Some operational notes

[mark]

When doing a global master change...

* don't forget MogileFS
* stop all jobs, if anybody is still up with the old config (TheSchwartz workers!) they will continue writing to the old master
* take the old master hard-offline and leave it that way to flush out any bad behavers? at least monitor show processlist
* don't forget all slaves (mailserver slave)

Todo: figure out if there's a way to put a MySQL server fully read-only (except for replication?), I know there's a read-only config file option but I don't think it works as expected.

Comments

[alierak]

So this MySQL replication thing seems to be sort of medium-okay-availability, not what I would call high availability. Why shouldn't there be a way for clients to fail over automatically? Stick it behind a failover-capable frontend maybe? I need to find out what the state of the art is on this, I guess. For work I'm thinking of enforcing use of a DNS cname to connect to databases, so I could do failover with a DNS change (and closing all connections so clients have to retry). But it should be possible to do better than that with something like a load balancer.

[mark]

Oh yeah, I wrote proxy for MySQL in Perl (http://code.google.com/p/proximo/, http://xb95.dreamwidth.org/tag/proximo) which is actually a really damn nifty bit of software. It's just pretty slow and not that useful in production, if it was sped, it could be amazingly useful.

It can do the sort of fallback you're talking about. It can also force stupid applications that don't know how to split up reads and writes to different databases to actually take advantage of a slave. And some other neat things.

But anyway.

Yeah, MySQL replication doesn't give HA. If you want true HA, you can look at MySQL Cluster, which is a certain technology and supposedly works pretty well. I haven't played with that since 2004 or so, when it didn't work very well at all.

Reading Filters

[morgandawn]

Hi mark

is there any ETA on when we can set up reading filters on Dreamwidth?