In order to not lose data intermittently when using log compaction, Kafka topics should be configured with a non-zero min.compaction.lag.ms. The default of 0 appears, in some circumstances, to result in delete events being compacted by their tombstone before the consumer can process the delete.

This isn't a bug in Debezium as much as it's a undocumented configuration land mine. There isn't another bug filed describing this problem, so I figured I'd put it in here.