[release-23.0] Change connection pool idle expiration logic (#19004) (#19014)

Signed-off-by: Arthur Schreiber <[email protected]>
Co-authored-by: vitess-bot[bot] <108069721+vitess-bot[bot]@users.noreply.github.com>

Author: vitess-bot[bot]

go/pools/smartconnpool/pool.go

@@ -781,35 +781,74 @@ func (pool *ConnPool[C]) closeIdleResources(now time.Time) {
 	mono := monotonicFromTime(now)
 
 	closeInStack := func(s *connStack[C]) {
-		// Do a read-only best effort iteration of all the connection in this
-		// stack and atomically attempt to mark them as expired.
-		// Any connections that are marked as expired are _not_ removed from
-		// the stack; it's generally unsafe to remove nodes from the stack
-		// besides the head. When clients pop from the stack, they'll immediately
-		// notice the expired connection and ignore it.
-		// see: timestamp.expired
-		for conn := s.Peek(); conn != nil; conn = conn.next.Load() {
-			if conn.timeUsed.expired(mono, timeout) {
-				pool.Metrics.idleClosed.Add(1)
+		conn, ok := s.Pop()
+		if !ok {
+			// Early return to skip allocating slices when the stack is empty
+			return
+		}
 
-				conn.Close()
-				pool.closedConn()
+		activeConnections := pool.Active()
+
+		// Only expire up to ~half of the active connections at a time. This should
+		// prevent us from closing too many connections in one go which could lead to
+		// a lot of `.Get` calls being added to the waitlist if there's a sudden spike
+		// coming in _after_ connections were popped off the stack but _before_ being
+		// returned back to the pool. This is unlikely to happen, but better safe than sorry.
+		//
+		// We always expire at least one connection per stack per iteration to ensure
+		// that idle connections are eventually closed even in small pools.
+		//
+		// We will expire any additional connections in the next iteration of the idle closer.
+		expiredConnections := make([]*Pooled[C], 0, max(activeConnections/2, 1))
+		validConnections := make([]*Pooled[C], 0, activeConnections)
+
+		// Pop out connections from the stack until we get a `nil` connection
+		for ok {
+			if conn.timeUsed.expired(mono, timeout) {
+				expiredConnections = append(expiredConnections, conn)
 
-				// Using context.Background() is fine since MySQL connection already enforces
-				// a connect timeout via the `db-connect-timeout-ms` config param.
-				c, err := pool.getNew(context.Background())
-				if err != nil {
-					// If we couldn't open a new connection, just continue
-					continue
+				if len(expiredConnections) == cap(expiredConnections) {
+					// We have collected enough connections for this iteration to expire
+					break
 				}
+			} else {
+				validConnections = append(validConnections, conn)
+			}
 
-				// opening a new connection might have raced with other goroutines,
-				// so it's possible that we got back `nil` here
-				if c != nil {
-					// Return the new connection to the pool
-					pool.tryReturnConn(c)
-				}
+			conn, ok = s.Pop()
+		}
+
+		// Return all the valid connections back to waiters or the stack
+		//
+		// The order here is not important - because we can't guarantee to
+		// restore the order we got the connections out of the stack anyway.
+		//
+		// If we return the connections in the order popped off the stack:
+		//   * waiters will get the newest connection first
+		//   * stack will have the oldest connections at the top of the stack.
+		//
+		// If we return the connections in reverse order:
+		//  * waiters will get the oldest connection first
+		//  * stack will have the newest connections at the top of the stack.
+		//
+		// Neither of these is better or worse than the other.
+		for _, conn := range validConnections {
+			pool.tryReturnConn(conn)
+		}
+
+		// Close all the expired connections and open new ones to replace them
+		for _, conn := range expiredConnections {
+			pool.Metrics.idleClosed.Add(1)
+
+			conn.Close()
+
+			err := pool.connReopen(context.Background(), conn, mono)
+			if err != nil {
+				pool.closedConn()
+				continue
 			}
+
+			pool.tryReturnConn(conn)
 		}
 	}
 

go/pools/smartconnpool/pool_test.go

@@ -1365,16 +1365,22 @@ func TestIdleTimeoutConnectionLeak(t *testing.T) {
 
 	// Try to get connections while they're being reopened
 	// This should trigger the bug where connections get discarded
+	wg := sync.WaitGroup{}
+
 	for i := 0; i < 2; i++ {
-		getCtx, cancel := context.WithTimeout(t.Context(), 50*time.Millisecond)
-		defer cancel()
+		wg.Go(func() {
+			getCtx, cancel := context.WithTimeout(t.Context(), 300*time.Millisecond)
+			defer cancel()
 
-		conn, err := p.Get(getCtx, nil)
-		require.NoError(t, err)
+			conn, err := p.Get(getCtx, nil)
+			require.NoError(t, err)
 
-		p.put(conn)
+			p.put(conn)
+		})
 	}
 
+	wg.Wait()
+
 	// Wait a moment for all reopening to complete
 	require.EventuallyWithT(t, func(c *assert.CollectT) {
 		// Check the actual number of currently open connections
@@ -1405,3 +1411,82 @@ func TestIdleTimeoutConnectionLeak(t *testing.T) {
 	assert.Equal(t, int64(0), state.open.Load())
 	assert.Equal(t, int64(4), state.close.Load())
 }
+
+func TestIdleTimeoutDoesntLeaveLingeringConnection(t *testing.T) {
+	var state TestState
+
+	ctx := context.Background()
+	p := NewPool(&Config[*TestConn]{
+		Capacity:    10,
+		IdleTimeout: 50 * time.Millisecond,
+		LogWait:     state.LogWait,
+	}).Open(newConnector(&state), nil)
+
+	defer p.Close()
+
+	var conns []*Pooled[*TestConn]
+	for i := 0; i < 10; i++ {
+		conn, err := p.Get(ctx, nil)
+		require.NoError(t, err)
+		conns = append(conns, conn)
+	}
+
+	for _, conn := range conns {
+		p.put(conn)
+	}
+
+	require.EqualValues(t, 10, p.Active())
+	require.EqualValues(t, 10, p.Available())
+
+	// Wait a bit for the idle timeout worker to refresh connections
+	assert.Eventually(t, func() bool {
+		return p.Metrics.IdleClosed() > 10
+	}, 500*time.Millisecond, 10*time.Millisecond, "Expected at least 10 connections to be closed by idle timeout")
+
+	// Verify that new connections were created to replace the closed ones
+	require.EqualValues(t, 10, p.Active())
+	require.EqualValues(t, 10, p.Available())
+
+	// Count how many connections in the stack are closed
+	totalInStack := 0
+	for conn := p.clean.Peek(); conn != nil; conn = conn.next.Load() {
+		totalInStack++
+	}
+
+	require.LessOrEqual(t, totalInStack, 10)
+}
+
+func BenchmarkPoolCleanupIdleConnectionsPerformanceNoIdleConnections(b *testing.B) {
+	var state TestState
+
+	capacity := 1000
+
+	p := NewPool(&Config[*TestConn]{
+		Capacity:    int64(capacity),
+		IdleTimeout: 30 * time.Second,
+		LogWait:     state.LogWait,
+	}).Open(newConnector(&state), nil)
+	defer p.Close()
+
+	// Fill the pool
+	connections := make([]*Pooled[*TestConn], 0, capacity)
+	for range capacity {
+		conn, err := p.Get(context.Background(), nil)
+		if err != nil {
+			b.Fatal(err)
+		}
+
+		connections = append(connections, conn)
+	}
+
+	// Return all connections to the pool
+	for _, conn := range connections {
+		conn.Recycle()
+	}
+
+	b.ResetTimer()
+
+	for b.Loop() {
+		p.closeIdleResources(time.Now())
+	}
+}