I met a check failed here:
incubator-doris/be/src/olap/cumulative_compaction.cpp

Lines 122 to 123 in 76a04de

| | CHECK(candidate_rowsets.size() == transient_size) |
| | << "tablet: " << _tablet->full_name() << ", "<< candidate_rowsets.size() << " vs. " << transient_size; |

The root cause is the wrong config:
config::min_cumulative_compaction_num_singleton_deltas > max_cumulative_compaction_num_singleton_deltas
So in pick_input_rowset(), we may get an _input_rowsets(size>= max_compaction_score). And clear _input_rowsets in here:
incubator-doris/be/src/olap/cumulative_compaction_policy.cpp

Lines 359 to 361 in a7422ee

| | if (last_delete_version->first == -1 && *compaction_score < min_compaction_score) { |
| | input_rowsets->clear(); |
| | } |

because min_compaction_score > max_compaction_score.

So we may get an empty _input_rowsets & transient_size=max_compaction_score<condidate_size & _last_delete_version.first == -1. And then, go to the CHECK(candidate_rowsets.size() == transient_size).

There are two approaches to this.

1. check min&max_compaction_score in pick_input_rowset(), cuz they are mutable configs, and can't check when update config.
2. delete this CHECK. Actually, I don't understand why we need candidate_rowsets.size() == transient_size). Dose it means we need to ensure it must loop through the candidate_rowsets? If so, it should be more readable.