fix(cosmos): HTTP/2 PING handler must observe child-stream read activity, not just parent-pipeline frames

Follow-up to #49095.

Root cause: Http2PingHandler is installed on the parent (TCP) channel and only updated lastReadActivityNanos from parent-pipeline channelRead -- which sees PING ACK / SETTINGS / GOAWAY but NOT HEADERS / DATA frames. With reactor-netty's Http2MultiplexHandler in the parent pipeline, all application H2 traffic is demuxed onto per-stream child channels, so a connection serving thousands of QPS still looked idle to the PING handler, causing unnecessary PINGs and measurable P95 overhead under load.

Fix:
- New package-private AttributeKey<AtomicLong> LAST_CHILD_STREAM_READ_ACTIVITY_NANOS seeded on the parent channel in handlerAdded.
- Http2PingCloseRewrapHandler (already installed in every child stream's pipeline) overrides channelReadComplete to stamp the parent attribute via accumulateAndGet(now, Math::max) -- once per read cycle, monotonic.
- Http2PingHandler.maybeSendPing now uses effectiveLastReadActivityNanos = max(field, child attr) for both the idle-threshold check AND the outstanding-PING early-return: if child-stream reads arrived after an outstanding PING, clear pingOutstandingSinceNanos and reset consecutiveFailures (defense against middleboxes that drop PING-ACK but pass application H2 frames).
- Names use lastReadActivity / read activity -- both signals are inbound reads, never writes.

Tests: two new unit tests in Http2PingHandlerTest covering the suppression path and the outstanding-PING reset path; both bypass real time via reflection.

Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>

Author: jeet1995

sdk/cosmos/azure-cosmos-tests/src/test/java/com/azure/cosmos/implementation/http/Http2PingHandlerTest.java

@@ -16,6 +16,7 @@
 import java.lang.reflect.Method;
 import java.util.concurrent.ScheduledFuture;
 import java.util.concurrent.TimeUnit;
+import java.util.concurrent.atomic.AtomicLong;
 
 import static org.assertj.core.api.Assertions.assertThat;
 
@@ -25,7 +26,7 @@
  * Covers state transitions that do not require advancing real time -- the handler's
  * timeout / threshold logic reads {@code System.nanoTime()}, so the tests sidestep
  * the clock by pre-setting {@code pingOutstandingSinceNanos} (or
- * {@code lastActivityNanos}) via reflection:
+ * {@code lastReadActivityNanos}) via reflection:
  * <ul>
  *   <li>PING ACK with matching payload resets the failure counter (RFC 9113 §6.7 payload echo).</li>
  *   <li>PING ACK with mismatched payload is ignored (late ACK after timeout cannot mask degradation).</li>
@@ -212,7 +213,7 @@ public void write(ChannelHandlerContext ctx, Object msg, ChannelPromise promise)
 
         // First tick: force idle -> PING send attempt -> write fails.
         // Expect consecutiveFailures=1, channel still open, task still scheduled.
-        setField(handler, "lastActivityNanos", longAgo);
+        setField(handler, "lastReadActivityNanos", longAgo);
         invokeMaybeSendPing(handler, ctx);
         channel.runPendingTasks(); // run the write + listener on the embedded event loop
 
@@ -222,7 +223,7 @@ public void write(ChannelHandlerContext ctx, Object msg, ChannelPromise promise)
         assertThat((Object) getField(handler, "pingTask")).isNotNull();
 
         // Second tick: write fails again -> threshold reached -> task cancelled, channel closed.
-        setField(handler, "lastActivityNanos", longAgo);
+        setField(handler, "lastReadActivityNanos", longAgo);
         invokeMaybeSendPing(handler, ctx);
         channel.runPendingTasks();
 
@@ -274,6 +275,100 @@ public void ackTimeout_incrementsConsecutiveFailuresAndClosesAtThreshold() throw
         channel.finishAndReleaseAll();
     }
 
+    @Test(groups = "unit")
+    public void childStreamReadActivity_suppressesIdlePing() throws Exception {
+        // Regression test for the v4.81.0-beta.1 fix: with parent-pipeline reads
+        // stale (would breach the idle threshold) but child-stream reads recent
+        // (via the LAST_CHILD_STREAM_READ_ACTIVITY_NANOS attribute updated by
+        // Http2PingCloseRewrapHandler.channelReadComplete on each H2 child stream),
+        // the handler must NOT send a PING.
+        //
+        // Without the fix (Http2PingHandler reading only lastReadActivityNanos), a
+        // saturated connection serving thousands of QPS via H2 child streams would
+        // still look "idle" to the handler -- because HEADERS/DATA frames are
+        // demuxed by Http2MultiplexHandler and never surface on the parent's
+        // channelRead -- and PINGs would fire on every interval, adding measurable
+        // latency overhead under load (the symptom that motivated this PR).
+        Http2PingHandler handler = new Http2PingHandler(1, 1, 3);
+        EmbeddedChannel channel = new EmbeddedChannel(handler);
+        ChannelHandlerContext ctx = channel.pipeline().context(handler);
+
+        // Parent-pipeline read-timestamp is stale (60s ago).
+        long longAgo = System.nanoTime() - TimeUnit.SECONDS.toNanos(60);
+        setField(handler, "lastReadActivityNanos", longAgo);
+
+        // Child-stream read-activity attribute is fresh: simulate Http2PingCloseRewrapHandler
+        // having just stamped the parent attribute after demuxing an inbound HEADERS/DATA batch.
+        AtomicLong childReadActivity = channel.attr(Http2PingHandler.LAST_CHILD_STREAM_READ_ACTIVITY_NANOS).get();
+        assertThat(childReadActivity).as("handlerAdded should have seeded the child-read-activity attribute").isNotNull();
+        childReadActivity.set(System.nanoTime());
+
+        invokeMaybeSendPing(handler, ctx);
+        channel.runPendingTasks();
+
+        // PING must NOT have fired: effectiveLastReadActivityNanos = max(stale, fresh) = fresh.
+        assertThat((int) getField(handler, "pingsSent"))
+            .as("recent child-stream reads must suppress idle-PING")
+            .isZero();
+        assertThat((long) getField(handler, "pingOutstandingSinceNanos")).isZero();
+        assertThat(channel.isOpen()).isTrue();
+
+        channel.finishAndReleaseAll();
+    }
+
+    @Test(groups = "unit")
+    public void childStreamReadActivityAfterOutstandingPing_resetsFailureState() throws Exception {
+        // Regression test for the v4.81.0-beta.1 fix: when a PING is outstanding
+        // (pingOutstandingSinceNanos != 0) AND child-stream reads flow AFTER the
+        // PING was sent, the handler must clear the outstanding state and reset
+        // consecutiveFailures -- the connection is demonstrably healthy (H2 codec is
+        // still delivering application data), even though the PING ACK itself was
+        // never observed on the parent pipeline.
+        //
+        // Without the fix, a middlebox that drops PING-ACK frames (but lets
+        // application H2 frames through) would accumulate consecutiveFailures up to
+        // the threshold and close a healthy, actively-serving connection.
+        Http2PingHandler handler = new Http2PingHandler(1, 1, 3);
+        EmbeddedChannel channel = new EmbeddedChannel(handler);
+        ChannelHandlerContext ctx = channel.pipeline().context(handler);
+
+        // PING sent 5s ago is still outstanding; 2 prior consecutive failures recorded.
+        long pingSentAt = System.nanoTime() - TimeUnit.SECONDS.toNanos(5);
+        setField(handler, "pingsSent", 7);
+        setField(handler, "pingOutstandingSinceNanos", pingSentAt);
+        setField(handler, "consecutiveFailures", 2);
+
+        // Parent-pipeline read-timestamp is even OLDER than the outstanding PING so
+        // it cannot itself trigger the early-return -- proving the reset is driven
+        // by the child-stream attribute specifically.
+        long parentLastRead = System.nanoTime() - TimeUnit.SECONDS.toNanos(10);
+        setField(handler, "lastReadActivityNanos", parentLastRead);
+
+        // Child-stream read-activity stamped AFTER the PING was sent (Http2PingCloseRewrapHandler
+        // saw inbound HEADERS/DATA on a child stream just now).
+        AtomicLong childReadActivity = channel.attr(Http2PingHandler.LAST_CHILD_STREAM_READ_ACTIVITY_NANOS).get();
+        assertThat(childReadActivity).isNotNull();
+        childReadActivity.set(System.nanoTime());
+
+        invokeMaybeSendPing(handler, ctx);
+        channel.runPendingTasks();
+
+        // Early-return cleared outstanding state and the failure counter; no new
+        // PING was sent; channel is still healthy and open.
+        assertThat((long) getField(handler, "pingOutstandingSinceNanos"))
+            .as("child-stream reads after outstanding PING must clear pingOutstandingSinceNanos")
+            .isZero();
+        assertThat((int) getField(handler, "consecutiveFailures"))
+            .as("child-stream reads after outstanding PING must reset consecutiveFailures")
+            .isZero();
+        assertThat((int) getField(handler, "pingsSent"))
+            .as("early-return must NOT send a new PING")
+            .isEqualTo(7);
+        assertThat(channel.isOpen()).isTrue();
+
+        channel.finishAndReleaseAll();
+    }
+
     private static void invokeMaybeSendPing(Http2PingHandler handler, ChannelHandlerContext ctx) throws Exception {
         Method m = Http2PingHandler.class.getDeclaredMethod("maybeSendPing", ChannelHandlerContext.class);
         m.setAccessible(true);

sdk/cosmos/azure-cosmos/CHANGELOG.md

@@ -12,6 +12,7 @@
 #### Bugs Fixed
 * Fixed region name normalization for preferred and excluded regions — non-canonical inputs (e.g., `"westus3"`, `"WEST US 3"`) are now mapped to the canonical form. Also fixed a case-sensitive exclude-region check in PPCB reevaluate logic. - See [PR 49090](https://github.com/Azure/azure-sdk-for-java/pull/49090)
 * Fixed `UnsupportedOperationException` when using `readManyByPartitionKeys` for empty pages. - See [PR 49311](https://github.com/Azure/azure-sdk-for-java/pull/49311)
+* Fixed the HTTP/2 PING keepalive handler so it observes actual HTTP/2 stream traffic (HEADERS / DATA) when deciding whether a connection is idle. The original handler (see [PR 49095](https://github.com/Azure/azure-sdk-for-java/pull/49095)) tracked only parent-pipeline frames (PING ACK, SETTINGS, GOAWAY) plus its own PING sends, because `Http2MultiplexHandler` routes stream frames to child channels. As a result, connections that were continuously serving thousands of requests per second appeared idle and were PINGed on every interval, adding measurable latency overhead under load. The handler now observes inbound child-stream activity via a parent-channel attribute updated by `Http2PingCloseRewrapHandler.channelReadComplete`, and an in-flight PING is cleared (rather than charged toward the close threshold) when child-stream traffic flows after the PING was sent — preventing spurious closes on healthy connections where PING-specific frames are dropped by a middlebox.
 
 #### Other Changes
 * Added HTTP/2 PING keepalive (default ON) for Gateway service endpoints to detect silently-broken connections. - See [PR 49095](https://github.com/Azure/azure-sdk-for-java/pull/49095)

sdk/cosmos/azure-cosmos/src/main/java/com/azure/cosmos/implementation/http/Http2PingCloseRewrapHandler.java

@@ -6,6 +6,9 @@
 import io.netty.channel.ChannelHandler;
 import io.netty.channel.ChannelHandlerContext;
 import io.netty.channel.ChannelInboundHandlerAdapter;
+import io.netty.handler.codec.http2.Http2MultiplexHandler;
+
+import java.util.concurrent.atomic.AtomicLong;
 
 /**
  * Per-request HTTP/2 child-stream handler that translates a parent-TCP-channel close
@@ -40,6 +43,46 @@ final class Http2PingCloseRewrapHandler extends ChannelInboundHandlerAdapter {
 
     private Http2PingCloseRewrapHandler() {}
 
+    /**
+     * Records inbound H2 stream reads on the parent channel so the PING handler
+     * can recognize a busy connection as non-idle and skip the PING send.
+     * <p>
+     * {@code channelReadComplete} fires at the end of each inbound read cycle on
+     * the child channel -- so this method stamps the parent attribute exactly once
+     * per batch of HEADERS/DATA frames demuxed by {@link Http2MultiplexHandler},
+     * not once per frame. Under normal load this is plenty of granularity to
+     * suppress unnecessary PINGs without becoming a hot-path allocator.
+     * <p>
+     * Edge cases:
+     * <ul>
+     *   <li><b>Backpressure / autoRead off</b>: {@link Http2MultiplexHandler} may
+     *       buffer frames in the child's inbound queue and defer dispatch. In that
+     *       case {@code channelReadComplete} may fire later than the wire-arrival
+     *       time, but it WILL fire when the read cycle ends. Worst case: one benign
+     *       extra PING during the brief buffering window.</li>
+     *   <li><b>Non-H2 channels</b>: this handler is only installed on H2 child
+     *       streams (where {@code ch.parent() != null}), so the parent reference
+     *       above is non-null in normal operation. Defensive null-check kept for
+     *       safety if the install topology ever changes.</li>
+     * </ul>
+     * Uses {@code accumulateAndGet(now, Math::max)} so concurrent updates from
+     * sibling child streams remain monotonic (in practice child streams sharing a
+     * parent run on the same event loop, so contention is nil; the monotonic
+     * semantics document intent and future-proof against multi-event-loop
+     * scenarios).
+     */
+    @Override
+    public void channelReadComplete(ChannelHandlerContext ctx) throws Exception {
+        Channel parent = ctx.channel().parent();
+        if (parent != null) {
+            AtomicLong holder = parent.attr(Http2PingHandler.LAST_CHILD_STREAM_READ_ACTIVITY_NANOS).get();
+            if (holder != null) {
+                holder.accumulateAndGet(System.nanoTime(), Math::max);
+            }
+        }
+        super.channelReadComplete(ctx);
+    }
+
     @Override
     public void channelInactive(ChannelHandlerContext ctx) throws Exception {
         Channel parent = ctx.channel().parent();