launchpad-dev team mailing list archive

[Robert Collins <robertc@xxxxxxxxxxxxxxxxx>]

On Tue, Jun 14, 2011 at 8:05 PM, Stuart Bishop
<stuart.bishop@xxxxxxxxxxxxx> wrote:
> On Tue, Jun 14, 2011 at 9:07 AM, Martin Pool <mbp@xxxxxxxxxxxxx> wrote:
>
>>  * actually doing lp-related work generally requires write access
>
> With time and effort, I think pretty much everything about read only
> mode could be fixed except this. If this is true and there is little
> point to a read-only mode during outages, then we might as well ditch
> it.
>
> If we invest some time in it, it can become more useful in general. If
> appservers automatically switched when they lose their master database
> connections, it becomes a failover mode if the master db dies or
> network connections between the data centers craps out. Again, if
> failover to read-only-mode is pointless it is pointless :-)

I think there is some nuance here.

Handling the master being awol gracefully is a good thing - and
readonly requests shouldn't need the master.

Going into the big 'its all readonly' lock that we do at the moment
however is quite a different thing - its not handling degradation
gracefully, its an explicit 'do not try' flag.

I guess I mean - I'm talking about the *existence* and use of the
readonly replica as a way to permit schema changes to happen as the
root cause of a number of our deploy reliability and post-deploy
performance issues.

-Rob