regarding PG on ZFS performance

Just re-ran some older tests with the current version of Ubunutu & ZFS (but older kernel thanks to a multi-way incompatibility with other things). Results are that with proper tuning, ZFS RAIDZ1 on 4 NVMe drives gives higher TPS on pgbench at scale 10,000 than XFS on one of the same NVMe--but the initial population of the db takes 25% longer.

Proper tuning: PG full_page_writes off (for ZFS, on for NVMe); ZFS lz4 compression, 64K recordsize, relatime

db created by: pgbench -i -s 10000 --foreign-keys test

benchmarked as: pgbench -c 100 -j 4 -t 1000 test

NVMe: 31,804 TPS
RAIDZ1: 50,228 TPS

Some other notes:

- the situation is reversed, single NVMe is faster when using 10 connections instead of 100
- these tests are all from within containers running on Kubernetes--pg server and client in same container, connected over domain sockets
- 256GB and 48 CPU pod limits--running where there's still the cgroup double-counting bug, so CPU is theoretically throttled to ~24, leaving ~20 to PG server
- the container is actually getting very slightly throttled at barely over 20 CPU--so not sure if it's CPU-bound or IO-bound
- PG settings are set up for a larger database, shared_buffers, work_mem, parallel workers, autovacuum, etc
- I'd read that because of the way ZFS handles RAIDZ1 compared to RAID5, that performance probably didn't suffer relative to RAID10, and this is the case--tests with ZFS RAID10 on the same drives were a tiny bit slower (2-3%) than RAIDZ1 for TPS, but a bit faster on initial population (6-8%)
- as an aside, WekaFS (https://www.aspsys.com/solutions/storage-solutions/weka-io/) is about 10% faster than RAIDZ1 (both TPS and initial fill)

I hope that experience from someone who actually bothered to read up on how to configure ZFS for PG can put to rest some "ZFS is too slow" misinformation. I am certain that ZFS is not nearly the fastest for all configurations (for instance, I am unable to configure the 4 NVMe drives into a hardware RAID10 to test, and it seems that ZFS may not scale well to larger numbers of disks) but "too slow to ever be consider for serious work" is flat-out wrong.

--
Scott Ribe
scott_ribe(at)elevated-dev(dot)com
https://www.linkedin.com/in/scottribe/