Status

Current state: Under Discussion

...

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

Currently there is not a native way to shutdown and entire KStreams application from a StreamThread. This functionality would be useful for immediately halting processing to prevent data from being corrupted. Though this is best effort it may have faults in certain network partition scenarios.

...

• source/sink topic deleted. A handler will be exposed to a the users created for error source topics in KIP-662.
• Serialization or Deserialization failures. This can add another option to prevent rolling thread death.
  • currently they have the option of returning a Continue or Fail Enum. In the future we can expand this to include Shutdown and then throw ShutdownRequestedException in response
• Add another option in KIP-399 if failing and having only the thread fail is not comprehensive enough
• User specific cases such as 

  Jira
  server ASF JIRA serverId 5aa69414-a9e9-3523-82ec-879b028fb15b key KAFKA-4748

Public Interfaces

KafkaStreams.java

/**
     * Attempts to shutdown an application.
     * If there is an alive StreamThread it will succeed, if there is not it will fail
     *
     * @return Will return true if shutdown is initiated false if it is not possible.
     */
    public boolean initiateClosingAllClients()package org.apache.kafka.streams;
 
public enum StreamsUncaughtExceptionHandlerResponse {
    SHUTDOWN_STREAM_THREAD,
    SHUTDOWN_KAFKA_STREAMS_CLIENT,
    SHUTDOWN_KAFKA_STREAMS_APPLICATION;
}
 
public interface StreamsUncaughtExceptionHandler {
    UncaughtExceptionHandlerResponse handleUncaughtException(Thread thread, Throwable exception);
}
KafkaStreams.java/**
* Set the handler invoked when a {@link StreamsConfig#NUM_STREAM_THREADS_CONFIG internal thread} abruptly
* terminates due to an uncaught exception.
*
* @param eh the uncaught exception handler of type {@link StreamsUncaughtExceptionHandler} for all internal threads; {@code null} deletes the current handler
* @throws IllegalStateException if this {@code KafkaStreams} instance is not in state {@link State#CREATED CREATED}.
*/public void setStreamsUncaughtExceptionHandler(final StreamsUncaughtExceptionHandler eh) ;

Proposed Changes

We propose to add a new method streams specific uncaught exception handler that will cause all clients in the application to be shutdown when called. We achieve this adding initiateClosingAllClients to `KafkaStreams.java`  which will search for an alive stream thread to start the shutdown. If there is not an alive thread the method will return false indicating that the shutdown was not initiated. The method will be non-blocking but will stop the alive thread it found from processingdo the following:

SHUTDOWN_STREAM_THREAD:

• The current stream thread is shutdown and transits to state DEAD.

• The Kafka Streams client transits to ERROR if no other stream thread is alive.

SHUTDOWN_KAFKA_STREAMS_CLIENT

• All Stream Threads in the client are shutdown and they transit to state DEAD

• The Kafka Streams client transits to state NOT_RUNNING.

• The State directory cleaner thread stop 

• The RocksDB metrics recording thread is not shutdown.

SHUTDOWN_KAFKA_STREAMS_APPLICATION

• The shutdown is communicated to the other Kafka Streams clients through the rebalance protocol.

• All Stream Threads across the entire application are shutdown and they transit to state DEAD

• All Kafka Streams clients, i.e., the entire Kafka Streams application, is shutdown.

• All Kafka Streams clients transit to state NOT_RUNNING. 

• The State directory cleaner thread stop

• The RocksDB metrics recording thread is not shutdown.

The clients will transition to state NOT_RUNNING because that is the current behavior of close.

In order to communicate the shutdown request from one client to the others we propose to update the SubcriptionInfoData to include a short field which will encode an error code. The error will be propagated through the metadata during a rejoin event via the assignor. The actual shutdown will be handled by the StreamsRebalnceListener, this is where the INCOMPLETE_SOURCE_TOPIC_METADATA error can also be handled.

...

The SubcriptionInfoData will be upgraded to version 8 because we are adding a field for an error code to be propagated through the application.

No methods need to The pervious setUncaughtExceptionHandler will be deprecated in favor of this one.

Rejected Alternatives

• Two paths, Internal Error via exception and a request method for users to call
• Add a config option to shutdown when ever a user error is thrown - no flexible enough
• Throwing an Exception instead of shutdown Application