Comments on: PostgreSQL and MySQL: Millions of Queries per Second

By: Sveta Smirnova

Sveta Smirnova — Mon, 13 Jul 2020 10:57:55 +0000

Not

By: Holger Winkelmann

Holger Winkelmann — Mon, 13 Jul 2020 05:16:22 +0000

Hi, Have you any Metrics about the Latency behaviour while running the tests?

By: Bai Yang

Bai Yang — Wed, 18 Sep 2019 03:09:30 +0000

Seem that the binary log of mysql is disabled? I think we can look at the results after enable it. After all, mysql has two kinds of logs, and its master-slave replication relies on the binlog. PostgreSQL requires only wal to achieve both ACID compliance and replication.

Therefore, in the production environment, mysql almost needs to enable binlog, and the comparison result with binlog enabled is more real.

By: Sveta Smirnova

Sveta Smirnova — Wed, 17 Jul 2019 13:40:27 +0000

In reply to Reiso. No for both questions. This project was addition to my "normal" support job and it seems I do not have time to repeat this effort right now. Maybe in far future. However, since now Percona supports PostgreSQL there is a hope that benchmarking team at Percona will do something. Subscribe to posts by Vadim Tkachenko and Alexey Stroganov.

By: Reiso

Reiso — Wed, 17 Jul 2019 13:20:14 +0000

Have you considered or plan doing the comparison with mariadb 10.4 and postgres 11? Do you know of any benchmarks?

By: Reiso

Reiso — Wed, 17 Jul 2019 12:13:52 +0000

In reply to Reiso. what I meant was you can delete my comment :) my apologies

By: Reiso

Reiso — Wed, 17 Jul 2019 12:12:39 +0000

In reply to Reiso. no need to answer.

By: Sveta Smirnova

Sveta Smirnova — Wed, 17 Jul 2019 12:12:27 +0000

In reply to Reiso. What do you mean by how many tests were done on the MySQL side? Everything is in the blog post: sysbench job warmed up, then run for 300 seconds for each number of clients. Yes, PostgreSQL incremented number of clients by 10, because my partners did not use PgBouncer. For MySQL incrementing number of clients by 10 does not make sense, because it can handle 1024 connections out of the box. Thread Pool plugin needed only if number of connections exceeds 5K or even 10K.

By: Reiso

Reiso — Wed, 17 Jul 2019 12:05:17 +0000

Hi, having a doubt on OLTP_RW and even OLTP_RO, looking at the graphs from >150 clients how many tests where done on the mysql side? One? postgres seems to have had 10 clients increments done. Is my assumption correct?

By: Sveta Smirnova

Sveta Smirnova — Wed, 17 Jul 2019 11:48:19 +0000

In reply to Alicja Kucharczyk (@StiepanTrofimo). Sorry, missed your comment. It is in the blog post: 300 seconds. But data set was warmed up before running tests.

By: Alicja Kucharczyk (@StiepanTrofimo)

Alicja Kucharczyk (@StiepanTrofimo) — Sun, 19 May 2019 23:38:24 +0000

Thank you for the post. Could you please disclose how long took you to do those tests? at least approximately if possible

By: Yousef

Yousef — Sat, 02 Feb 2019 01:02:53 +0000

A real life test revealed that Percona is,at least, 3x faster then postgres, not using any if those benchmarking tools

By: Inceptus

Inceptus — Tue, 02 Jan 2018 16:49:06 +0000

Nice explanation & thanks for sharing.

By: Sveta Smirnova

Sveta Smirnova — Thu, 02 Mar 2017 10:01:34 +0000

In reply to zhao. Oh, looks like Dimitri's graphs misplaced! trx=2 Dimitri is certainly faster than trx=1 Dimitri. Thank you for noticing this! I will ask to replace graph with correct one. In any case reason is obvious: with trx=2 flushing occurs once per second while with trx=1 at every transaction.

By: zhao

zhao — Thu, 02 Mar 2017 00:59:22 +0000

In the graph of “OLTP RW”,why the performance of MySQL-5.7 trx=1 Dimitri is better than MySQL-5.7 trx=2 Dimitri ?
why the the performance of MySQL-5.7 trx=2 Sveta is better than MySQL-5.7 trx=1 Sveta ?

By: Anastasia Raspopina

Anastasia Raspopina — Mon, 23 Jan 2017 22:43:20 +0000

In reply to Adam Scott. All the credits should go to Sveta, actually. All the tests are hers, my assistance was minor and was mainly about organizational things, intros and such. Thank you for supporting the idea though - hopefully we'll see more open source databases tested under the umbrella of this project.

By: Adam Scott

Adam Scott — Mon, 23 Jan 2017 22:07:23 +0000

Thank you Anastasia and Sveta.

By: Sveta Smirnova

Sveta Smirnova — Sat, 21 Jan 2017 00:33:12 +0000

In reply to Mark Callaghan. > Can you summarize what is in the concurrency_kit sysbench branch? This branch uses LuaJIT instead of Lua and has scaling improvements. Actually looks like while this post was under review Alexey merged concurrency_kit into upstream. At least this is that commit logs say. > Is prepared statement support in sysbench now? You can use prepared statements with any version of SysBench! Difference between 0.4 and 0.5/1.0 versions is that 0.4 had --oltp-ps-mode option while in 0.5/1.0 you have to implement prepared statements in Lua code. Alexey wants to re-implement --oltp-ps-mode for modern versions: https://github.com/akopytov/sysbench/issues/95 I believe this is the reason why he does not advertise prepared statements support. But I did not wait and just re-wrote standard tests. You may use them as a workaround for now: https://github.com/akopytov/sysbench/pull/94 > What features do you and Alexey want to add to sysbench? To be able to continue collaboration with PostgreSQL we need to solve SysBench+PostgreSQL issue. For MySQL I am happy so far.

By: Mark Callaghan

Mark Callaghan — Sat, 21 Jan 2017 00:12:14 +0000

You did a great job tuning, running and documenting your work.

Can you summarize what is in the concurrency_kit sysbench branch?
Is prepared statement support in sysbench now?
What features do you and Alexey want to add to sysbench?

By: Sveta Smirnova

Sveta Smirnova — Wed, 18 Jan 2017 00:25:04 +0000

In reply to Sveta Smirnova. *used last time