db_repl_test failed under TSAN. And one reason was because of concurrent writes of _pings in ElapsedTracker. But _pings is an int32_t and it should be harmless except for imprecise counts. So maybe we should fix ElapsedTracker to make it thread-safe, with AtomicWord or mutexes.

ElapsedTracker is not owned by repl. However, more data races could come up in the test once we fix ElapsedTracker. So it is hard to estimate how long it would take to make the test pass.