Creating a repository took 26 s while Earth was busy issue #335 closed: completed andymccall opened this 9 days ago · 1 comment

What /new waits on (2026-10-26): app.repos.create → Repo::init_bare_unbound, which does 16 fsyncs (strace of oxy init --bare): each of ~8 small files is written atomically (fsync the temp file, rename, fsync the directory), .oxy is fsynced 4 times and objects/loose 3 times. The work itself takes under a millisecond. On ext4 with data=ordered, every fsync forces a journal commit that waits for the disk's other dirty data, which at 12:01 was a 6 GB image export. So each fsync waited about a second, and the create took 26 s. It also runs synchronously on an async worker, so other requests on that thread waited with it.

Proposed, in order of value:

1. Run heavy disk jobs on Earth (image exports and imports, big copies) under ionice -c3 nice -n19, so the hub's fsyncs are not stuck behind them. This was the cause here; I'll do it from now on.
2. Hub: run repos.create (and create_on, fork) in spawn_blocking, so a slow disk stalls only the request that writes, not the workers.
3. oxy-core: init makes its directory durable once at the end (fsync each new file, then each directory once), instead of after every file. That is about 8 fsyncs instead of 16; worth it only if creates are still slow under load after 1 and 2.