poorbarcode commented on code in PR #24799:
URL: https://github.com/apache/pulsar/pull/24799#discussion_r2415448600
##########
pulsar-broker/src/main/java/org/apache/pulsar/broker/service/ServerCnxThrottleTracker.java:
##########
@@ -18,125 +18,439 @@
*/
package org.apache.pulsar.broker.service;
-import java.util.concurrent.atomic.AtomicIntegerFieldUpdater;
import lombok.extern.slf4j.Slf4j;
+import org.apache.pulsar.broker.ServiceConfiguration;
/**
- * Tracks the state of throttling for a connection. The throttling happens by
pausing reads by setting
- * Netty {@link io.netty.channel.ChannelConfig#setAutoRead(boolean)} to false
for the channel (connection).
- * <p>
- * There can be multiple rate limiters that can throttle a connection. Each
rate limiter will independently
- * call the {@link #incrementThrottleCount()} and {@link
#decrementThrottleCount()} methods to signal that the
- * connection should be throttled or not. The connection will be throttled if
the counter is greater than 0.
- * <p>
- * Besides the rate limiters, the connection can also be throttled if the
number of pending publish requests exceeds
- * a configured threshold. This throttling is toggled with the {@link
#setPendingSendRequestsExceeded} method.
- * There's also per-thread memory limits which could throttle the connection.
This throttling is toggled with the
- * {@link #setPublishBufferLimiting} method. Internally, these two methods
will call the
- * {@link #incrementThrottleCount()} and {@link #decrementThrottleCount()}
methods when the state changes.
+ * Manages and tracks throttling state for server connections in Apache Pulsar.
+ *
+ * <p>This class provides a centralized mechanism to control connection
throttling by managing
+ * multiple throttling conditions simultaneously. When throttling is active,
it pauses incoming
+ * requests by setting Netty's {@link
io.netty.channel.ChannelConfig#setAutoRead(boolean)} to
+ * {@code false} for the associated channel.
+ *
+ * <h3>Throttling Mechanism</h3>
+ * <p>The tracker maintains independent counters for different types of
throttling conditions
+ * defined in {@link ThrottleType}. A connection is considered throttled if
any of these
+ * conditions are active (counter > 0). The connection will only resume normal
operation
+ * when all throttling conditions are cleared.
+ *
+ * <h3>Supported Throttling Types</h3>
+ * <ul>
+ * <li><b>Connection-level:</b> Max pending publish requests, outbound
buffer limits</li>
+ * <li><b>Thread-level:</b> IO thread memory limits for in-flight
publishing</li>
+ * <li><b>Topic-level:</b> Topic publish rate limiting</li>
+ * <li><b>Resource Group-level:</b> Resource group publish rate limiting</li>
+ * <li><b>Broker-level:</b> Global broker publish rate limiting</li>
+ * <li><b>Flow Control:</b> Channel writability and cooldown rate
limiting</li>
+ * </ul>
+ *
+ * <h3>Reentrant vs Non-Reentrant Types</h3>
+ * <p>Some throttling types support multiple concurrent activations
(reentrant):
+ * <ul>
+ * <li>{@link ThrottleType#TopicPublishRate} - Reentrant because multiple
producers may share
+ * the same rate limiter which relates to the same topic</li>
+ * <li>{@link ThrottleType#ResourceGroupPublishRate} - Reentrant because
multiple producers may share
+ * the same rate limiter which relates to the same resource group</li>
+ * </ul>
+ * <p>Other types are non-reentrant and can only be activated once at a time.
The reentrant types
+ * use counters to track how many producers are affected by the same shared
rate limiter, while
+ * non-reentrant types use simple boolean states.
+ *
+ * <h3>Thread Safety</h3>
+ * <p>This class is designed to be used from a single thread (the connection's
IO thread)
+ * and is not thread-safe for concurrent access from multiple threads.
+ *
+ * <h3>Usage Example</h3>
+ * <pre>{@code
+ * ServerCnxThrottleTracker tracker = new ServerCnxThrottleTracker(serverCnx);
+ *
+ * // Mark connection as throttled due to rate limiting
+ * tracker.markThrottled(ThrottleType.TopicPublishRate);
+ *
+ * // Later, when rate limiting condition is cleared
+ * tracker.unmarkThrottled(ThrottleType.TopicPublishRate);
+ * }</pre>
+ *
+ * @see ThrottleType
+ * @see ThrottleRes
+ * @see ServerCnx
*/
@Slf4j
-final class ServerCnxThrottleTracker {
-
- private static final AtomicIntegerFieldUpdater<ServerCnxThrottleTracker>
THROTTLE_COUNT_UPDATER =
- AtomicIntegerFieldUpdater.newUpdater(
- ServerCnxThrottleTracker.class, "throttleCount");
-
- private static final AtomicIntegerFieldUpdater<ServerCnxThrottleTracker>
- PENDING_SEND_REQUESTS_EXCEEDED_UPDATER =
- AtomicIntegerFieldUpdater.newUpdater(
- ServerCnxThrottleTracker.class,
"pendingSendRequestsExceeded");
- private static final AtomicIntegerFieldUpdater<ServerCnxThrottleTracker>
PUBLISH_BUFFER_LIMITING_UPDATER =
- AtomicIntegerFieldUpdater.newUpdater(
- ServerCnxThrottleTracker.class, "publishBufferLimiting");
+public final class ServerCnxThrottleTracker {
+
private final ServerCnx serverCnx;
- private volatile int throttleCount;
- private volatile int pendingSendRequestsExceeded;
- private volatile int publishBufferLimiting;
+ private final int[] states = new int[ThrottleType.values().length];
+
+ /**
+ * Enumeration of different throttling conditions that can be applied to a
server connection.
+ *
+ * <p>Each type represents a specific resource constraint or rate limiting
condition
+ * that may require throttling the connection to maintain system stability
and fairness.
+ *
+ * <p>Some types support reentrant behavior (can be activated multiple
times concurrently),
+ * while others are non-reentrant (single activation only).
+ */
+ public static enum ThrottleType {
+ /**
+ * Throttling due to excessive pending publish requests on the
connection.
+ *
+ * <p>This throttling is activated when the number of in-flight
publish requests
+ * exceeds the configured limit. It helps prevent memory exhaustion
and ensures
+ * fair resource allocation across connections.
+ *
+ * <p><b>Type:</b> Non-reentrant
+ * <p><b>Configuration:</b> {@link
ServiceConfiguration#getMaxPendingPublishRequestsPerConnection()}
+ */
+ ConnectionMaxQuantityOfInFlightPublishing,
+ /**
+ * Throttling due to excessive memory usage by in-flight publish
operations on the IO thread.
+ *
+ * <p>This throttling is activated when the total memory used by
pending publish operations
+ * on a shared IO thread exceeds the configured limit. Multiple
connections may share the
+ * same IO thread, so this limit applies across all connections on
that thread.
+ *
+ * <p><b>Type:</b> Non-reentrant
+ * <p><b>Configuration:</b> {@link
ServiceConfiguration#getMaxMessagePublishBufferSizeInMB()}
+ */
+ IOThreadMaxBytesOfInFlightPublishing,
- public ServerCnxThrottleTracker(ServerCnx serverCnx) {
- this.serverCnx = serverCnx;
+ /**
+ * Throttling due to topic-level publish rate limiting.
+ *
+ * <p>This throttling is activated when publish operations exceed the
configured
+ * rate limits for a specific topic. Multiple producers on the same
topic may
+ * contribute to triggering this throttling condition.
+ *
+ * <p><b>Type:</b> Reentrant (supports multiple concurrent activations)
+ * <br><b>Reason for reentrancy:</b> Multiple producers may share the
same rate limiter
+ * which relates to the same topic. Each producer can independently
trigger throttling
+ * when the shared topic rate limiter becomes active, requiring a
counter to track
+ * how many producers are affected by the same rate limiter.
+ *
+ * <p><b>Configuration:</b> Topic-level publish rate policies
+ */
+ TopicPublishRate,
+
+ /**
+ * Throttling due to resource group-level publish rate limiting.
+ *
+ * <p>This throttling is activated when publish operations exceed the
configured
+ * rate limits for a resource group. Resource groups allow
fine-grained control
+ * over resource allocation across multiple topics and tenants.
+ *
+ * <p><b>Type:</b> Reentrant (supports multiple concurrent activations)
+ * <br><b>Reason for reentrancy:</b> Multiple producers may share the
same rate limiter
+ * which relates to the same resource group. Each producer can
independently trigger
+ * throttling when the shared resource group rate limiter becomes
active, requiring
+ * a counter to track how many producers are affected by the same rate
limiter.
+ *
+ * <p><b>Configuration:</b> Resource group publish rate policies
+ */
+ ResourceGroupPublishRate,
+ /**
+ * Throttling due to broker-level publish rate limiting.
+ *
+ * <p>This throttling is activated when publish operations exceed the
global
+ * broker-level rate limits. This provides a safety mechanism to
prevent
+ * the entire broker from being overwhelmed by publish traffic.
+ *
+ * <p><b>Type:</b> Non-reentrant
+ * <p><b>Configuration:</b> {@link
ServiceConfiguration#getBrokerPublisherThrottlingMaxMessageRate()}
+ * and {@link
ServiceConfiguration#getBrokerPublisherThrottlingMaxByteRate()}
+ */
+ BrokerPublishRate,
+
+ /**
+ * Throttling due to channel outbound buffer being full.
+ *
+ * <p>This throttling is activated when the Netty channel's outbound
buffer
+ * reaches its high water mark, indicating that the client cannot keep
up
+ * with the rate of outgoing messages. This prevents memory exhaustion
+ * and provides backpressure to publishers.
+ *
+ * <p><b>Type:</b> Non-reentrant
+ * <p><b>Reference:</b> PIP-434: Expose Netty channel configuration
WRITE_BUFFER_WATER_MARK
+ */
+ ConnectionOutboundBufferFull,
+
+ /**
+ * Throttling due to connection pause/resume cooldown rate limiting.
+ *
+ * <p>This throttling is activated during cooldown periods after a
connection
+ * has been resumed from a throttled state. It prevents rapid
oscillation
+ * between throttled and unthrottled states.
+ *
+ * <p><b>Type:</b> Non-reentrant
+ */
+ ConnectionPauseReceivingCooldownRateLimit
+ }
+
+ /**
+ * Enumeration representing the result of a throttling state change
operation.
+ *
+ * <p>This enum indicates what happened when a throttling condition was
marked or unmarked,
+ * helping callers understand whether the overall connection state changed
or if the
+ * operation was ignored.
+ */
+ enum ThrottleRes {
+ /**
+ * The operation resulted in a change to the overall connection
throttling state.
+ *
+ * <p>This occurs when:
+ * <ul>
+ * <li>The connection transitions from unthrottled to throttled
(first throttle type activated)</li>
+ * <li>The connection transitions from throttled to unthrottled
(last throttle type deactivated)</li>
+ * </ul>
+ *
+ * <p>When this result is returned, the connection's auto-read setting
will be updated
+ * accordingly to pause or resume request processing.
+ */
+ ConnectionStateChanged,
+
+ /**
+ * The operation changed the state of the specific throttle type but
did not affect
+ * the overall connection throttling state.
+ *
+ * <p>This occurs when:
+ * <ul>
+ * <li>A throttle type is activated, but the connection was already
throttled by other types</li>
+ * <li>A throttle type is deactivated, but the connection remains
throttled by other types</li>
+ * <li>A reentrant throttle type's counter is incremented or
decremented</li>
+ * </ul>
+ */
+ TypeStateChanged,
+
+ /**
+ * The operation was dropped because it would violate the throttle
type's constraints.
+ *
+ * <p>This occurs when:
+ * <ul>
+ * <li>Attempting to mark a non-reentrant throttle type that is
already active</li>
+ * <li>Attempting to unmark a throttle type that is not currently
active</li>
+ * <li>Attempting to unmark a reentrant throttle type with an
invalid counter state</li>
+ * </ul>
+ */
+ Dropped
}
/**
- * See {@link Producer#incrementThrottleCount()} for documentation.
+ * Checks if the connection is currently throttled by any throttle type.
+ *
+ * <p>This method examines all throttle type states and returns {@code
true}
+ * if any of them are active (counter > 0).
+ *
+ * @return {@code true} if any throttling condition is active, {@code
false} otherwise
*/
- public void incrementThrottleCount() {
- int currentThrottleCount =
THROTTLE_COUNT_UPDATER.incrementAndGet(this);
- if (currentThrottleCount == 1) {
- changeAutoRead(false);
+ private boolean hasThrottled() {
+ for (int stat : states) {
+ if (stat > 0) {
+ return true;
+ }
}
+ return false;
}
/**
- * See {@link Producer#decrementThrottleCount()} for documentation.
+ * Returns the total count of active throttling conditions across all
types.
+ *
+ * <p>This method sums up all the individual counters for each throttle
type,
+ * providing a measure of the overall throttling pressure on the
connection.
+ * For reentrant types, this includes the full counter value (not just 0
or 1).
+ *
+ * @return the total number of active throttling conditions
*/
- public void decrementThrottleCount() {
- int currentThrottleCount =
THROTTLE_COUNT_UPDATER.decrementAndGet(this);
- if (currentThrottleCount == 0) {
- changeAutoRead(true);
+ public int throttledCount() {
+ int i = 0;
+ for (int stat : states) {
+ i += stat;
}
+ return i;
}
- private void changeAutoRead(boolean autoRead) {
- if (isChannelActive()) {
+ /**
+ * Marks the connection as throttled for the specified throttle type.
+ *
+ * <p>This method activates throttling for the given type and may pause
the connection's
+ * request processing if this is the first active throttling condition.
For reentrant
+ * types ({@link ThrottleType#TopicPublishRate} and {@link
ThrottleType#ResourceGroupPublishRate}),
+ * this increments the counter. For non-reentrant types, this sets the
state to active.
+ *
+ * <p>If the connection transitions from unthrottled to throttled, this
method will
+ * set the Netty channel's auto-read to {@code false}, effectively pausing
incoming
+ * request processing.
+ *
+ * <p>Metrics are automatically recorded to track throttling events and
connection state changes.
+ *
+ * @param type the type of throttling condition to activate
+ * @throws IllegalArgumentException if type is null
+ *
+ * @see #unmarkThrottled(ThrottleType)
+ * @see ThrottleType
+ */
+ public void markThrottled(ThrottleType type) {
+ ThrottleRes res = doMarkThrottled(type);
+ recordMetricsAfterThrottling(type, res);
+ if (res == ThrottleRes.ConnectionStateChanged && isChannelActive()) {
if (log.isDebugEnabled()) {
- log.debug("[{}] Setting auto read to {}",
serverCnx.toString(), autoRead);
+ log.debug("[{}] Setting auto read to false",
serverCnx.toString());
}
- // change the auto read flag on the channel
- serverCnx.ctx().channel().config().setAutoRead(autoRead);
- }
- // update the metrics that track throttling
- if (autoRead) {
- serverCnx.getBrokerService().recordConnectionResumed();
- } else if (isChannelActive()) {
- serverCnx.increasePublishLimitedTimesForTopics();
- serverCnx.getBrokerService().recordConnectionPaused();
+ serverCnx.ctx().channel().config().setAutoRead(false);
}
}
- private boolean isChannelActive() {
- return serverCnx.isActive() && serverCnx.ctx() != null &&
serverCnx.ctx().channel().isActive();
+ /**
+ * Unmarks the connection as throttled for the specified throttle type.
+ *
+ * <p>This method deactivates throttling for the given type and may resume
the connection's
+ * request processing if this was the last active throttling condition.
For reentrant
+ * types ({@link ThrottleType#TopicPublishRate} and {@link
ThrottleType#ResourceGroupPublishRate}),
+ * this decrements the counter. For non-reentrant types, this clears the
active state.
+ *
+ * <p>If the connection transitions from throttled to unthrottled, this
method will
+ * set the Netty channel's auto-read to {@code true}, effectively resuming
incoming
+ * request processing.
+ *
+ * <p>Metrics are automatically recorded to track unthrottling events and
connection state changes.
+ *
+ * @param type the type of throttling condition to deactivate
+ * @throws IllegalArgumentException if type is null
+ *
+ * @see #markThrottled(ThrottleType)
+ * @see ThrottleType
+ */
+ public void unmarkThrottled(ThrottleType type) {
+ ThrottleRes res = doUnmarkThrottled(type);
Review Comment:
Added the check, see also
https://github.com/apache/pulsar/pull/24799#discussion_r2415448306
--
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]
