data inconsistency after injection of cdc network loss

[fubinzh]

What did you do?

1. TiDB cluster with 3 cdc
2. Create mysql sink changefeed
3. Pause changefeed
4. Run sysbench oltp_update_non_index prepare

sysbench --db-driver=mysql --mysql-host=upstream-tidb.cdc-testbed-tps-1816047-1-21 --mysql-port=4000 --mysql-user=root --mysql-db=workload --tables=100 --table-size=500000 --create_secondary=off --time=3600 --threads=100 oltp_update_non_index prepare

5. resume changefeed
6. run oltp_update_non_index run, and at the same time inject randon cdc node network loss for 10s every 30 minutes, totally 4 chaos injection is done in 2 hours.

sysbench --db-driver=mysql --mysql-host=upstream-tidb.cdc-testbed-tps-1816047-1-21 --mysql-port=4000 --mysql-user=root --mysql-db=workload --tables=100 --table-size=500000 --create_secondary=off --time=3600 --threads=100 oltp_update_non_index run

7. Send finishmark and Wait finishmark sync to downstream
8. Do data consistency check b/w upstream and downstream.

What did you expect to see?

Data consistency should pass

What did you see instead?

Data inconsistency seen. One table out of 100 tables in total are inconsistent, only 2 rows are different.

Versions of the cluster

[2023/07/04 23:09:33.079 +00:00] [INFO] [version.go:47] ["Welcome to Change Data Capture (CDC)"] [release-version=v7.3.0-alpha] [git-hash=41fda209483e6d0d94cd5e09ce784b692b2a614a] [git-branch=heads/refs/tags/v7.3.0-alpha] [utc-build-time="2023-07-04 11:03:24"] [go-version="go version go1.20.5 linux/amd64"] [failpoint-build=false]

[fubinzh]

cdc log:
ticdc_0_1816047.log
ticdc_1_1816047.log
ticdc_2_1816047.log

[fubinzh]

/found automation
/severity critical

[fubinzh]

This issue also seen when running case tikv_scale_kafka_sync, in which tikv/cdc scale up/in is done while changefeed running:
Steps:

1. create 2 kafka changefeed one for open protocol, one for canal json
2. run kafka consumer
3. run sysbench prepare "sysbench --db-driver=mysql --mysql-host=nslookup upstream-tidb.cdc-testbed-tps-1816831-1-572 | awk -F: '{print $2}' | awk 'NR==5' | sed s/[[:space:]]//g --mysql-port=4000 --mysql-user=root --mysql-db=workload --tables=30 --table-size=1000000 --create_secondary=off --debug=true --threads=20 --mysql-ignore-errors=2013,1213,1105,1205,8022,8027,8028,9004,9007,1062 oltp_write_only prepare"
4. scale tikv from 3 to 6, and scale cdc from 3 to 1
5. scale tikv from 6 to 3 nad cdc from 1 to 3 when running sysbench workload "sysbench --db-driver=mysql --mysql-host=nslookup upstream-tidb.cdc-testbed-tps-1816831-1-572 | awk -F: '{print $2}' | awk 'NR==5' | sed s/[[:space:]]//g --mysql-port=4000 --mysql-user=root --mysql-db=workload --tables=30 --table-size=1000000 --create_secondary=off --time=7200 --debug=true --threads=20 --mysql-ignore-errors=2013,1213,1105,1205,8022,8027,8028,9004,9007,1062 oltp_write_only run"
6. send finishmark, and do data consistency for open protocal when cdc sync is done.

[fubinzh]

This issue also seen during 7.1.1 release testing.