BadgerDB control/fine-tune disk writes

gq0 · November 4, 2017, 1:58pm

Greetings everyone,

I just started using badger and was wondering if there is a way to control when the vlog is synced to disk. It looks like that immediately after a transaction is committed the sync starts.

I set SyncWrites to false, and MaxTableSize is set to 64 MB, but still after every second Set operation, approx. 170 bytes written to the DB, I got another disk write.

Can I fine tune when a disk write happens?

mrjn · November 6, 2017, 12:58am

github.com

dgraph-io/badger/blob/6c5c93205b5571967bd2efb350697947ad942221/value.go#L120


		nbr = int64(valsz)
	}
	y.NumReads.Add(1)
	y.NumBytesRead.Add(nbr)
	return buf, err
}

func (lf *logFile) doneWriting(offset uint32) error {
	// Sync before acquiring lock.  (We call this from write() and thus know we have shared access
	// to the fd.)
	if err := lf.fd.Sync(); err != nil {
		return errors.Wrapf(err, "Unable to sync value log: %q", lf.path)
	}
	// Close and reopen the file read-only.  Acquire lock because fd will become invalid for a bit.
	// Acquiring the lock is bad because, while we don't hold the lock for a long time, it forces
	// one batch of readers wait for the preceding batch of readers to finish.
	//
	// If there's a benefit to reopening the file read-only, it might be on Windows.  I don't know
	// what the benefit is.  Consider keeping the file read-write, or use fcntl to change
	// permissions.
	lf.lock.Lock()

This is the place where I see sync happening, irrespective of the value of SyncWrites. This would happen if the value log file gets filled up, which if you’re using default options is set to 1GB. This indicates that your values might be big.

If you’re seeing syncs happening more often (170 bytes), can you create a runnable Go code, and create a Github issue, so we can investigate.

Syncs happen only as they’re needed on SyncWrites=false. Any more fine tuning than that would lead to data corruption issues; so it’s better to investigate why syncs are happening as often as you say they’re.

Update: Note that disk writes != file syncs. Disk writes need to happen, but file syncs are what you control with SyncWrites option.

gq0 · November 6, 2017, 6:42pm

Thanks @mrjn for your reply.

I observed y.NumWrites and this happens for every write to the file descriptor. The doneWriting operation correctly happens if the ValueLogFileSize is reached. Sorry, I misunderstood y.NumWrites. Hence, the actual number of writes depends on the mmap configuration of the file descriptor, correct?

mrjn · November 7, 2017, 12:00am

We don’t do writes to the mmap area. We write directly to the underlying file. So, it all depends on the OS when it wants to flush this out to the disk.

github.com

dgraph-io/badger/blob/89689ef36cae26ae094cb5ea79b7400d839f2d68/value.go#L682


func (vlog *valueLog) write(reqs []*request) error {
	vlog.filesLock.RLock()
	curlf := vlog.filesMap[vlog.maxFid]
	vlog.filesLock.RUnlock()

	toDisk := func() error {
		if vlog.buf.Len() == 0 {
			return nil
		}
		vlog.elog.Printf("Flushing %d blocks of total size: %d", len(reqs), vlog.buf.Len())
		n, err := curlf.fd.Write(vlog.buf.Bytes())
		if err != nil {
			return errors.Wrapf(err, "Unable to write to value log file: %q", curlf.path)
		}
		y.NumWrites.Add(1)
		y.NumBytesWritten.Add(int64(n))
		vlog.elog.Printf("Done")
		atomic.AddUint32(&vlog.writableLogOffset, uint32(n))
		vlog.buf.Reset()

		if vlog.writableOffset() > uint32(vlog.opt.ValueLogFileSize) {

gq0 · November 7, 2017, 2:52pm

I see, thank @mrjn

So if I want to fine-tune this, I have to go to the OS/driver configuration. Thanks a lot for the support. Nice DB BTW

system · December 7, 2017, 2:52pm

This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.

Topic		Replies	Views
BadgerDB file size growing too much too fast Badger	2	1155	December 17, 2019
DB Space preallocation control Badger	1	101	March 11, 2024
Is a badger Commit lazy? even with SyncWrite = true? Badger	1	566	August 9, 2020
BadgerDB without touching disk? Badger	5	999	August 23, 2019
WARNING: LOG Compact FAILED with error: Writes are blocked, possibly due to DropAll or Close Badger	3	802	July 17, 2020

BadgerDB control/fine-tune disk writes

Related Topics