LWN: Comments on "An update on torn-write protection"

Errors are not stupid

ju3Ceemi — Mon, 19 May 2025 21:55:13 +0000

Is it ? This feature would be used by every databases, which would stop doing double writes with their WAL

I cannot understand how "brick the system" is a good solution

Slowing down and pushing an error on dmesg seems quite sane to me: users would have the information, and the performance would be .. what we have today (or a bit better, perhaps ?)

Errors are not stupid

tux3 — Thu, 10 Apr 2025 18:57:42 +0000

Yeah.. The users for this feature are going to be people who seriously care about perf. They will certainly want to fix the misconfiguration, and not just run in whatever slow path.

Silently taking a fallback path just makes the problem harder to diagnose.

This would be a good choice for a simple high-level API that targets a wide audience, where people might not want to be bothered with any low-level details, perhaps.

Errors are not stupid

Cyberax — Wed, 09 Apr 2025 18:20:30 +0000

> Hellwig said that the developers should fix the instrumentation of the system ""instead of creating stupid interfaces"".

Facepalm. Silently failing behind the scenes is the _worst_ possible interface. And the suggested "fix" is to add a source-code specific tracepoint? Really?

Non-reordered writes

epa — Wed, 09 Apr 2025 16:10:26 +0000

Could databases make use of a weaker guarantee, that the write may be written in pieces, but the earlier blocks are guaranteed to be written before (or at the same time as) the later ones?