Re: ML-based indexing ("The Case for Learned Index Structures", a paper from Google)

Mi., 21. Apr. 2021, 11:16 Uhr, Oleg Bartunov <obartunov(at)postgrespro(dot)ru> wrote:
> Have you seen recent paper "Benchmarking Learned Indexes" ?

Yes. I skipped it after that this benchmark "just" compares the
algorithm implementations.

What's needed - and what many here as well as the "ML-In-Databases"
paper from Kraska et al. (2021) are saying - is, that a new index
(like a learned index) should be implemented as a PostgreSQL
extension.

Mi., 21. Apr. 2021, 15:46 Uhr, Bruce Momjian <bruce(at)momjian(dot)us> wrote:
> The issue is that some index structures, like bitmap indexes, have very
> poor concurrent performance. This means that some indexes perform very
> well for a single user but poorly for multiple users.

I see now. That looks to me like a second step of an experiment to
implement a possible new index.

~Stefan

Am Mi., 21. Apr. 2021 um 15:46 Uhr schrieb Bruce Momjian <bruce(at)momjian(dot)us>:
>
> On Wed, Apr 21, 2021 at 10:52:19AM +0200, Stefan Keller wrote:
> > Di., 20. Apr. 2021 23:50 Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> wrote:
> > > There's enough support these days that you can build a new index
> > > type as an extension, without touching the core code at all.
> >
> > Thanks. I'm ramping up knowledge about extending PG with C++.
> >
> > I'm still interested to understand in principle what an index has to
> > do with concurrency control, in order to divide
> > concerns/reponsibilities of code.
>
> The issue is that some index structures, like bitmap indexes, have very
> poor concurrent performance. This means that some indexes perform very
> well for a single user but poorly for multiple users.
>
> --
> Bruce Momjian <bruce(at)momjian(dot)us> https://momjian.us
> EDB https://enterprisedb.com
>
> If only the physical world exists, free will is an illusion.
>