Non-goal: this line must never move into a {@code finally} block, and this method + * must never grow one around it. "Every time" above means every time the drain + * finishes, not every time this method exits. The absence of the line is the signal + * that the drain died, so a {@code finally} would destroy the signal and print confident + * partial counts in the same edit — the line would appear after a drain that threw, carrying + * whatever {@code tally} it had reached. Both halves of the value are lost at once. The line + * has to be the last statement of the successful path and reachable only from it. + * + *
This is written down because it is the obvious review comment ("shouldn't we always log + * the drain result?"), it sounds like thoroughness, and the paragraph above reads as an + * invitation to it. Raised by the fleet01 lead on 2026-09-12, from their 2026-09-10 incident: + * we only know that drain died on that host because it threw, and a + * {@code NoClassDefFoundError} reached the JVM's uncaught handler. A drain that hung on one + * session, or returned early on a condition rather than an exception, would leave no stack + * trace, no {@code ERROR} token and no priority — only a missing line. That makes the loud + * variant the one we have seen and the quiet variants the ones this line exists to catch. + * + *
Related: {@code released} and {@code abandoned} are counted incrementally inside {@link + * #drainSnapshot}'s loop and folded with {@link DrainTally#plus}, rather than derived from a + * collection read at the end, for the same reason. If a partial report is ever wanted it must + * be a different line with a different verb. One line must not serve both, or a reader cannot + * tell a finished drain from an interrupted one by its wording. */ void drainAll(long timeoutNanos) { long deadline = System.nanoTime() + timeoutNanos;