Hi Yerry,

> ext4, xfs are older file systems when comparing them with zfs. Zfs does a lot more and is more sophisticated and advanced. If you look only at performance then those older file systems are faster. But do you really need this in all situations?

That’s why I wrote please test for yourself.

> Testing postgres and zfs, you must tune some parameters, like the record size. A very good explanation is written by Klara
> https://klarasystems.com/articles/5-reasons-why-your-zfs-storage-benchmarks-are-wrong/

Thanks for thehint.

> An other example is checksums in zfs. Every time zfs reads data, it checks on corruption and it will repair it automatically. I do not think that ext4 wlll do that? But those actions takes iops. Maybe disable it when testing.
I’ve disabled checksums in PostgreSQL for that reason.

> With newer types of file storage (nvme’s, …) zfs has some performance problems. The main reason is that zfs is optimized for spinning disks. But the community is working on it.
Thank you.

Cheers,
Daniel

Comparing file systems with benchmarks tools like fio, pgbench, … is very difficult. (comparing apples and oranges)

ext4, xfs are older file systems when comparing them with zfs. Zfs does a lot more and is more sophisticated and advanced. If you look only at performance then those older file systems are faster. But do you really need this in all situations?

Testing postgres and zfs, you must tune some parameters, like the record size. A very good explanation is written by Klara

https://klarasystems.com/articles/5-reasons-why-your-zfs-storage-benchmarks-are-wrong/

An other example is checksums in zfs. Every time zfs reads data, it checks on corruption and it will repair it automatically. I do not think that ext4 wlll do that? But those actions takes iops. Maybe disable it when testing.

Also worth reading: https://bun.uptrace.dev/postgres/tuning-zfs-aws-ebs.html

With newer types of file storage (nvme’s, …) zfs has some performance problems. The main reason is that zfs is optimized for spinning disks. But the community is working on it.

Regards,
Yerry