Re: [PATCH] Compression dictionaries for JSONB - Mailing list pgsql-hackers

From Matthias van de Meent
Subject Re: [PATCH] Compression dictionaries for JSONB
Date
Msg-id CAEze2WjyxVnKEDv8s5eP=dwYgtBKrQzVF4wvWk2Nb1FRgGGmTw@mail.gmail.com
Whole thread Raw
In response to Re: [PATCH] Compression dictionaries for JSONB  (Aleksander Alekseev <aleksander@timescale.com>)
Responses Re: [PATCH] Compression dictionaries for JSONB
List pgsql-hackers
On Mon, 6 Feb 2023 at 15:03, Aleksander Alekseev
<aleksander@timescale.com> wrote:
>
> Hi,
>
> I see your point regarding the fact that creating dictionaries on a
> training set is too beneficial to neglect it. Can't argue with this.
>
> What puzzles me though is: what prevents us from doing this on a page
> level as suggested previously?

The complexity of page-level compression is significant, as pages are
currently a base primitive of our persistency and consistency scheme.
TOAST builds on top of these low level primitives and has access to
catalogs, but smgr doesn't do that and can't have that, respectively,
because it need to be accessible and usable without access to the
catalogs during replay in recovery.

I would like to know how you envision we would provide consistency
when page-level compression would be implemented - wouldn't it
increase WAL overhead (and WAL synchronization overhead) when writing
out updated pages to a new location due to it changing compressed
size?

> More similar data you compress the more space and disk I/O you save.
> Additionally you don't have to compress/decompress the data every time
> you access it. Everything that's in shared buffers is uncompressed.
> Not to mention the fact that you don't care what's in pg_attribute,
> the fact that schema may change, etc. There is a table and a
> dictionary for this table that you refresh from time to time. Very
> simple.

You cannot "just" refresh a dictionary used once to compress an
object, because you need it to decompress the object too.

Additionally, I don't think block-level compression is related to this
thread in a meaningful way: TOAST and datatype -level compression
reduce the on-page size of attributes, and would benefit from improved
compression regardless of the size of pages when stored on disk, but a
page will always use 8kB when read into memory. A tuple that uses less
space on pages will thus always be the better option when you're
optimizing for memory usage, while also reducing storage size.

> Of course the disadvantage here is that we are not saving the memory,
> unlike the case of tuple-level compression. But we are saving a lot of
> CPU cycles

Do you have any indication for how attribute-level compares against
page-level compression in cpu cycles?

> and doing less disk IOs.

Less IO bandwidth, but I doubt it uses less operations, as each page
would still need to be read; which currently happens on a page-by-page
IO operation. 10 page read operations use 10 syscalls to read data
from disk - 10 IO ops.

> I would argue that saving CPU
> cycles is generally more preferable. CPUs are still often a bottleneck
> while the memory becomes more and more available, e.g there are
> relatively affordable (for a company, not an individual) 1 TB RAM
> instances, etc.

But not all systems have that 1TB RAM, and we cannot expect all users
to increase their RAM.

> So it seems to me that doing page-level compression would be simpler
> and more beneficial in the long run (10+ years). Don't you agree?

Page-level compression can not compress patterns that have a length of
more than 1 page. TOAST is often used to store values larger than 8kB,
which we'd prefer to compress to the greatest extent possible. So, a
value-level compression method specialized to the type of the value
does make a lot of sense, too.

I'm not trying to say that compressing pages doesn't make sense or is
useless, I just don't think that we should ignore attribute-level
compression just because page-level compression could at some point be
implemented too.


Kind regards,

Matthias van de Meent



pgsql-hackers by date:

Previous
From: Tom Lane
Date:
Subject: Re: run pgindent on a regular basis / scripted manner
Next
From: Andrew Dunstan
Date:
Subject: Re: run pgindent on a regular basis / scripted manner