Re: [HACKERS][Proposal] LZ4 Compressed Storage Manager

From: Konstantin Knizhnik <k(dot)knizhnik(at)postgrespro(dot)ru>
To: Николай Петров <nik(dot)petrov(dot)ua(at)yandex(dot)ru>, pgsql-hackers(at)postgresql(dot)org
Subject: Re: [HACKERS][Proposal] LZ4 Compressed Storage Manager
Date: 2019-04-01 09:30:22
Message-ID: eb28e65e-580e-3aff-ee93-7c5c7814a918@postgrespro.ru
Views: Raw Message | Whole Thread | Download mbox | Resend email
Thread:
Lists: pgsql-hackers

On 31.03.2019 17:25, Николай Петров wrote:
> Hello everyone!
> Thank you for your interest to this topic.
>
> I would like to propose Compressed Storage Manager for PostgreSQL.
>
> The problem:
> In cases when you store some log-like data in your tables, or when you
> store time-series data you may face with high disk space consumption
> because of a lot of data. It is a good idea to compress tables,
> especially if you have a compressible data and OLAP
> WORM (write once read many) usage scenarios.
>
> Current ways to solve this problem:
> Now this could be solved via a compressible file system such as BTRFS
> or ZFS. This approach has a contradictory impact on performance and
> connected with difficulties of administration.
>
> Other's DB approaches:
> Postgres Pro Enterprise has embedded CFS [1][2] for this purposes.
> MySQL InnoDB has two options of compression - table level compression
> (zlib only) [3] and transparency pages compression (zlib, LZ4) [4]
> via hole punching [5].
>
> My offer:
> Implement LZ4 Compressed Storage Manager. It should compress pages on
> writing to block files and decompress on reading. I would like to
> offer LZ4 at first, because it has low CPU consumption and it is
> available under BSD 2 clause license.
>
> Compressed Storage Manager operation description (TLDR: algorithm could
> be similar to MySQL table level compression):
> - It should store compressed pages in a block file, but because of
> different size of compressed data, it should have an additional
> file with offset for each pages.
> - When it reads a page, it translates upper PostgreSQL layers
> file/offset query to actual page offset, read compressed page
> bytes, decompress them and fill the requested buffer with
> decompressed page.
> - New pages writing quite a simple, it has to compress the page,
> write it to block file and write page offset into a file with
> pointers.
> - In cases when it's necessary to write changed page, it has to
> check that the size of the compressed page smaller or equal to
> previous version. If it's bigger, it is should to write page
> to the end of the block file and change the page pointer. The
> old page version became dead.
> - There is an ability to make free space release mechanism, for instance,
> MySQL use hole punching (what contradictory impact on
> performance [6]). At first time dead pages could be freed
> via VACUUM FULL.
>
> pointers file
> +====+====+====+
> | p1 | p2 | p3 |
> +=|==+==|=+==|=+
> | | |_________________________________
> | |____________________ |
> | | | block file
> +=|======+=================+=|===============+=|==================+
> | p1 len | p1 ####data#### | p2 len | p2 #d# | p3 len | p3 #data# |
> +========+=================+=================+====================+
>
>
> Test of possible compression (database [7], table ticket_flights [8]):
> 547M 47087 <- uncompressed
> 200M 47087.lz4.1.pages.compressed <-- pages compression (37%)
>
> Pros:
> - decreases disk space usage
> - decreases disk reads
> Cons:
> - possible increases random access I/O
> - increases CPU usage
> - possible conflicts with PostgreSQL expectations
> of Storage Manager behaviour
> - could conflict with pg_basebackup and pg_upgrade utilities
> - compression requires additional memory
>
> Why it should be implemented on Storage Manager level instead of usage
> Pluggable storage API [9]?
> - From my perspective view Storage Manager level implementation
> allows to focus on proper I/O operations and compression.
> It allows to write much more simple realization. It's because of
> Pluggable storage API force you to implement more complex
> interfaces. To be honest, I am really hesitating about this point,
> especially because of Pluggable storage API allows to create
> extension without core code modification and it potentially allows
> to use more perfective compression algorithms (Table Access Manager
> allows you to get more information about storing data).
>
> I would like to implement a proof of concept
> and have a couple of questions:
> - your opinion about necessity of this feature
> (Compressed Storage Manager)
> - Is it good idea to implement DB compression on Storage Manager
> level? Perhaps it is better to use Pluggable storage API.
> - Is there any reason to refuse this proposal?
> - Are there any circumstances what didn't allow to implement
> Compressed Storage Manager?
>
> Regards,
> Nikolay P.
>
> [1] - https://postgrespro.com/docs/enterprise/9.6/cfs
> [2] - https://afiskon.github.io/static/2017/postgresql-in-core-compression-pgconf2017.pdf (page 17)
> [3] - https://dev.mysql.com/doc/refman/8.0/en/innodb-table-compression.html
> [4] - https://dev.mysql.com/doc/refman/8.0/en/innodb-page-compression.html
> [5] - https://lwn.net/Articles/415889/
> [6] - https://www.percona.com/blog/2017/11/20/innodb-page-compression/
> [7] - https://postgrespro.com/education/demodb
> [8] - https://postgrespro.com/docs/postgrespro/10/apjs02
> [9] - https://commitfest.postgresql.org/22/1283/
>
>
>

I can shared my experience of development of CFS for PostgresPro.
First of all I want to notice that most likely it will be not possible
to isolate all changes in Postgres at Storage Manager level.
There are many places in Postgres (basebackup,vacuum,...) which makes
some assumptions on content of Postgres data directory.
So if compressed storage manager will provide some alternative files
layout, then other parts of the Postgres should know about it.

The most difficult thing in CFS development is certainly
defragmentation. In CFS it is done using background garbage collection,
by one or one
GC worker processes. The main challenges were to minimize its
interaction with normal work of the system, make it fault tolerant and
prevent unlimited growth of data segments.

CFS is not introducing its own storage manager, it is mostly embedded in
existed Postgres file access layer (fd.c, md.c). It allows to reused
code responsible for mapping relations and file descriptors cache. As it
was recently discussed in hackers, it may be good idea to separate the
questions "how to map blocks to filenames and offsets" and "how to
actually perform IO". In this it will be easier to implement compressed
storage manager.

--
Konstantin Knizhnik
Postgres Professional: http://www.postgrespro.com
The Russian Postgres Company

In response to

Browse pgsql-hackers by date

  From Date Subject
Next Message Rafia Sabih 2019-04-01 09:39:21 Re: explain plans with information about (modified) gucs
Previous Message Matsumura, Ryo 2019-04-01 09:29:06 RE: SQL statement PREPARE does not work in ECPG