Merge pull request 'De-flake TestDrainPendingSerializesPerHost (CI stability)' (#33) from fix-flaky-server-http-tests into main

Reviewed-on: #33

internal/server/http/pending_drain_test.go

@@ -512,11 +512,27 @@ func TestDrainPendingSerializesPerHost(t *testing.T) {
 	// Connect the agent so DrainPending can dispatch.
 	c := agentDial(t, srv, ts, hostID, token)
 	sendHello(t, c, "serialise-host")
-	// Drain the on-hello goroutine's pass first (no pending rows yet),
+	// Wait for the on-hello push to settle.
-	// then wait for the schedule.set so the connection is fully settled.
 	_ = drainUntil(t, c, api.MsgScheduleSet)
 
-	// Insert 5 pending rows now that the on-hello drain has already run.
+	// A real agent is always in a read loop. Keep this test client
+	// reading in the background for the rest of the test: without an
+	// active reader the server-side conn can be dropped under parallel
+	// load, which unregisters it from the hub and makes DrainPending
+	// no-op (conn == nil) — the historical source of this test's
+	// flakiness (it would observe 0 or a partial drain). The reader also
+	// consumes the command.run envelopes our drains emit.
+	readerCtx, stopReader := context.WithCancel(context.Background())
+	defer stopReader()
+	go func() {
+		for {
+			if _, _, err := c.Read(readerCtx); err != nil {
+				return
+			}
+		}
+	}()
+
+	// Insert 5 due pending rows.
 	now := time.Now().UTC()
 	for i := range 5 {
 		pid := ulid.Make().String()
@@ -533,7 +549,8 @@ func TestDrainPendingSerializesPerHost(t *testing.T) {
 		}
 	}
 
-	// Spawn 10 goroutines all calling DrainPending concurrently.
+	// Fire 10 concurrent DrainPending calls. The per-host mutex must
+	// ensure each row is dispatched at most once (no double-dispatch).
 	var wg sync.WaitGroup
 	for range 10 {
 		wg.Add(1)
@@ -544,24 +561,26 @@ func TestDrainPendingSerializesPerHost(t *testing.T) {
 	}
 	wg.Wait()
 
-	// Drain any envelopes the agent received so we don't block below.
+	// Drain to completion. The fire-and-forget on-hello DrainPending
-	// We read with short timeouts and stop when the connection goes quiet.
+	// shares the same per-host mutex and can hold it during the burst,
-	drainDeadline := time.Now().Add(500 * time.Millisecond)
+	// leaving rows for a later pass — exactly how production drains
-	for time.Now().Before(drainDeadline) {
+	// (repeatedly, via the 30s tick / on reconnect). Re-drain until the
-		ctx, cancel := context.WithTimeout(context.Background(), 100*time.Millisecond)
+	// queue is empty; because every drain is still serialised, each row
-		_, _, err := c.Read(ctx)
+	// is dispatched at most once, so the exactly-5 job count below proves
-		cancel()
+	// there was no double-dispatch.
-		if err != nil {
+	deadline := time.Now().Add(5 * time.Second)
-			break
+	for countPendingForHost(t, st, hostID) > 0 && time.Now().Before(deadline) {
-		}
+		srv.DrainPending(context.Background(), hostID)
+		time.Sleep(10 * time.Millisecond)
 	}
 
-	// All 5 pending rows must be gone.
+	// All 5 pending rows must be drained.
 	if n := countPendingForHost(t, st, hostID); n != 0 {
-		t.Errorf("pending rows after concurrent drain: got %d, want 0", n)
+		t.Errorf("pending rows after drain-to-completion: got %d, want 0", n)
 	}
 
-	// Exactly 5 backup job rows (one per pending row), not 10+ from a race.
+	// Exactly 5 backup job rows (one per pending row) — never more, which
+	// would mean the per-host mutex failed to prevent double-dispatch.
 	var n int
 	_ = st.DB().QueryRow(
 		`SELECT COUNT(*) FROM jobs WHERE host_id = ? AND kind = 'backup' AND actor_kind = 'schedule'`,