There is a data loss happening (for some of the column families) when we do the replay logs.
The bug seems to be from the fact that during replay-logs we only choose to replay
the logs from the maximumSequenceID across ALL the stores. This is wrong. If a
column family is ahead of others (because the crash happened before all the column
families were flushed), then we lose data for the column families that have not yet
The correct logic for replay should begin the replay from the minimum across the
maximum in each store.