On Thu, Feb 4, 2016 at 11:41 PM, Tom Lane <tgl@xxxxxxxxxxxxx> wrote:
=?UTF-8?Q?Filip_Rembia=C5=82kowski?= <filip.rembialkowski@xxxxxxxxx> writes:
> A table has a trigger.
> The trigger sends a NOTIFY.
> Test with COPY FROM shows non-linear correlation between number of inserted
> rows and COPY duration.
No surprise, see AsyncExistsPendingNotify. You would have a lot of other
performance issues with sending hundreds of thousands of distinct notify
events from one transaction anyway, so I can't get terribly excited about
this.
What kind of issues? Do you mean, problems in postgres or problems in client?
Is there an additional non-linear cost on COMMIT (extra to the cost I already showed)?
The 8GB internal queue (referenced in a Note at http://www.postgresql.org/docs/current/static/sql-notify.html) should be able to keep ~ 1E8 such notifications (assumed one notification will fit in 80 bytes).
On client side, this seems legit - the LISTENer deamon will collect these notifications and process them in line.
There might be no LISTENer running at all.
Still, the main problem I get with this approach is quadratic cost on big insert transactions.
I wonder if this behavior is possible to change in future postgres versions. And how much programming work does it require.
Is duplicate-elimination a fundamental, non-negotiable requirement?
Thank you,
Filip
