#345: a cache hit made four blocking std::fs calls on a tokio worker,
and the cache was on shfs (FUSE). A page asks for about fifty images at
once, so hits ran 11 ms at p50 and over a second at p90. Reading,
keeping and the reap job's trim now run in spawn_blocking. Production's
cache also moves to /mnt/fast in the compose file, outside the repo.
#349: two requests for the same uncached image both fetched it from its
publisher. A per-address lock now makes the second wait for the first,
then find the image kept. The lock is forgotten once its last holder is
done, whichever way the request ended.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
rays
cb4ee5a570