Problem

SERVER-123974 introduced a performance regression of 10–27% in insert throughput on the perf-mongo-perf-repl-all-feature-flags sys-perf variant, observed on the insert_read_commands task.

Root Cause

SERVER-123974 registers an appliedOpTimeObserver. After this, every primary write calls _setMyLastAppliedOpTimeAndWallTime => _appliedOpTimeDispatcher.notify() => _event.fetchAndAdd(1) + _event.notifyAll(). This triggers a futex_wake syscall. Although the background thread drains quickly (one atomic store) and re-blocks again, at ~300K inserts/second (ARM, replica set): 300K × 400ns = 120ms/sec of pure syscall overhead that previously cost nothing. That's a ~10–15% throughput regression.

Fix

Rate-limit OpTimeObserverDispatcher::notify() to fire at most once per wall-clock second by comparing the opTime seconds field against the seconds encoded in _pendingTs (the existing pending-timestamp field).