Re: Seeking datacenter PITR backup procedures [RESENDING] - Mailing list pgsql-general

"Joey K." <pguser@gmail.com> wrote:
>
> Greetings,

I'm guessing you're in a hurry or in a pinch that you need to repost
after one day on a weekend.  I was waiting to let someone more
knowledgeable answer, but I've had some experience with this, so
I'll answer to the best of my ability.

> We have several web applications with Pg 8.2.x running on isolated servers
> (~25). The database size on each machines (du -h pgdata) is ~2 GB. We have
> been using nightly filesystem backup (stop pg, tar backup to ftp, start pg)
> and it worked well.

Any reason why you haven't been using pg_dump?  There are a LOT of
drawbacks to doing filesystem level backups.  For example, you can't
restore to disparate hardware (a filesystem backup made from PG on
an i386 system won't work on an amd64 system, for example)

> We would like to move to PITR backups since the database size will increase
> moving forward and our current backup method might increase server
> downtimes.

How much do you expect it to increase?  2G is _tiny_ by modern
standards.  Even if you expect it to increase an order of magnitude,
it's still a reasonable size for pg_dump.

Some huge advantages to pg_dump:
*) architecture-neutral dumps
*) No need to stop the database
*) Extremely simple procedure for backup and restore
*) Human parseable backups (you may not even need to restore, just
   look through the data to see what was there in some cases)
*) Can restore a database without shutting down a server, thus you
   can move a database from one server to another (for example)
   without affecting work occurring on the second server.

> We have a central ftp backup server (yes, ftp :-) which we would like to use
> for weekly full and daily incremental PITR backups.
>
> After reading the docs, PITR is still fuzzy. Our ideas for backup are (do
> not worry about the syntax),
>
> ** START **
>
> tmpwal = "/localhost/tmp"   # tmp space on server 1 for storing wal files
> before ftp
> Configure $pgdata/postgresql.conf archive_command = "cp %p $tmpwal/%f"
>
> Day 1:
> % psql pg_start_backup(); tar pgdata.tar --exclude pg_xlog/ pgdata
> % psql pg_stop_backup()
> % ftp put pgdata.tar ftpserver:/server1/day1/pgdata
> % ftp put $tmpwal/* ftpserver:/server1/day1/wal
> % rm -f $tmpwal/* pgdata.tar
>
> Day 2:
> % ftp put $tmpwal/* ftpserver:/server1/day2/wal
> % rm -f $tmpwal/*
>
> Day 3:
> ...
> ...
>
> Day 7:
> % rm -f $tmpwal/*
> Start over
>
> Recovery on server1 (skeleton commands),
> % rm -f $tmpwal/*

Why are you removing this day's WAL logs before recovery?  If the
disaster occurs close to your backup time, this will result in the
loss of an entire day's data changes.

> % mv pgdata pgdata.hosed
> % ftp get ftpbackup:/server1/day1/pgdata.tar  .
> % tar -xvf pgdata.tar
> % ftp get ftpbackup:/server1/day1/wal/*  $tmpwal
> % ftp get ftpbackup:/server1/day2/wal/*  $tmpwal
> .....
> .....
> % cp -r pgdata.hosed/pg_xlog pgdata/
> % echo "cp $tmpwal/%f %p" > pgdata/recovery.conf
> % start pg (recovery begins)
>
> ** END **

Again, how much WAL traffic are you generating?  Make sure you have
enough free space on the recovery system to hold all of the WAL logs
in the event you need to recover.

Note that this procedure does not do a good job of protecting you
from catastrophic hardware failure.  In the event that your RAID
system goes insane, you can lose as much as an entire day's worth
of updates, and there's no reason to.

Is there a reason you can't use (for example) NFS to transfer each WAL
log to the backup system on the fly?  This would better protect you
from IO system failures, and reduce the amount of disk space you need
on each server, since you don't need to keep WAL logs there.

> Assumptions:
> a. After pg_stop_backup(), Pg immediately recycles log files and hence wal
> logs can be copied to backup. This is a clean start.

I don't believe so.  ARAIK, all pg_stop_backup() does is remove the
marker that pg_start_backup() put in place to tell the recovery process
when the filesystem backup started.  By not backing up pg_xlog, you are
going to be behind by however many transactions are in the most recent
transaction log that has not yet been archived.  Depending on how often
your databases are updated, this is likely acceptable.  If you need
anything more timely than that, you'll probably want to implement
Slony or some other replication system.

> b. New wal files since (a) are incremental backups

Yes.

> We are not sure if WAL log filenames are unique and possibly overwrite older
> wal files during recovery.

They are unique.  Eventually they will recycle, but if you look at the
names and the number of possible names, you'll see that it will take a
long time before the names recycle.

> I'm seeking suggestions from others with experience performing PostgreSQL
> PITR backups from multiple servers to a central backup server.

Again, I recommend the use of NFS (or similar) for real-time backup
to protect you from hardware failure.

I _highly_ recommend you stage some disaster scenarios and actually
use your procedure to restore some databases before you trust it.
Getting PITR working effectively is tricky, no matter how many questions
you ask of knowledgeable people.  You're going to want to have first-
hand experience going through the process.

HTH.

--
Bill Moran
http://www.potentialtech.com