Depending on the configuration, the messages are either logged before they start or after the first response is received. Logging the message only when it fully completes in all cases would make the latency measurement more accurate and provide similar semantics for the data in all cases. This also solves the problem where queries that fail due to network errors are logged but not really executed.