Re: Addled index

Oleg Alexeev <oalexeev(at)gmail(dot)com> writes:
> * it is varchar columns, 256 and 32 symbols length
> * encoding, collation and ctype: UTF8, en_US.utf8, en_US.utf8
> * autovacuum, fsync off, full_page_writes = on, wal_writer_delay = 500ms,
> commit_delay = 100, commit_siblings = 10, checkpoint_timeout = 20min,
> checkpoint_completion_target = 0.7
> * postgres 9.2.3 installed via yum repository for version 9.2
> * 64 bit Centos 6, installed and updated from yum repository

fsync off? Have you had any power failures or other system crashes?
ext4 is *way* more prone than ext3 was to corrupt data when fsync is
disabled, because it caches and reorders writes much more aggressively.

> Database located on software md raid 1 based on two SSD disks array. Ext4
> filesystem. Database is master node.

Meh. I quote from the RHEL6 documentation (Storage Administration
Guide, Chapter 20: Solid-State Disk Deployment Guidelines):

> Red Hat also warns that software RAID levels 1, 4, 5, and 6 are not
> recommended for use on SSDs.

https://access.redhat.com/knowledge/docs/en-US/Red_Hat_Enterprise_Linux/6/html/Storage_Administration_Guide/newmds-ssdtuning.html

The part of the docs I'm looking at only asserts that performance is
bad, but considering that it's a deprecated combination, it may well be
that there are data-loss bugs in there. I'd certainly suggest making
sure you are on a *recent* kernel. If that doesn't help, reconsider
your filesystem choices.

(Disclaimer: I work for Red Hat, but not in the filesystem group,
so I don't necessarily know what I'm talking about. But I have the
feeling you have chosen a configuration that's pretty bleeding-edge
for RHEL6.)

regards, tom lane