2009/7/14 Alvaro Herrera <alvherre@commandprompt.com>:
> Jeff Davis wrote:
>
>> The only problem there is telling the btree AM whether or not to do the
>> insert or not (i.e. fake versus real insert). Perhaps you can just do
>> that with careful use of a global variable?
>>
>> Sure, all of this is a little ugly, but we've already acknowledged that
>> there is some ugliness around the existing unique constraint and the
>> btree code that supports it (for one, the btree AM accesses the heap).
>
Well the ugliness referred to here (btree accessing the heap) seems
like a necessary evil. I don't think I want to add to it by
introducing global variables.
> My 2c on this issue: if this is ugly (and it is) and needs revisiting to
> extend it, please by all means let's make it not ugly instead of moving
> the ugliness around. I didn't read the original proposal in detail so
> IMBFOS, but it doesn't seem like using our existing deferred constraints
> to handle uniqueness checks unuglifies this code enough ... For example
> I think we'd like to support stuff like "UPDATE ... SET a = -a" where
> the table is large.
>
This patch works OK for around 1M rows. 10M is a real stretch (for me
it took around 1.7GB of backend memory). Any larger than that is not
going to be feasible. There is a separate TODO item to tackle this
scalability limit for deferred triggers, and I'd like to tackle that
in a separate patch. I think more discussion needs to be had on ways
to fix this (and hopefully unuglify that code in the process).
ITSM that it is not simply a matter of spooling the current queues to
disk. There is code in there which scans whole queues shuffling things
around. So perhaps a queue per trigger would help optimise this,
allowing us to move a whole queue cheaply, or drop it in favour of a
bulk check. I've not thought it through much more than that so far.
- Dean