Bug: TSO proxy will cause pd follower OOM

[guoxiangCN]

Bug Report

What did you do?

Enable tso proxy, the pd follower will OOM in several days.

What did you expect to see?

pd follower works well.

What did you see instead?

the pd follower will OOM in several days (2~5).

What version of PD are you using (pd-server -V)?

pd 5.4.3, but master also have this trouble.

[guoxiangCN]

The problematic code is here: https://github.com/tikv/pd/blob/master/pkg/utils/tsoutil/tso_dispatcher.go#L72

val, loaded := s.dispatchChs.LoadOrStore(req.getForwardedHost(), make(chan Request, maxMergeRequests))

every foward tso request will make a channel, which need 80KB!
maybe we should use a RWMutex+map or a map implementation with LoadOrCompute instead of sync.Map

[guoxiangCN]

PTAL @rleungx @disksing

[guoxiangCN]

[rleungx]

Introduced by #4085. The affected version starts from v5.3.0.

[guoxiangCN]

Introduced by #4085. The affected version starts from v5.3.0.

thanks for your reply! any repair plan ? Is the 5. x version still under maintenance, Is there a patch version that will continue to be released in the future?

[rleungx]

@guoxiangCN I have raised a PR to fix it. For TiDB EOL, please see https://www.pingcap.com/tidb-release-support-policy/ for detailed information. Once the PR is merged, we may cherry-pick the fix for other release branches, but 5.x may not.

[guoxiangCN]

@rleungx it seems that v5.4 supports unitl 2025-02-15, If the official can release a fixed version, that would be great！

[rleungx]

@rleungx it seems that v5.4 supports unitl 2025-02-15, If the official can release a fixed version, that would be great！

We prefer to upgrade to a new version for better support.