Improving PostgreSQL's mysterious SLRU subsystem - Thomas Munro - PGCon 2022

Improve PostgreSQL's SLRU subsystem, a mysterious and complex part of the database, with proposals for cache-friendly buffer pool design, intelligent replacement algorithms, checksums, and performance enhancements.

Key takeaways

Improving PostgreSQL’s SLRU subsystem involves turning SLRU into a cache-friendly buffer pool.
The current SLRU design can cause bugs and does not handle high contention scenarios.
The speaker worries about the buffer replacement policy, advocating for a more intelligent replacement algorithm.
The design of the main buffer pool was driven by the need to store arbitrary data, which led to complexity.
The SLRU subsystem has no checksums, which can make errors in the data invisible.
The speaker proposes implementing checksums in the buffer pool to detect and correct errors.
The background writer process can be slow and can cause delays in writing data to disk.
The speaker proposes improving the performance of the background writer.
The neconference track and ‘shared Nothing’ approach may be useful in improving the performance of PostgreSQL.
The C log is a fundamental part of the Postgres system.
The speaker has created a commit-fest entry and a mailing list thread to discuss improving the SLRU subsystem.
The issue of how to handle raw pages is still open.
The speaker proposes using atomix for locking, which may simplify the buffer pool locking mechanism.
The buffer pool should benefit from future developments in storage and I/O.

Improving PostgreSQL's mysterious SLRU subsystem - Thomas Munro - PGCon 2022

More talks