Presented by:

Screen shot 2019 02 12 at 14.50.50

Nikolay Samokhvalov

Nombox / Postgres.ai

More than 17 years of experience with various DBMSes, more than 13 years with PostgreSQL. M.S., MIPT and ISP RAS, specialty "Database systems".

Founder of Postgres.ai.

Founder of #RuPostgres (PostgreSQL user group for Russian-speaking users, the second in size on Meetup.com, 1900+ members).

Committee Chair for Database/Architecture Sections of Highload++, RITFest, and PGDay Russia.

Russian press contact at PostgreSQL Global Development Group.

Twitter: @postgresmen

No video of the event yet, sorry!

What is the optimal value of shared_buffers for your database and workload? 8 GiB? Or 16? Is it worth using 70% of RAM if you run Postgres 11 on a server with 512 GiB of memory?

Of course, you might check dozens of various values of shared_buffers, right on production, but if your database is mission-critical for your business, it’s not what you want.

To find the optimal value, we will conduct database experiments. We’ll discuss:

  • how to reproduce the production workload in the “lab” environment,
  • how to reduce the amount of time of a single experimental run,
  • what (latency? throughput? both? anything else?) and how we need to measure,
  • how the optimal value depends on workload, and finally,
  • how machine learning can help to predict the optimal value even without experiments!

During our journey to find the optimal shared_buffers value, we will use Nancy CLI — an open source framework we have built specifically to automate conducting numerous database experiments in a systematic way.

In the talk, we'll also dive into some technical details of shared buffers implementation in Postgres discussing why lower or higher shared_buffers values lead to much worse performance than one might expect.

Date:
2019 March 21 09:00
Duration:
50 min
Room:
Riverside Suite
Conference:
Postgres Conference
Language:
Track:
Use Cases
Difficulty:
Medium