Time-series ordered inserts fall back on unordered, one-at-a-time inserts, if an error is encountered while performing the ordered insert. In both the ordered and unordered insert, we perform the same check check for statements that may have already been executed within the context of retryable writes.

I don't believe that we account for the fact that we could double count executed statements this way if we fall back to the unordered case. For example, if we inserted 3 measurements with the second one having already been executed, and then errored on the third measurement, upon falling back to unordered inserts it seems that we would increment the retriedStatementsCount counter again, as if we were seeing it for the first time.