Add the ability to disable client heartbeat

[aofei]

What would you like to be added:

Please consider making juicefs mount --heartbeat 0 ... to disable the client heartbeat feature.

Why is this needed:

My use case:

$ juicefs format \
	--storage sqlite3 \
	--bucket /litestream/data-dir.jfs-bucket.db \
	sqlite3:///litestream/data-dir.jfs-meta.db \
	data-dir
$ litestream replicate -sync-interval 1 /litestream/data-dir.jfs-meta.db s3://mybkt/data-dir.jfs-meta.db &
$ litestream replicate -sync-interval 1 /litestream/data-dir.jfs-bucket.db s3://mybkt/data-dir.jfs-bucket.db &
$ juicefs mount sqlite3:///litestream/data-dir.jfs-meta.db /data

The files in /data rarely change, but I still expect JuiceFS to reflect to those two databases when it does, and Litestream to stream WALs to my S3 bucket in time. They actually do work pretty well together. Until I got my S3 bill.

Again, the files in /data are rarely change, so I'd like JuiceFS to be quiet until it does, which means it shouldn't be doing any writes to those two databases. But here comes the problem. It turns out that there is a heartbeat feature for the client here. The client always send heartbeats to meta storage at regular intervals (which defaults to 12s, and fixed in the [1, 60] interval). This means that by default every 12s Litestream will stream two WALs to my S3 bucket (I don't know why two). In other words, two S3 class A operations every 12s, or 432,000 class A operations every month, for nothing. And I have hundreds of the above setups currently active...

I know that for network meta storage, such as Redis, the client heartbeat feature is indispensable. But when it comes to SQLite meta storage, I believe this feature is unnecessary, given that there is only local disk I/O.

So I think it's reasonable to add the ability to disable the client heartbeat.

[solracsf]

@aofei there is a --heartbeat option.
Have you set it to 0?

https://juicefs.com/docs/community/command_reference/#mount

[aofei]

Hi @solracsf! And yes, I'm aware of the --heartbeat option. But, like I said, --heartbeat 0 is actually equivalent to --heartbeat 1.

... which defaults to 12s, and fixed in the [1, 60] interval...

[solracsf]

True

juicefs/pkg/meta/config.go

Lines 57 to 59 in c123b05

	if c.Heartbeat < time.Second {
	logger.Warnf("heartbeat should not be less than 1 second")
	c.Heartbeat = time.Second

Indeed, for SQLite and BadgerDB 0 should be allowed.

[SandyXSD]

Fixed by #3471

[SandyXSD]

There is still some background updates with low frequency per volume, e.g.

juicefs/pkg/meta/base.go

Line 424 in f1c475d

if ok, err := m.en.setIfSmall("lastCleanupFiles", time.Now().Unix(), int64(time.Minute.Seconds())*9/10); err != nil {

[aofei]

Fixed by #3471

Thanks for the fix! ❤️

There is still some background updates with low frequency per volume, e.g.

juicefs/pkg/meta/base.go

Line 424 in f1c475d

if ok, err := m.en.setIfSmall("lastCleanupFiles", time.Now().Unix(), int64(time.Minute.Seconds())*9/10); err != nil {

I'd take that as acceptable, given that --no-bgjob is already there.

I'm going to close this issue.