[Date Prev][Date Next] [Chronological] [Thread] [Top]

Re: upgrade 2.2.x -> 2.3.x, sync repl replica compatibility?



At 02:21 PM 3/28/2006, Howard Chu wrote:

Thanks for taking time to respond Howard.

...If you're running recent enough 2.2 code, any order should be fine. Logically speaking, since you probably have many consumers and only a few providers, the path of minimal impact is to change consumers one at a time and change the provider last. Of course, it may make sense to upgrade the least stable systems first, regardless of consumer or provider role.

All our instances are 2.2.29 (or at least 2.2.28). Is that recent enough? All our replicas are running the Sync replication protocol in refresh only mode.


We have one relevant production master and many (~6) slaves. Some of the slaves are under other administrative control. For we who run the master and some of the slaves it's probably best to upgrade the master first after testing to insure that an upgraded master seems to function with all the slave types we know are out there (different OSs and hardware). At least if we upgrade the master first we have a fairly simple backoff strategy. Announce, upgrade, and wait to see if there's any fallout. If there are problems then backup to our old 2.2.29 master and regroup. If there are no problems (watchdog timer resync, etc.) then we can upgrade the slaves one at a time. It's the slaves that carry our production read load, so we can function with the master down for some time.

We expect to dump and restore the master database to and from LDIF format across the master upgrade. All our slaves will have their watchdog timers go off, but that will be expected from the announcement. I will be delighted if we get to the upgraded master as we've had continuing problems (perhaps once a month) with the master seemingly getting into some sort of CPU bound deadlock where we have to kill it hard and end up having to restart all the replicas - very embarrassing. We're hoping that the upgrade to 2.3.20 will correct that problem.

--Jed http://www.nersc.gov/~jed/