v5.4.1: PD tso client Panic after inject 2s network loss to pd leader

[mayjiang0203]

Bug Report

tcms plan: [plan/784399] [endless-oltp-sybench-write-only-rel@v5.4.1]
tcms case: oltp-pd-rel-leader-network-loss

What did you do?

Inject 2s network loss to pd leader:

[2022/05/02 18:41:19.672 +08:00] [INFO] [chaos.go:93] ["Run chaos"] [name="pd network loss"] [selectors="[endless-oltp-tps-784399-1-760/tc-pd-0]"] [SelectorsRetainPolicy(selectors)=
"[endless-oltp-tps-784399-1-760/tc-pd-0]"] [targetSelectors="[nil]"] [TargetSelectorsRetainPolicy(targetSelectors)="[nil]"] [experimentSpec="NetworkLossSpec{Duration: "2s", Schedu
ler: ExperimentScheduler{Cron: "@every 36s"}, Loss: "100", Correlation: "70"}"]

What did you expect to see?

No panic in all components.

What did you see instead?

Panic happen in one tidb.

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

UTC Build Time: 2022-04-29 06:35:41
Rust Version: rustc 1.56.0-nightly (2faabf579 2021-07-27)
Enable Features: jemalloc mem-profiling portable sse test-engines-rocksdb cloud-aws cloud-gcp cloud-azure
Profile: dist_release
2022-05-02T17:49:18.363+0800 INFO k8s/client.go:107 it should be noted that a long-running command will not be interrupted even the use case has ended. For more informat
ion, please refer to https://github.com/pingcap/test-infra/discussions/129
./pd-server -V
Release Version: v5.4.1
Edition: Community
Git Commit Hash: 18098e9
Git Branch: heads/refs/tags/v5.4.1
UTC Build Time: 2022-04-29 01:18:56
2022-05-02T17:49:18.595+0800 INFO k8s/client.go:107 it should be noted that a long-running command will not be interrupted even the use case has ended. For more informat
ion, please refer to https://github.com/pingcap/test-infra/discussions/129
./tidb-server -V
Release Version: v5.4.1
Edition: Community
Git Commit Hash: cd60925897337d790469cc8293f0cbb3a2bdcb36
Git Branch: heads/refs/tags/v5.4.1
UTC Build Time: 2022-04-29 06:38:01
GoVersion: go1.16.4
Race Enabled: false
TiKV Min Version: v3.0.0-60965b006877ca7234adaced7890d7b029ed1306
Check Table Before Drop: false

[mayjiang0203]

/assign JmPotato
/severity Major

[mayjiang0203]

/label affects-5.0
/label affects-5.1
/label affects-5.2
/label affects-5.3
/label affects-5.4
/label affects-6.0

[JmPotato]

This bug will occur when the following conditions are met:

1. PD-0 resigns the leader.
2. PD-0 updates the TSO right after resetting it concurrently and leaves the non-empty TSO in memory.
3. PD-0 becomes the leader again.
4. PD-0 receives a TSO request after the leadership is available and before the TSO synchronization is done.
5. A smaller TSO will be generated with this request.
6. BOOM!!!