Binary diffs for BSONColumn containing interleaved mode data can be very large as the offset point will be inside the first interleaved stream. This can lead to extremely large oplog entries for users with data models where the objects themselves contain a large amount of data or types that compresses poorly.

One solution could be to fallback and replicate such cases using the old uncompressed logical docDiff.

Another potential solution could be to limit the amount of elements we allow to be appended in interleaved mode, where we'd restart interleaved mode if we go over the threshold.