During startup, we randomly sample records in the oplog to generate oplog stones. It is possible to sample multiple identical records as oplog stone candidates during startup.

This would normally be pretty hard, but if the size storer is very inaccurate, this may cause us to sample more records than necessary, increasing the probability of sampling the same record multiple times.

If two identical oplog stones do exist, an oplog record may be truncated in one iteration. In a subsequent iteration, the server may crash when it discovers that the last record in a stone ending in the same record does not exist.