"dead dmlSink" error in sink workers will cause changefeed restart

[hicqu]

What did you do?

• run a changefeed;
• if some sink errors are met in fetchFromCache, they won't be treated as special SinkInternalErrors;
• so the whole changefeed will be restarted, instead of the processor sink module.

What did you expect to see?

No response

What did you see instead?

Whole changefeed will be restarted, instead of the processor sink module.

Versions of the cluster

Master branch.

Analysis

In v6.5.3 an improvement is introduced to reduce checkpoint lag when downstream sink fails. With the new mechanism, processor can re-establish failed sinks only instead of re-initialize the whole changefeed.

However, this bug breaks this mechanism, which will lead the whole changefeed get re-initialized in all cases. It will increase
checkpoint lag, especially for the cases with lots of tables and regions.

This bug only affects users who configured consistent.level = eventual.