value 67728890 in pg_toast_2619

The pg_toast error occurred on another node too. Sadly these were are production systems and I didn't know if the pg_toast table was the same for all three nodes, and it's quite possible that all three nodes would have exhibited the same set of problems if I'd let them run production queries long enough.

Anyway, naturally this is making me nervous to do more upgrades, so I'm hoping I can figure out if it's the upgrade that's killing me, or if 9.1.10 is exposing underlying problems, or what. Unfortunately I have no idea what to look for.

These system are ubuntu precise, and I did recall this from my memory:

http://www.postgresql.org/message-id/21697.1343141116@xxxxxxxxxxxxx

... but it appears that the fix for that has already been applied to my postgresql-common package. Looking over the postgres logs before the upgrade, I don't see anything abnormal.

Does this behavior sound familiar to anybody else out there, or does anybody have suggestions of what to look for or try?