t386: pin the per-member drift baseline the fix's own tests left open
CI / contract (push) Successful in 37s
CI / build (push) Successful in 1m57s

The two tests merged with #386 both start with the member already BUSY, so a
single global drift baseline passes them. This one sleeps the host while nothing
is busy and only then starts a turn, which fails without the per-member map.
This commit is contained in:
Dai Ha
2026-09-10 06:59:42 +07:00
parent fd8650cda4
commit b9d09e044e
@@ -30,6 +30,7 @@ import java.util.concurrent.ScheduledFuture;
import java.util.concurrent.ScheduledThreadPoolExecutor;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicLong;
import java.util.concurrent.atomic.AtomicReference;
import java.util.function.BiConsumer;
import java.util.function.LongSupplier;
@@ -277,6 +278,58 @@ class FleetHealthMonitorTest {
}
}
/**
* fleetd #386 follow-up, added on merge. The fix carries a PER-MEMBER drift baseline, so drift
* from a sleep that happened BEFORE a member went busy is never charged to that member. The
* two tests shipped with the fix both start with the member already BUSY, so a single global
* baseline passes them — this one fails without the per-member map.
*
* <p>Order matters: the host sleeps while nothing is busy, and only then does a member take a
* turn. Its stall clock must start at zero.
*/
@Test void driftFromASleepBeforeAMemberWentBusyIsNotChargedToThatMember() {
Logger logger = (Logger) LoggerFactory.getLogger(FleetHealthMonitor.class);
ListAppender<ILoggingEvent> appender = new ListAppender<>();
appender.start();
logger.addAppender(appender);
try {
AtomicLong mono = new AtomicLong(0);
AtomicLong real = new AtomicLong(0);
AtomicReference<List<MemberSession>> roster = new AtomicReference<>(List.of());
FakeHerdr herdr = new FakeHerdr().withAgent("busy", "term_busy", "pane_busy", "tab_busy");
var scheduler = Executors.newSingleThreadScheduledExecutor();
AgentControl agents = new AgentControl(herdr);
FleetHealthMonitor monitor = new FleetHealthMonitor(agents, roster::get,
new MessageService(agents, new Injector(agents), new Rendezvous(), new InMemoryReplyInbox()),
scheduler, mono::get, real::get, 60, 600, (_, _) -> { });
monitor.tick(); // baseline, no members yet
// The host sleeps for 700s with nobody busy: the monotonic clock stands still.
real.set(TimeUnit.SECONDS.toNanos(700));
monitor.tick();
// Awake again. Only NOW does a member start a turn, with a fresh activity stamp taken
// from the monotonic clock. Both clocks advance together from here.
mono.set(TimeUnit.SECONDS.toNanos(10));
real.set(TimeUnit.SECONDS.toNanos(710));
roster.set(List.of(member("term_busy", MemberSession.State.BUSY,
0, TimeUnit.SECONDS.toNanos(10))));
monitor.tick();
mono.set(TimeUnit.SECONDS.toNanos(20));
real.set(TimeUnit.SECONDS.toNanos(720));
monitor.tick();
monitor.stop();
assertEquals(0, appender.list.stream().filter(event -> event.getFormattedMessage()
.contains("member=term_busy state=STALL_SUSPECTED")).count(),
"the member has been busy for 10s, not 710s — the earlier sleep is not its stall");
} finally {
logger.detachAppender(appender);
}
}
private static FleetHealthMonitor monitorWithClocks(FakeHerdr herdr, List<MemberSession> roster,
java.util.concurrent.ScheduledExecutorService scheduler, LongSupplier clock,
LongSupplier realtimeClock, long intervalSeconds, long workingSuspectAfterSeconds,