This perspective completely ignores the use case of off-the-shelf software where you have limited or no capability to add "application metrics". Logs are invaluable in that scenario, which is arguably the majority of software deployments.
I've done this for Cassandra. While there are tons of jmx metrics exposed for Cassandra, last I checked the recommended way of logging compaction times was to tail the log. I even have a little project that tails logs and sends timestamps and values to graphite based on regexp:
