From: Justin Tobler Date: Tue, 17 Nov 2026 23:35:44 GMT Subject: Re: [PATCH 2/4] object-file: only flush batched fsync writes when enabled Message-ID: In-Reply-To: On 26/11/11 02:16PM, Patrick Steinhardt wrote: > On Mon, Nov 09, 2026 at 06:22:49PM -0600, Justin Tobler wrote: > > When committing an ODB transaction for the "files" backend, a full > > hardware flush is issued against a temporary file before the staged > > objects are migrated into the main object database. This flush acts as a > > barrier for objects that have been written with batched fsync, where > > `odb_transaction_files_fsync()` only issues a writeout request for each > > object instead of fully syncing it. > > Hmm, okay. The difference here is bascially that we don't issue a > hardware flush. So we know that the data is not in the page cache > anymore, but it may still only be cached by the disk. Performing the hardware sync here is really only intended for batch fsync Right, that's what writeout-only gives us for each object with batched fsync, and why the single hardware flush at commit time is needed as a barrier. That state only arises with batched fsync though. Otherwise, objects are either fully synced individually, in which case they are already durable and the extra flush is redundant, or not synced at all, in which case there is nothing to make durable. So the flush is only useful if there were writeout-only syncs. > > The flush is issued whenever the transaction has a temporary object > > directory though, regardless of whether batched fsync is actually > > enabled. For transactions configured for git-receive-pack(1), the > > temporary object directory is always created, so the flush is also > > performed when objects are synced individually or not synced at all. In > > those cases there are no pending writeout requests and the flush is > > unnecessary. Only issue the flush when batched fsync is enabled > > accordingly. > > So this is essentially a performance optimization? Yes and I also think its not really the correct behavior. Even is fsync is disabled, the fsync here happens. It doesn't really harm anything, but I think it makes sense to address as we begin to roll out transactions more widely. [snip] > I assume that we only support batched fsyncs for loose objects in the > first place? In other words, if a transaction were to write a bunch of > packfiles, it would not be possible to batch-fsync those right now? Correct, this mechanism was intended to batch fsync loose objects. > If so this new condition correct to me. But I think this should be > documented as part of the commit message, because it's an implicit > assumption. Makes sense. I'll update the commit message in the next version. Thanks, -Justin