Fix(openfile): avoid returning stale slice caches

[polyrabbit]

Some users cache opened files for a long time for performance reasons, which causes them read stale slices from openfiles cache - we only check expiration when opening a file, not reading slices.

One scenario is one client opens a file and reads occasionally - the read slices are cached forever. Another client compacts the file and old slices are deleted after trash-days. Now when the first client reads the file again, it gets stale slices. The penalty of false positive cache is one-second delay of a successful read. We could avoid this by applying expire to slices cache as well.

juicefs/pkg/chunk/cached_store.go

Lines 711 to 720 in ad6cf90

	for err != nil && tried < 2 {
	time.Sleep(time.Second * time.Duration(tried*tried))
	if tried > 0 {
	logger.Warnf("GET %s: %s; retrying", key, err)
	store.objectReqErrors.Add(1)
	start = time.Now()
	}
	in, err = store.storage.Get(key, 0, -1, object.WithRequestID(&reqID), object.WithStorageClass(&sc))
	tried++
	}

[SandyXSD]

The lastCheck is used to indicate the check time of attribute, so should not be updated here.

[polyrabbit]

@SandyXSD Do you think chunk cache should also obey the open-cache expire flag? If not, I'm OK to close this PR and keep it privately.

[davies]

The current behavior is 1 seconds delay in the case of stale slice (rarely), so it's better to remove the retry or retry without delay.

[polyrabbit]

Agreed, I suppose it's better to remove the inner retry. The current read retry is twice as user expected.

[polyrabbit]

Taken over by #5199 as discussed.