Paul B. Henson via illumos-discuss wrote:
On Wed, Jul 23, 2014 at 01:27:47PM +0200, Saso Kiselkov via illumos-discuss
wrote:
For async replication, incremental snapshots are the way to go. Just
transfer an incremental snapshot every 30-60s. They're relatively
inexpensive and robust, so pool corruption doesn't immediately get
propagated to the replication pair.
We were looking to something like this recently, but ran into
scalability issues with large filesystem counts. It took something like
5 minutes to do a zfs recv of 1000 filesystems (with no changes since
last snapshot). The zfs send took only 16 seconds, crossing the network
took seconds, but the recv performance broke it horribly. One of my
colleagues that was working on it determined it was setting properties
that was killing it, so rather that doing a send -R of everything
everytime, we tried just sending filesystems that had changed. But while
a send -R of 1000 filesystems took 16 seconds, generating individual
send streams separately for 100 took 3 minutes 8-/. We ended up having
to give up on the idea of near real time replication and go another way.
That really is the beauty of DRBD and HAST - they're mirroring disk
writes, as they happen. Everything up to date, all the time, across two
machines (not as good for n-machine clusters, though - you can stack
them, but that gets baroque).
Miles Fidelman
--
In theory, there is no difference between theory and practice.
In practice, there is. .... Yogi Berra
-------------------------------------------
illumos-discuss
Archives: https://www.listbox.com/member/archive/182180/=now
RSS Feed: https://www.listbox.com/member/archive/rss/182180/21175430-2e6923be
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21175430&id_secret=21175430-6a77cda4
Powered by Listbox: http://www.listbox.com
