Status
Current state: Discuss
Discussion thread: here
JIRA: KAFKA-2120
Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).
Motivation
In old producer/consumer, we have a socket timeout associated with each request, the producer/consumer will send a request and then wait for response. If no response is received from a broker within specified timeout, the request will fail.
In the NetworkClient of new producer/consumer, currently we don't have a similar timeout for requests. Adding a client side request timeout in NetworkClient would be useful for the following reasons:
1. For KafkaProducer.close() and KafkaProducer.flush() we need the request timeout as implict timeout.
2. Currently the producer sends requests in a non-blocking way and does not wait for response. If a socket is disconnected the network client will fail the requests with a NetworkException. However, It might take very long for TCP connection to timeout. Adding a client side request timeout will help solve such kind of issues.
This KIP is only trying to solve the problem after a request has been constructed. For new producer specifically, there are situations where record batches can sit in accumulator for very long before they are drained out of RecordAccumulator (KAKFA-1788). Since the handling for those situations are more state-based instead of time-based, we are not trying to address them together in this KIP.
Public Interfaces
Add the following new configuration to org.apache.kafka.clients.producer.ProducerConfig and org.apache.kafka.clients.consumer.ConsumerConfig:
public static final String REQUEST_TIMEOUT_CONFIG = "request.timeout.ms";
private static final String REQUEST_TIMEOUT_DOC = "The configuration controls the maximum amount of time the producer will wait for the response of a request. If the "
+ "response is not received before the timeout elapses the producer will resend the request if necessary or fail the request if retries are exhausted."
If a request is timeout and retries have not been exhausted, the request will be retried.
When a request is timeout and retries are exhausted, an org.apache.kafka.common.errors.TimeoutException will be put in the returned future of a request and callback will be fired.
There is an existing configuration TIMEOUT_CONFIG, which is really a server side timeout specifying the maximum amount of time brokers can wait before send back the response. In order to avoid future confusion, for KafkaProducer, we might also want to change TIMEOUT_CONFIG to REPLICATION_TIMEOUT_CONFIG.
Proposed Changes
Because the NetworkClient works in an async way. We need to keep track of the send time for each request. So the implementation will the following:
- associate a time stamp with each inflight requests and retry
- The poll() will go through the inflight requests and expire those requests if necessary.
- The timeout for poll() will need to consider the request timeout for inflight requests.
The request timeout will be set to a reasonable value, say 60 seconds.
Actions after request timeout
When the request timeout has been reached, we do the following:
- Refresh metadata
- disconnect and create a new TCP connection.
- Retry the request on the new TCP connection.
In most cases, metadata refresh should be able to pick up the new leader if a broker is down. If a broker was not down but just slow, as long as request timeout is set to a reasonable value, we should not see dramatic increase in TCP connections when a broker was not down but just slow.
Plan to deprecate TIMEOUT_CONFIG
To replace TIMEOUT_CONFIG with REPLICATION_TIMEOUT_CONFIG, we will do the following;
- In 0.8.2.2 or 0.8.3, we will add REPLICATION_TIMEOUT_CONFIG but leave TIMEOUT_CONFIG there. If user sets TIMEOUT_CONFIG, we show a deprecation warning and use TIMEOUT_CONFIG to override REPLICATION_TIMEOUT_CONFIG.
- In 0.9, we will remove TIMEOUT_CONFIG from the configuration.
Compatibility, Deprecation, and Migration Plan
The changes should be backward compatible.
Rejected Alternatives
None