Hello team, we are seeing a blocking issue in our cluster. Shared-nothing-3.3.18

compaction_manager.cpp

keeps scheduling compaction for the same tablet across its 3 replicas. Compaction succeeds on be-7 and be-9, but on be-5 it repeatedly throws:

Corruption: Bad page: too small size (0)
Internal error: reader get_next error: Corruption: Bad page: too small size (0)

From the logs: •

size_tiered_compaction_policy.cpp

keeps picking the same tablet again and again. •

CompactionManager

re-submits it each time since the compaction score remains high. •

vertical_compaction_task.cpp

fails every time with

state: COMPACTION_FAILED

. When we query the data, we get the same

Bad page: too small size (0)

error at query time as well. Question: Why doesn’t the BE mark this replica as bad and trigger a repair automatically, instead of getting stuck in an endless loop of compaction retries and failures? Two log files are attached which show the full issue cycle.