1996fanrui commented on code in PR #28652:
URL: https://github.com/apache/flink/pull/28652#discussion_r3550813004


##########
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/consumer/EndOfFetchedChannelStateEvent.java:
##########
@@ -0,0 +1,75 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements.  See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership.  The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License.  You may obtain a copy of the License at
+ *
+ *     http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+package org.apache.flink.runtime.io.network.partition.consumer;
+
+import org.apache.flink.core.memory.DataInputView;
+import org.apache.flink.core.memory.DataOutputView;
+import org.apache.flink.runtime.event.RuntimeEvent;
+
+/**
+ * Marks the tail of recovered buffers that the spill drain pushed into a 
{@link
+ * RecoverableInputChannel}. The consume path polls this sentinel to learn the 
exact moment all
+ * recovered buffers have been consumed; it is never delivered to the 
operator. It is distinct from
+ * {@link EndOfInputChannelStateEvent} (which terminates the {@link 
RecoveredInputChannel} read
+ * stream) so the two recovery handoffs cannot be confused.
+ */
+public class EndOfFetchedChannelStateEvent extends RuntimeEvent {
+
+    /** The singleton instance of this event. */
+    public static final EndOfFetchedChannelStateEvent INSTANCE =
+            new EndOfFetchedChannelStateEvent();
+
+    // ------------------------------------------------------------------------
+
+    // not instantiable
+    private EndOfFetchedChannelStateEvent() {}
+
+    // ------------------------------------------------------------------------
+
+    @Override
+    public void write(DataOutputView out) {
+        throw new UnsupportedOperationException(
+                "EndOfFetchedChannelStateEvent must be serialized via 
EventSerializer's dedicated"
+                        + " type-tag path, not reflective write().");
+    }
+
+    @Override
+    public void read(DataInputView in) {
+        throw new UnsupportedOperationException(
+                "EndOfFetchedChannelStateEvent must be deserialized via 
EventSerializer's dedicated"
+                        + " type-tag path, not reflective read().");
+    }
+
+    // ------------------------------------------------------------------------
+
+    @Override
+    public int hashCode() {
+        return 20250814;

Review Comment:
   Just a random number 😂



##########
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/consumer/LocalInputChannel.java:
##########
@@ -504,18 +772,33 @@ void releaseAllResources() throws IOException {
         if (!isReleased) {
             isReleased = true;
 
+            upstreamReady.completeExceptionally(new 
CancelTaskException("Channel released."));
+
+            // Recovery will never be consumed on a released channel; unblock 
anyone gating on it.
+            stateConsumedFuture.complete(null);

Review Comment:
   After looking into the code, `completeExceptionally(...)` makes more sense.



##########
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/consumer/LocalInputChannel.java:
##########
@@ -113,48 +159,211 @@ public LocalInputChannel(
 
         this.partitionManager = checkNotNull(partitionManager);
         this.taskEventPublisher = checkNotNull(taskEventPublisher);
-        this.channelStatePersister = new ChannelStatePersister(stateWriter, 
getChannelInfo());
-
-        // Migrate recovered buffers from RecoveredInputChannel if provided.
-        // These buffers have been filtered but not yet consumed by the Task.
-        if (!initialRecoveredBuffers.isEmpty()) {
-            final int expectedCount = initialRecoveredBuffers.size();
-            // Sequence number starts at Integer.MIN_VALUE, consistent with 
RecoveredInputChannel.
-            int seqNum = Integer.MIN_VALUE;
-            while (!initialRecoveredBuffers.isEmpty()) {
-                Buffer buffer = initialRecoveredBuffers.poll();
-                // Determine next data type based on the next buffer in the 
queue
-                Buffer.DataType nextDataType =
-                        initialRecoveredBuffers.isEmpty()
-                                ? Buffer.DataType.NONE
-                                : initialRecoveredBuffers.peek().getDataType();
-                // buffersInBacklog is set to 0 as these are recovered buffers
-                BufferAndBacklog bufferAndBacklog =
-                        new BufferAndBacklog(buffer, 0, nextDataType, 
seqNum++);
-                toBeConsumedBuffers.add(bufferAndBacklog);
+        this.channelStatePersister =
+                new ChannelStatePersister(checkNotNull(stateWriter), 
getChannelInfo());
+        this.inRecovery = needsRecovery;
+        this.bufferManager =
+                needsRecovery
+                        ? new 
BufferManager(inputGate.getMemorySegmentProvider(), this, 0, true)
+                        : null;
+        this.networkBuffersPerChannel = networkBuffersPerChannel;
+        this.needsRecovery = needsRecovery;
+        if (!needsRecovery) {
+            stateConsumedFuture.complete(null);
+        }
+    }
+
+    @Override
+    void setup() throws IOException {
+        if (needsRecovery && networkBuffersPerChannel > 0) {
+            bufferManager.requestExclusiveBuffers(networkBuffersPerChannel);
+        }
+    }
+
+    // ------------------------------------------------------------------------
+    // RecoverableInputChannel implementation
+    // ------------------------------------------------------------------------
+
+    @Override
+    public void onRecoveredStateBuffer(Buffer buffer) {
+        boolean wasEmpty;
+        synchronized (recoveredBuffers) {
+            if (isReleased) {
+                buffer.recycleBuffer();
+                return;
             }
-            checkState(
-                    toBeConsumedBuffers.size() == expectedCount,
-                    "Buffer migration failed: expected %s buffers but got %s",
-                    expectedCount,
-                    toBeConsumedBuffers.size());
+            // Migrate recovered buffers from RecoveredInputChannel. These 
buffers have been
+            // filtered but not yet consumed by the Task.
+            wasEmpty = offerRecoveredBuffer(buffer);
+        }
+        if (wasEmpty) {
+            notifyChannelNonEmpty();
+        }
+    }
+
+    @Override
+    public void finishRecoveredBufferDelivery() throws IOException {
+        upstreamReady.join();
+        boolean wasEmpty;
+        synchronized (recoveredBuffers) {
+            checkState(inRecovery, "Recovery delivery already finished.");
+            // Append the sentinel after the last recovered buffer. The 
consume path flips out of
+            // recovery only once it polls this sentinel, guaranteeing all 
recovered buffers are
+            // consumed first.
+            wasEmpty =
+                    offerRecoveredBuffer(
+                            EventSerializer.toBuffer(
+                                    EndOfFetchedChannelStateEvent.INSTANCE, 
false));
+        }
+        if (wasEmpty) {
+            notifyChannelNonEmpty();
+        }
+    }
+
+    @Override
+    public Buffer requestRecoveryBufferBlocking() throws InterruptedException, 
IOException {
+        checkState(
+                bufferManager != null,
+                "requestRecoveryBufferBlocking called on a Local channel 
constructed with"
+                        + " needsRecovery=false");
+        upstreamReady.join();
+        return bufferManager.requestBufferBlocking();
+    }
+
+    @Override
+    public void insertRecoveryCheckpointBarrierIfInRecovery(long checkpointId) 
throws IOException {
+        boolean wasEmpty = false;
+        synchronized (recoveredBuffers) {
+            if (!isReleased && inRecovery) {
+                wasEmpty =
+                        offerRecoveredBuffer(
+                                EventSerializer.toBuffer(
+                                        new 
RecoveryCheckpointBarrier(checkpointId), false));
+            }
+        }
+        if (wasEmpty) {
+            notifyChannelNonEmpty();
+        }
+    }
+
+    /**
+     * Flips out of recovery the moment the consume path polls the {@code
+     * EndOfFetchedChannelStateEvent} sentinel, i.e. once all recovered 
buffers have been consumed.
+     * Live upstream data may flow again afterwards.
+     */
+    @Override
+    public void onRecoveredStateConsumed() {
+        synchronized (recoveredBuffers) {
+            checkState(inRecovery, "Recovery already finished.");
+            inRecovery = false;
         }
+        notifyChannelNonEmpty();
+        stateConsumedFuture.complete(null);
+    }
+
+    @Override
+    public CompletableFuture<Void> getStateConsumedFuture() {
+        return stateConsumedFuture;
+    }
+
+    /**
+     * Appends a recovered buffer (or {@code RecoveryCheckpointBarrier} / 
{@code
+     * EndOfFetchedChannelStateEvent} sentinel) to {@link #recoveredBuffers}.
+     *
+     * @return {@code true} iff {@link #recoveredBuffers} transitioned from 
empty to non-empty.
+     */
+    private boolean offerRecoveredBuffer(Buffer buffer) {
+        assert Thread.holdsLock(recoveredBuffers);
+        checkState(inRecovery, "Push into recovered buffers after recovery 
finished.");
+        boolean wasEmpty = recoveredBuffers.isEmpty();
+        recoveredBuffers.add(buffer);
+        return wasEmpty;
+    }
+
+    private int nextRecoverySequenceNumber() {
+        assert Thread.holdsLock(recoveredBuffers);
+        return recoverySequenceNumber++;
+    }
+
+    /**
+     * Walks {@link #recoveredBuffers} up to the {@link 
RecoveryCheckpointBarrier} sentinel matching
+     * {@code checkpointId}, retaining each pre-barrier recovered data buffer 
and removing the
+     * sentinel.
+     *
+     * @throws IOException if no sentinel matching {@code checkpointId} is 
found (the snapshot
+     *     protocol guarantees one must be present while the channel is in 
recovery).
+     */
+    private List<Buffer> collectPreRecoveryBarrier(long checkpointId) throws 
IOException {
+        assert Thread.holdsLock(recoveredBuffers);
+        List<Buffer> retained = new ArrayList<>();
+        try {
+            Iterator<Buffer> it = recoveredBuffers.iterator();
+            while (it.hasNext()) {
+                Buffer b = it.next();
+                if (isRecoveryCheckpointBarrier(b, checkpointId)) {
+                    it.remove();
+                    b.recycleBuffer();
+                    return retained;
+                }
+                if (b.isBuffer()) {
+                    retained.add(b.retainBuffer());

Review Comment:
   > What happens if b is a RecoveryCheckpointBarrier for a different 
checkpoint id than the one we're collecting for?
   
   As I understand it, this situation won't actually occur; defensive 
programming isn't currently being used here. If it were to happen, would you be 
inclined to throw an exception?
   
   > It falls through both branches here (not a data buffer, not the matching 
sentinel) and is left in the deque without being removed by the iterator. Is 
that intentional?
   
   Yes, it is intentional.
   
   CDR only snapshot data buffers, and removing `RecoveryCheckpointBarrier`. 
And does not snapshot and remove events.
   
   Regarding a normal unaligned checkpoint, it only snapshot data buffers, and 
never touch the event buffer as well.



##########
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/consumer/RecoveredInputChannel.java:
##########
@@ -115,23 +115,49 @@ public void setChannelStateWriter(ChannelStateWriter 
channelStateWriter) {
         this.channelStateWriter = checkNotNull(channelStateWriter);
     }
 
-    public final InputChannel toInputChannel() throws IOException {
-        Preconditions.checkState(
-                bufferFilteringCompleteFuture.isDone(), "buffer filtering is 
not complete");
-        if (!inputGate.isCheckpointingDuringRecoveryEnabled()) {
-            Preconditions.checkState(
-                    stateConsumedFuture.isDone(), "recovered state is not 
fully consumed");
+    public final InputChannel toInputChannel(boolean needsRecovery) throws 
IOException {
+        if (needsRecovery) {
+            return toInputChannelInRecovery();
+        }
+        synchronized (receivedBuffers) {
+            Preconditions.checkState(receivedBuffers.isEmpty(), "Received 
buffer should be empty.");
         }
 
-        // Extract remaining buffers before conversion.
-        // These buffers have been filtered but not yet consumed by the Task.
+        final InputChannel inputChannel = 
toInputChannelInternal(needsRecovery);
+        inputChannel.setup();
+        inputChannel.checkpointStopped(lastStoppedCheckpointId);
+        return inputChannel;
+    }
+
+    /**
+     * FLINK-38544 transitional: removed when the spilling backend lands. 
Creates the physical
+     * channel in recovery state and synchronously hands every queued 
recovered buffer over through
+     * the push interface. The legacy {@link EndOfInputChannelStateEvent} in 
the queue is dropped in
+     * translation; the {@link EndOfFetchedChannelStateEvent} sentinel takes 
its place. The sentinel
+     * is appended directly instead of via {@link
+     * RecoverableInputChannel#finishRecoveredBufferDelivery()} because that 
method waits for
+     * upstream readiness, which cannot happen while the mailbox thread is 
still converting channels
+     * (partitions are requested only after conversion).
+     */
+    private InputChannel toInputChannelInRecovery() throws IOException {
         final ArrayDeque<Buffer> remainingBuffers;
         synchronized (receivedBuffers) {
             remainingBuffers = new ArrayDeque<>(receivedBuffers);
             receivedBuffers.clear();
         }
 
-        final InputChannel inputChannel = 
toInputChannelInternal(remainingBuffers);
+        final InputChannel inputChannel = toInputChannelInternal(true);
+        inputChannel.setup();
+        final RecoverableInputChannel recoverableChannel = 
(RecoverableInputChannel) inputChannel;
+        for (Buffer buffer : remainingBuffers) {
+            if (isEndOfInputChannelStateEvent(buffer)) {
+                buffer.recycleBuffer();

Review Comment:
   No further entries after `EndOfInputChannelStateEvent` in remainingBuffers 
in theory.
   
   Adding a break or defensive check here works, or ignore here since this is 
only a transitional change.
   
   In the subsequent pr, recovered buffers are only consumed in local or remote 
input channel when cdr is enabled, so the `RecoveredInputChannel` never has any 
buffer or event, and this transitional change(`toInputChannelInRecovery` method 
) will be removed.
   
   I prefer to keep it as is, wdyt?



##########
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/consumer/LocalInputChannel.java:
##########
@@ -113,48 +159,211 @@ public LocalInputChannel(
 
         this.partitionManager = checkNotNull(partitionManager);
         this.taskEventPublisher = checkNotNull(taskEventPublisher);
-        this.channelStatePersister = new ChannelStatePersister(stateWriter, 
getChannelInfo());
-
-        // Migrate recovered buffers from RecoveredInputChannel if provided.
-        // These buffers have been filtered but not yet consumed by the Task.
-        if (!initialRecoveredBuffers.isEmpty()) {
-            final int expectedCount = initialRecoveredBuffers.size();
-            // Sequence number starts at Integer.MIN_VALUE, consistent with 
RecoveredInputChannel.
-            int seqNum = Integer.MIN_VALUE;
-            while (!initialRecoveredBuffers.isEmpty()) {
-                Buffer buffer = initialRecoveredBuffers.poll();
-                // Determine next data type based on the next buffer in the 
queue
-                Buffer.DataType nextDataType =
-                        initialRecoveredBuffers.isEmpty()
-                                ? Buffer.DataType.NONE
-                                : initialRecoveredBuffers.peek().getDataType();
-                // buffersInBacklog is set to 0 as these are recovered buffers
-                BufferAndBacklog bufferAndBacklog =
-                        new BufferAndBacklog(buffer, 0, nextDataType, 
seqNum++);
-                toBeConsumedBuffers.add(bufferAndBacklog);
+        this.channelStatePersister =
+                new ChannelStatePersister(checkNotNull(stateWriter), 
getChannelInfo());
+        this.inRecovery = needsRecovery;
+        this.bufferManager =
+                needsRecovery
+                        ? new 
BufferManager(inputGate.getMemorySegmentProvider(), this, 0, true)
+                        : null;
+        this.networkBuffersPerChannel = networkBuffersPerChannel;
+        this.needsRecovery = needsRecovery;
+        if (!needsRecovery) {
+            stateConsumedFuture.complete(null);
+        }
+    }
+
+    @Override
+    void setup() throws IOException {
+        if (needsRecovery && networkBuffersPerChannel > 0) {
+            bufferManager.requestExclusiveBuffers(networkBuffersPerChannel);
+        }
+    }
+
+    // ------------------------------------------------------------------------
+    // RecoverableInputChannel implementation
+    // ------------------------------------------------------------------------
+
+    @Override
+    public void onRecoveredStateBuffer(Buffer buffer) {
+        boolean wasEmpty;
+        synchronized (recoveredBuffers) {
+            if (isReleased) {
+                buffer.recycleBuffer();
+                return;
             }
-            checkState(
-                    toBeConsumedBuffers.size() == expectedCount,
-                    "Buffer migration failed: expected %s buffers but got %s",
-                    expectedCount,
-                    toBeConsumedBuffers.size());
+            // Migrate recovered buffers from RecoveredInputChannel. These 
buffers have been
+            // filtered but not yet consumed by the Task.
+            wasEmpty = offerRecoveredBuffer(buffer);
+        }
+        if (wasEmpty) {
+            notifyChannelNonEmpty();
+        }
+    }
+
+    @Override
+    public void finishRecoveredBufferDelivery() throws IOException {
+        upstreamReady.join();
+        boolean wasEmpty;
+        synchronized (recoveredBuffers) {
+            checkState(inRecovery, "Recovery delivery already finished.");
+            // Append the sentinel after the last recovered buffer. The 
consume path flips out of
+            // recovery only once it polls this sentinel, guaranteeing all 
recovered buffers are
+            // consumed first.
+            wasEmpty =
+                    offerRecoveredBuffer(
+                            EventSerializer.toBuffer(
+                                    EndOfFetchedChannelStateEvent.INSTANCE, 
false));
+        }
+        if (wasEmpty) {
+            notifyChannelNonEmpty();
+        }
+    }
+
+    @Override
+    public Buffer requestRecoveryBufferBlocking() throws InterruptedException, 
IOException {
+        checkState(
+                bufferManager != null,
+                "requestRecoveryBufferBlocking called on a Local channel 
constructed with"
+                        + " needsRecovery=false");
+        upstreamReady.join();
+        return bufferManager.requestBufferBlocking();
+    }
+
+    @Override
+    public void insertRecoveryCheckpointBarrierIfInRecovery(long checkpointId) 
throws IOException {
+        boolean wasEmpty = false;
+        synchronized (recoveredBuffers) {
+            if (!isReleased && inRecovery) {
+                wasEmpty =
+                        offerRecoveredBuffer(
+                                EventSerializer.toBuffer(
+                                        new 
RecoveryCheckpointBarrier(checkpointId), false));
+            }
+        }
+        if (wasEmpty) {
+            notifyChannelNonEmpty();
+        }
+    }
+
+    /**
+     * Flips out of recovery the moment the consume path polls the {@code
+     * EndOfFetchedChannelStateEvent} sentinel, i.e. once all recovered 
buffers have been consumed.
+     * Live upstream data may flow again afterwards.
+     */
+    @Override
+    public void onRecoveredStateConsumed() {
+        synchronized (recoveredBuffers) {
+            checkState(inRecovery, "Recovery already finished.");
+            inRecovery = false;
         }
+        notifyChannelNonEmpty();
+        stateConsumedFuture.complete(null);
+    }
+
+    @Override
+    public CompletableFuture<Void> getStateConsumedFuture() {
+        return stateConsumedFuture;
+    }
+
+    /**
+     * Appends a recovered buffer (or {@code RecoveryCheckpointBarrier} / 
{@code
+     * EndOfFetchedChannelStateEvent} sentinel) to {@link #recoveredBuffers}.
+     *
+     * @return {@code true} iff {@link #recoveredBuffers} transitioned from 
empty to non-empty.
+     */
+    private boolean offerRecoveredBuffer(Buffer buffer) {
+        assert Thread.holdsLock(recoveredBuffers);
+        checkState(inRecovery, "Push into recovered buffers after recovery 
finished.");
+        boolean wasEmpty = recoveredBuffers.isEmpty();
+        recoveredBuffers.add(buffer);
+        return wasEmpty;
+    }
+
+    private int nextRecoverySequenceNumber() {
+        assert Thread.holdsLock(recoveredBuffers);
+        return recoverySequenceNumber++;
+    }
+
+    /**
+     * Walks {@link #recoveredBuffers} up to the {@link 
RecoveryCheckpointBarrier} sentinel matching
+     * {@code checkpointId}, retaining each pre-barrier recovered data buffer 
and removing the
+     * sentinel.
+     *
+     * @throws IOException if no sentinel matching {@code checkpointId} is 
found (the snapshot
+     *     protocol guarantees one must be present while the channel is in 
recovery).
+     */
+    private List<Buffer> collectPreRecoveryBarrier(long checkpointId) throws 
IOException {
+        assert Thread.holdsLock(recoveredBuffers);
+        List<Buffer> retained = new ArrayList<>();
+        try {
+            Iterator<Buffer> it = recoveredBuffers.iterator();
+            while (it.hasNext()) {
+                Buffer b = it.next();
+                if (isRecoveryCheckpointBarrier(b, checkpointId)) {
+                    it.remove();
+                    b.recycleBuffer();
+                    return retained;
+                }
+                if (b.isBuffer()) {
+                    retained.add(b.retainBuffer());

Review Comment:
   For example, event can be EndOfOutputChannelStateEvent or 
EndOfFetchedChannelStateEvent



##########
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/consumer/RemoteInputChannel.java:
##########
@@ -596,6 +754,11 @@ public void onBuffer(Buffer buffer, int sequenceNumber, 
int backlog, int subpart
             throws IOException {
         boolean recycleBuffer = true;
 
+        // The first buffer from the producer proves the upstream reader is 
registered and the
+        // connection is live; release any recovery-side awaiter. On later 
buffers this is a cheap
+        // idempotent no-op (the latch count is already zero).
+        upstreamReady.countDown();

Review Comment:
   yes, we have ran data processing and checkpoint related benchmarks.
   
   No regression on data processing.



##########
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/consumer/RemoteInputChannel.java:
##########
@@ -123,6 +129,45 @@ public class RemoteInputChannel extends InputChannel {
 
     private final ChannelStatePersister channelStatePersister;
 
+    /**
+     * Whether the channel is still replaying recovered state. Recovered 
buffers delivered by the
+     * spill drain are appended directly to {@link #receivedBuffers}, so the 
consume path needs no
+     * recovery-specific branch. Starts {@code false} for channels that do not 
need recovery and is
+     * flipped to {@code false} the moment the consume path polls the {@code
+     * EndOfFetchedChannelStateEvent} sentinel that the drain appended after 
the last recovered
+     * buffer (see {@link #onRecoveredStateConsumed()}).
+     */
+    @GuardedBy("receivedBuffers")
+    private boolean inRecovery;
+
+    private final CompletableFuture<Void> stateConsumedFuture = new 
CompletableFuture<>();
+
+    /**
+     * Sequence number assigned to recovered buffers, starting at {@link 
Integer#MIN_VALUE},
+     * consistent with {@link RecoveredInputChannel}.
+     */
+    @GuardedBy("receivedBuffers")
+    private int recoverySequenceNumber = Integer.MIN_VALUE;
+
+    /**
+     * Ordinary (non-priority) upstream events received while recovery is 
still in progress. They
+     * cannot enter {@link #receivedBuffers} ahead of the recovered buffers, 
so they are stashed
+     * here and appended once recovery delivery finishes. Credit is suppressed 
during recovery, so
+     * the upstream can only send events (never data buffers) before {@link
+     * #finishRecoveredBufferDelivery()}.
+     */
+    @GuardedBy("receivedBuffers")
+    private final ArrayDeque<SequenceBuffer> recoveryEventStash = new 
ArrayDeque<>();
+
+    /**
+     * One-shot latch that opens once the upstream reader is registered and 
the connection is live
+     * (signalled by the first {@link #onBuffer} or by {@link 
#releaseAllResources()}).
+     * Recovery-side awaiters block on it before handing off; once open, {@link
+     * CountDownLatch#countDown()} on the hot path is a cheap idempotent 
no-op, unlike completing a
+     * {@code CompletableFuture}.
+     */
+    private final CountDownLatch upstreamReady = new CountDownLatch(1);

Review Comment:
   > And more fundamentally, why do we need to block on upstream readiness at 
all here (as opposed to LocalInputChannel, which apparently doesn't need to)?
   
   LocalInputChannel has same block mechanism. The purpose is avoid complex 
race condition or synchronization during recovery or starting.
   
   We could unify both to `CompletableFuture` or `CountDownLatch` to ensure 
consistency.



##########
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/consumer/RemoteInputChannel.java:
##########
@@ -188,6 +217,11 @@ void setExpectedSequenceNumber(int expectedSequenceNumber) 
{
         this.expectedSequenceNumber = expectedSequenceNumber;
     }
 
+    @VisibleForTesting
+    void completeUpstreamReadyForTest() {

Review Comment:
   Since exposing the latch/future would hand tests broader control than they 
need right now, would you be okay keeping this narrow trigger for the moment, 
and exposing the primitive later if a test ever needs finer-grained control?



##########
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/consumer/RemoteInputChannel.java:
##########
@@ -713,30 +889,110 @@ private void checkAnnouncedOnlyOnce(SequenceBuffer 
sequenceBuffer) {
     }
 
     /**
-     * Spills all queued buffers on checkpoint start. If barrier has already 
been received (and
-     * reordered), spill only the overtaken buffers.
+     * Persists inflight data on checkpoint start. During recovery, persists 
recovered buffers
+     * before the matching RecoveryCheckpointBarrier sentinel; after recovery, 
uses the normal
+     * remote-channel barrier sequence tracking and persists overtaken live 
buffers.
      */
     public void checkpointStarted(CheckpointBarrier barrier) throws 
CheckpointException {
-        synchronized (receivedBuffers) {
-            if (barrier.getId() < lastBarrierId) {
-                throw new CheckpointException(
-                        String.format(
-                                "Sequence number for checkpoint %d is not 
known (it was likely been overwritten by a newer checkpoint %d)",
-                                barrier.getId(), lastBarrierId),
-                        CheckpointFailureReason
-                                .CHECKPOINT_SUBSUMED); // currently, at most 
one active unaligned
-                // checkpoint is possible
-            } else if (barrier.getId() > lastBarrierId) {
-                // This channel has received some obsolete barrier, older 
compared to the
-                // checkpointId
-                // which we are processing right now, and we should ignore 
that obsoleted checkpoint
-                // barrier sequence number.
-                resetLastBarrier();
+        try {
+            List<Buffer> toPersist;
+            synchronized (receivedBuffers) {
+                if (inRecovery) {
+                    toPersist = collectPreRecoveryBarrier(barrier.getId());
+                } else {
+                    if (barrier.getId() < lastBarrierId) {
+                        // Currently, at most one active unaligned checkpoint 
is possible.
+                        throw new CheckpointException(
+                                String.format(
+                                        "Sequence number for checkpoint %d is 
not known (it was likely been overwritten by a newer checkpoint %d)",
+                                        barrier.getId(), lastBarrierId),
+                                CheckpointFailureReason.CHECKPOINT_SUBSUMED);
+                    } else if (barrier.getId() > lastBarrierId) {
+                        // This channel has received some obsolete barrier, 
older compared to the
+                        // checkpointId which we are processing right now, and 
we should ignore that
+                        // obsoleted checkpoint barrier sequence number.
+                        resetLastBarrier();
+                    }
+                    toPersist = getInflightBuffersUnsafe(barrier.getId());
+                }
+                channelStatePersister.startPersisting(barrier.getId(), 
toPersist);
+            }
+        } catch (IOException e) {
+            throw new CheckpointException(
+                    "Failed to extract recovered buffers for checkpoint " + 
barrier.getId(),
+                    CheckpointFailureReason.CHECKPOINT_DECLINED,
+                    e);
+        }
+    }
+
+    /**
+     * Walks {@link #receivedBuffers} (skipping priority events) up to the 
{@link
+     * RecoveryCheckpointBarrier} sentinel matching {@code checkpointId}, 
retaining each pre-barrier
+     * recovered data buffer and removing the sentinel. During recovery the 
upstream has no credit,
+     * so {@code receivedBuffers} holds only recovered buffers, sentinels, and 
priority events — no
+     * live data buffers.
+     *
+     * @throws IOException if no sentinel matching {@code checkpointId} is 
found (the snapshot
+     *     protocol guarantees one must be present while the channel is in 
recovery).
+     */
+    @GuardedBy("receivedBuffers")
+    private List<Buffer> collectPreRecoveryBarrier(long checkpointId) throws 
IOException {
+        assert Thread.holdsLock(receivedBuffers);
+        List<Buffer> retained = new ArrayList<>();
+        SequenceBuffer sentinel = null;
+        try {
+            Iterator<SequenceBuffer> it = receivedBuffers.iterator();
+            // Priority events are stored separately at the head and never 
carry recovered data.
+            Iterators.advance(it, receivedBuffers.getNumPriorityElements());
+            while (it.hasNext()) {
+                SequenceBuffer sb = it.next();
+                if (isRecoveryCheckpointBarrier(sb.buffer, checkpointId)) {
+                    sentinel = sb;
+                    break;
+                }
+                // Skip non-data events (e.g. the 
EndOfFetchedChannelStateEvent sentinel appended
+                // after the recovered buffers): only recovered data buffers 
are snapshotted.
+                if (sb.buffer.isBuffer()) {

Review Comment:
   > encounter the EndOfFetchedChannelStateEvent sentinel before finding the 
matching RecoveryCheckpointBarrier
   
   This case happens when checkpoint is trigger after all recovered buffers are 
delivered to input channel.
   
   and this comment is related to  
https://github.com/apache/flink/pull/28652#discussion_r3546635750. 
   
   



##########
flink-runtime/src/main/java/org/apache/flink/runtime/io/network/partition/consumer/RemoteInputChannel.java:
##########
@@ -123,6 +129,45 @@ public class RemoteInputChannel extends InputChannel {
 
     private final ChannelStatePersister channelStatePersister;
 
+    /**
+     * Whether the channel is still replaying recovered state. Recovered 
buffers delivered by the
+     * spill drain are appended directly to {@link #receivedBuffers}, so the 
consume path needs no
+     * recovery-specific branch. Starts {@code false} for channels that do not 
need recovery and is
+     * flipped to {@code false} the moment the consume path polls the {@code
+     * EndOfFetchedChannelStateEvent} sentinel that the drain appended after 
the last recovered
+     * buffer (see {@link #onRecoveredStateConsumed()}).
+     */
+    @GuardedBy("receivedBuffers")
+    private boolean inRecovery;
+
+    private final CompletableFuture<Void> stateConsumedFuture = new 
CompletableFuture<>();
+
+    /**
+     * Sequence number assigned to recovered buffers, starting at {@link 
Integer#MIN_VALUE},
+     * consistent with {@link RecoveredInputChannel}.
+     */
+    @GuardedBy("receivedBuffers")
+    private int recoverySequenceNumber = Integer.MIN_VALUE;
+
+    /**
+     * Ordinary (non-priority) upstream events received while recovery is 
still in progress. They
+     * cannot enter {@link #receivedBuffers} ahead of the recovered buffers, 
so they are stashed
+     * here and appended once recovery delivery finishes. Credit is suppressed 
during recovery, so
+     * the upstream can only send events (never data buffers) before {@link
+     * #finishRecoveredBufferDelivery()}.
+     */
+    @GuardedBy("receivedBuffers")
+    private final ArrayDeque<SequenceBuffer> recoveryEventStash = new 
ArrayDeque<>();
+
+    /**
+     * One-shot latch that opens once the upstream reader is registered and 
the connection is live
+     * (signalled by the first {@link #onBuffer} or by {@link 
#releaseAllResources()}).
+     * Recovery-side awaiters block on it before handing off; once open, {@link
+     * CountDownLatch#countDown()} on the hot path is a cheap idempotent 
no-op, unlike completing a
+     * {@code CompletableFuture}.
+     */
+    private final CountDownLatch upstreamReady = new CountDownLatch(1);

Review Comment:
   > Why is this a CountDownLatch here but a CompletableFuture for the 
equivalent state in LocalInputChannel?
   
   Reminder: this refactoring stems from your previous review comments.
   
   From the code comment, it was refactored from `CompletableFuture` to 
`CountDownLatch#countDown()` since hot path performance.
   
   The `CountDownLatch#countDown()` is cheaper than `CompletableFuture` when it 
is already done.



-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


Reply via email to