On 5/15/24 10:42 PM, David Rowley wrote:
> Thanks for writing that up. It's always exciting to see this each year.
>
> On Thu, 16 May 2024 at 13:45, Jonathan S. Katz <jkatz@postgresql.org> wrote:
>> * Please indicate if you believe there's a notable omission, or if we
>> should omit any of these callouts
>
> I'd say the streaming read stuff added in b5a9b18cd and subsequent
> commits like b7b0f3f27 and 041b96802 are worth a mention. I'd be happy
> to see this over the IS NOT NULL qual stuff that I worked on in there
> or even the AVX512 bit counting. Speeding up a backwater aggregate
> function is nice, but IMO, not compatible with reducing the number
> reads.
> There's some benchmarking in a youtube video:
> https://youtu.be/QAYzWAlxCYc?si=L0UT6Lrf067ZBv46&t=237
Nice! Definitely agree on including this - it wasn't initially clear to
me on the read of the release notes. I'll update it. Please see in the
next revision (will posted upthread), proposed text here for
convenience, as I'm not sure I'm appropriately capturing it:
==
This release introduces adds an interface to stream I/O, and can show
performance improvements when performing sequential scans and running
[`ANALYZE`](https://www.postgresql.org/docs/17/sql-analyze.html).
==
The AVX-512 bit counting showed solid impact[1] on the binary distance
functions in pgvector (I have to re-run again w/v17, as I seem to recall
seeing some numbers that boosted it 5-7x [but recall isn't 100% ;)]).
>> * Please indicate if a description is confusing - I'm happy to rewrite
>> to ensure it's clearer.
>>
>> Please provide feedback no later than Wed 2024-05-22 18:00 UTC.
>
> The only other thing I saw from a quick read was a stray "the" in "the
> copy proceed even if the there is an error inserting a row."
Thanks!
Jonathan
[1] https://github.com/pgvector/pgvector/pull/519
