When the secondary falls off the oplog there are frequent schema changes in the diagnostic data (every few seconds) which results in very poor compression and much reduced retention period for diagnostic data.

The cause is that the replica set status infoMessage field with the string "could not find member to sync from" comes and goes periodically about every 5 seconds.

I believe this could be fixed by ignoring fields that do not contribute to ftdc data, such as strings, before comparing the schema of the current sample with the reference sample, instead of ignoring them after the comparison.