On Wed, Oct 18, 2017 at 6:57 PM, John <[email protected]> wrote:
> Would it not be better to trial the switch over with the *.labsdb aliases
> before risking catastrophic failure when rebooting? Doing a short test (up
> to 24 hours) would allow users to identify anything that may break before it
> becomes an unbreak now situation? If any critical systems are affected the
> change can be rolled back, issues identified and fixed before the final
> switch over.

Switching the "c1.labsdb" and "c3.labsdb" aliases will break 100% of
the Tools and other users that are using user created databases on the
hosts they currently point to. The server reboots have a small chance
of non-recoverable hardware failure, but there is really nothing we
can do to prevent that. We considered putting the reboots off and just
waiting for the servers to be decommissioned, but ultimately it was
decided that the risk of continuing to run out of date Linux kernels
was worse than the risk of hardware failure.

There is a tool at <https://tools.wmflabs.org/tool-db-usage/> where
you can see the local databases on the c1 and c3 servers. Anything
that is using these tables will need to change behavior somehow.

It is not a happy thing for us to force anyone to change their
software, but as explained in the wiki page [0] we can not find a
reliable method to ensure that the same user created tables are
available on all three of the new backend servers, and we feel that in
order to be able to properly maintain the new servers we need to have
more flexibility in choosing which traffic goes to which physical
server at any given point in time. If we recreate the current state
where certain service names are guaranteed to point to specific
physical servers Tools will break at arbitrary times when we are doing
otherwise invisible maintenance.

[0]: https://wikitech.wikimedia.org/wiki/Wiki_Replica_c1_and_c3_shutdown

Bryan
-- 
Bryan Davis              Wikimedia Foundation    <[email protected]>
[[m:User:BDavis_(WMF)]] Manager, Cloud Services          Boise, ID USA
irc: bd808                                        v:415.839.6885 x6855

_______________________________________________
Wikimedia Cloud Services mailing list
[email protected] (formerly [email protected])
https://lists.wikimedia.org/mailman/listinfo/cloud

Reply via email to