TiFlash Alert Rules

This document introduces the alert rules of the TiFlash cluster.

`TiFlash_schema_error`

Alert rule:
increase(tiflash_schema_apply_count{type="failed"}[15m]) > 0
Description:
When the schema apply error occurs, an alert is triggered.
Solution:
The error might be caused by some wrong logic. Contact TiFlash R&D for support.

Alert rule:
histogram_quantile(0.99, sum(rate(tiflash_schema_apply_duration_seconds_bucket[1m])) BY (le, instance)) > 20
Description:
When the probability that the apply duration exceeds 20 seconds is over 99%, an alert is triggered.
Solution:
It might be caused by the internal problems of the TiFlash TMT engine. Contact TiFlash R&D for support.

Alert rule:
histogram_quantile(0.99, sum(rate(tiflash_raft_read_index_duration_seconds_bucket[1m])) BY (le, instance)) > 3
Description:
When the probability that the read index duration exceeds 3 seconds is over 99%, an alert is triggered.
Note
read index is the kvproto request sent to the TiKV leader. TiKV region retries, busy store, or network problems might lead to long request time of read index.
Solution:
The frequent retries might be caused by frequent splitting or migration of the TiKV cluster. You can check the TiKV cluster status to identify the retry reason.

Alert rule:
histogram_quantile(0.99, sum(rate(tiflash_raft_wait_index_duration_seconds_bucket[1m])) BY (le, instance)) > 2
Description:
When the probability that the waiting time for Region Raft Index in TiFlash exceeds 2 seconds is over 99%, an alert is triggered.
Solution:
It might be caused by a communication error between TiKV and the proxy. Contact TiFlash R&D for support.