[KAFKA-4024] First metadata update always take retry.backoff.ms milliseconds to complete

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 0.9.0.1, 0.10.0.0
• Fix Version/s: 0.10.2.0
• Component/s: clients
• Assignee: Yuto Kawamura
• Reporter: Yuto Kawamura
• Created: 06/Aug/16 09:14
• Updated: 04/Nov/16 06:20
• Resolved: 04/Nov/16 06:20

Description

Recently I updated our KafkaProducer configuration, specifically we adjusted retry.backoff.ms from default(100ms) to 1000ms.
After that we observed that the first send() start taking longer than before, investigated then found following facts.

Environment:

• Kafka broker 0.9.0.1
• Kafka producer 0.9.0.1

Our current version is 0.9.0.1 but it reproduced with latest build from trunk branch as well.

TL;DR

The first KafkaProducer.send() always blocked retry.backoff.ms milliseconds, due to unintentionally applied backoff on first metadata update.

Proof

I wrote following test code and placed under the clients/main/java/

import java.util.Properties;
import java.util.concurrent.TimeUnit;

import org.apache.kafka.clients.producer.KafkaProducer;
import org.apache.kafka.clients.producer.Producer;
import org.apache.kafka.clients.producer.ProducerConfig;
import org.apache.kafka.common.serialization.ByteArraySerializer;

public final class KafkaProducerMetadataUpdateDurationTest {
    public static void main(String[] args) {
        Properties props = new Properties();
        props.setProperty(ProducerConfig.BOOTSTRAP_SERVERS_CONFIG, "localhost:9092");
        props.setProperty(ProducerConfig.MAX_BLOCK_MS_CONFIG, "30000");
        String retryBackoffMs = System.getProperty("retry.backoff.ms");
        System.err.println("Experimenting with retry.backoff.ms = " + retryBackoffMs);
        props.setProperty(ProducerConfig.RETRY_BACKOFF_MS_CONFIG, retryBackoffMs);

        Producer<byte[], byte[]> producer =
                new KafkaProducer<>(props, new ByteArraySerializer(), new ByteArraySerializer());

        long t0 = System.nanoTime();
        try {
            producer.partitionsFor("test");
            long duration = System.nanoTime() - t0;
            System.err.println("Duration = " + TimeUnit.NANOSECONDS.toMillis(duration) + " ms");
        } finally {
            producer.close();
        }
    }
}

Here's experiment log:

# Start zookeeper & kafka broker
./bin/zookeeper-server-start.sh config/zookeeper.properties
./bin/kafka-server-start.sh config/server.properties

# Create test topic
./bin/kafka-topics.sh --zookeeper localhost:2181 --create --topic test --replication-factor 1 --partitions 1

$ ./bin/kafka-run-class.sh -Dretry.backoff.ms=100 KafkaProducerMetadataUpdateDurationTest
Experimenting with retry.backoff.ms = 100
Duration = 175 ms

$ ./bin/kafka-run-class.sh -Dretry.backoff.ms=1000 KafkaProducerMetadataUpdateDurationTest
Experimenting with retry.backoff.ms = 1000
Duration = 1066 ms

$ ./bin/kafka-run-class.sh -Dretry.backoff.ms=10000 KafkaProducerMetadataUpdateDurationTest
Experimenting with retry.backoff.ms = 10000
Duration = 10070 ms

As you can see, duration of partitionsFor() increases linearly in proportion to the value of retry.backoff.ms.

Here I describe the scenario that leads this behavior:
1. KafkaProducer initializes metadata with giving bootstrap.servers and the current timestamp: https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/clients/producer/KafkaProducer.java#L276
2. On the first send(), KafkaProducer requests metadata update due to missing partition info: https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/clients/producer/KafkaProducer.java#L527
3. But, DefaultMetadataUpdater doesn't actually send MetadataRequest, because metadata.timeToNextUpdate returns a value lager than zero: https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/clients/NetworkClient.java#L541-L548
4. Metadata.timeToNextUpdate returns lager one of time till metadata expiration or time till backing off expiration but practially needUpdate is always true at the first time so here the timeToAllowUpdate is always adopted, which never be zero until retry.backoff.ms elapsed since the first metadata.update(): https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/clients/Metadata.java#L116

This is because of kafka client tries to keep interval configured by retry.backoff.ms between each metadata update so it's basically works fine from the second update but for the first time, since it could never have the actual metadata(which is obtained by MetadaUpdate request), this backing off isn't making sense and in fact it's harming our application by blocking the first send() insanely long.

Issue Links

• links to: GitHub Pull Request #1707