Обсуждение: Backup/disaster recovery and bandwidth (long)

p { margin: 0; }

Hello, everyone. I want to throw a scenario out there to see what y'all think.

Soon, my cluster backups will be increasing in size inordinately. They're going to immediately go to 3x as large as they currently are with the potential to be about 20x within a year or so.

My current setup uses a single PG 8.x server doing nightly dumps (not ideal but sufficient for the moment, and one of the main reasons to move to PG 9) which are then downloaded from my hosting center to our offices for DR purposes. Each night I pull down roughly 5GB of compressed pg_dump data. Dumping this takes about 1.5hrs. Downloading this at 15Mbps takes about an hour. Soon I'll be looking at somewhere around 7hrs for the dumps to complete and downloading a 12GB file (which will take about 3 hrs). Oh, and I'll have to pay for significant bandwidth overage since I'm charged on a 95%, and while an hour a day does NOT kick me up to 15Mbps usage at 95%, 3hrs per night certainly will, so there's a real cost associated with this strategy as well.

While the time of the actual dumps is not a huge issue, the time of the download IS a large concern, especially since my support folks use that file daily to extract individual customer databases for restore in assisting customer support issues.

So, while now I have my pg_dumps completed around 2AM and downloaded to my local network at about 3AM, with the increase in our database sizes, what will be happening is that my pg_dump will not be completed until around 7AM, and the download would not be completed until around 10AM, best-case scenario. Add into that support trying to restore a database...more on that in a moment.

My _new_ setup will instead be 2 PG 9.x servers with hot-standby enabled (at my hosting center) and a 3rd PG 9.x server at my local office also replicating off of the master. Each one of those servers will perform his own pg_dumps of the individual databases for backup/disaster recovery purposes, and while each dump might not be consistent with one another, each SERVER will have dumps consistent to itself, which is viable for our situation, and does not require me to download 12GB (or more) each night with all of those associated nightmares, costs and other problems.

Alright, well, I've got that part all thought out, and it seems like a good way to do it to me, but I'm _still_ running into the situation that I've got to take 8hrs-ish to run the pg_dump no matter where it runs, and when my support folks need it (which they do daily), this basically means that if they have to have a customer database up NOW NOW NOW for support reasons, they simply cannot have it within an hour in many cases. Specifically, one database takes between 2 and 7.5hrs to pg_dump depending on which format I use, so if they need a CURRENT copy, they're at least 4 hours out. Additionally, they can't directly use the replicating server at my local office, because they need to test the problems the customers are having which include pesky things like INSERT, UPDATE and DELETE, so they have to restore this data to another internal PG backend.

Enter my outside-the-box thinking.

I rather assume that you cannot do a start/stop backup on a hot-standby server. HOWEVER, what if....

I set up a 4th database server internally at my office. Each night I stop PG on my 3rd server (the local one replicating off of the master) and rsync my pg_data directory to this new 4th server. I bring up the 4th server NOT as a standby, but as a master. They would then have all customer data on an internal, usable PG system from the time of the rsync, and while it might not reflect the immediate state of the database, that's pretty well always true, and they're used to that, since whenever they "clone" a site, they're using the dumps done around midnight anyway.

I believe, then, that when I restart server #3 (the standby who is replicating), he'll say "oh, geez, I was down, let me catch up on all that crap that happened while I was out of the loop," he'll replay the WAL files that were written while he was down, and then he'll catch back up.

Does this sound like a viable option? Or does someone have additional suggestions?

On Apr 25, 2012, at 10:11 AM, Scott Whitney wrote:

> I believe, then, that when I restart server #3 (the standby who is replicating), he'll say "oh, geez, I was down, let
mecatch up on all that crap that happened while I was out of the loop," he'll replay the WAL files that were written
whilehe was down, and then he'll catch back up. 
>
> Does this sound like a viable option? Or does someone have additional suggestions?

Perfectly viable. However, the WAL files must exist for this to happen. So you need to set wal_keep_segments
appropriately,or set up WAL archiving. 

Note that you could even provide a more up-to-date database for your people to work with. If the testbed is nearly up
todate, then an rsync to update it would take very little time. So you could shut down the replica, rsync, and bring
thereplica back up. 


--
Scott Ribe
scott_ribe@elevated-dev.com
http://www.elevated-dev.com/
(303) 722-0567 voice

p { margin: 0; }

...
My current setup uses a single PG 8.x...
My _new_ setup will instead be 2 PG 9.x ...

I set up a 4th database server internally at my office. Each night I stop PG on my 3rd server (the local one replicating off of the master) and rsync my pg_data directory to this new 4th server. I bring up the 4th server NOT as a standby, but as a master....

Does this sound like a viable option? Or does someone have additional suggestions?

I  mean bucardo (even though there are more tools like this one) just
for the replication stuff and the hot database backup only for the
backup stuff and only one bounce is needed to turn the archiving on, you
do not need to turn anything at all down during the backup.

A.A

On 04/25/2012 10:23 AM, Scott Ribe wrote:
> On Apr 25, 2012, at 10:11 AM, Scott Whitney wrote:
>
>> I believe, then, that when I restart server #3 (the standby who is replicating), he'll say "oh, geez, I was down,
letme catch up on all that crap that happened while I was out of the loop," he'll replay the WAL files that were
writtenwhile he was down, and then he'll catch back up. 
>>
>> Does this sound like a viable option? Or does someone have additional suggestions?
> Perfectly viable. However, the WAL files must exist for this to happen. So you need to set wal_keep_segments
appropriately,or set up WAL archiving. 
>
> Note that you could even provide a more up-to-date database for your people to work with. If the testbed is nearly up
todate, then an rsync to update it would take very little time. So you could shut down the replica, rsync, and bring
thereplica back up. 
>
>