On 01/24/2014 10:29 PM, Josh Berkus wrote:
> On 01/24/2014 12:47 PM, Heikki Linnakangas wrote:
>> ISTM the consensus is that we need better monitoring/administration
>> interfaces so that people can script the behavior they want in external
>> tools. Also, a new synchronous apply replication mode would be handy,
>> but that'd be a whole different patch. We don't have a patch on the
>> table that we could consider committing any time soon, so I'm going to
>> mark this as rejected in the commitfest app.
> I don't feel that "we'll never do auto-degrade" is determinative;
> several hackers were for auto-degrade, and they have a good use-case
> argument.
Auto-degrade may make sense together with synchronous apply
mentioned by Heikki.
I do not see much use for synchronous-(noapply)-if-you-can mode,
though it may make some sense in some scenarios if sync failure
is accompanied by loud screaming ("hey DBA, we are writing checks
with no money in the bank, do something fast!")
Perhaps some kind of sync-with-timeout mode, where timing out
results with a "weak error" (something between current
warning and error) returned to client and/or where it causes and
external command to be run which could then be used to flood
admins mailbox :)
> However, we do have consensus that we need more scaffolding
> than this patch supplies in order to make auto-degrade *safe*.
>
> I encourage the submitter to resumbit and improved version of this patch
> (one with more monitorability) for 9.5 CF1. That'll give us a whole
> dev cycle to argue about it.
>
Cheers
--
Hannu Krosing
PostgreSQL Consultant
Performance, Scalability and High Availability
2ndQuadrant Nordic OÜ