Client Configuration Properties for Confluent Platform¶
Client configuration properties for an Apache Kafka® Producer or Consumer determine how the client interacts with a Kafka cluster. You can tweak several default configuration property settings to achieve better performance based on the workload. This document will help you understand how to configure your Kafka Producer and Consumer clients to optimize client performance based on your workload.
Why tuning client configurations is important¶
Kafka client configurations provide flexibility and control over various aspects of the client’s behavior, performance, security, and reliability. Properly tuning these configurations helps optimize the client’s interactions with the Kafka cluster and ensures efficient message processing. The following are two specific areas where ensuring correct settings positively impacts the workload:
- Performance: Client configurations can be adjusted to optimize performance. Adjusting properties that control batching, compression, linger, and prefetch can significantly impact client throughput, latency, and resource utilization.
- Error handling: Kafka clients need to handle errors with retries, or fail gracefully until a solution can be implemented to resolve the error. Ensuring the configuration is correct can enhance workload resilience and ensure reliability for mission-critical applications.
Configuration categories¶
Client configuration properties are grouped into the following configuration categories:
- Connection and network properties: A Kafka client must establish a connection with Confluent clusters to produce and consume messages. This category includes settings for bootstrap servers, connection timeout, and network buffer sizes. Optimizing these settings can ensure reliable and efficient communication between the client and the Kafka cluster.
- Security and authentication properties: Kafka supports various security mechanisms, such as SSL/TLS encryption, SASL authentication, and authorization using Access Control Lists (ACLs). This category includes security-related settings, such as SSL certificates, authentication protocols, and user credentials. Properly configuring security settings ensures the confidentiality, integrity, and authenticity of the communication between clients and the Kafka cluster.
- Message processing properties: Kafka clients can process messages in
various ways, such as consuming messages from specific topics, committing
message offsets, or specifying how to handle message errors. This category
includes
max.poll.records,auto.commit.interval.ms,acks, and several others. Fine-tuning these property settings may improve client throughput, fault tolerance, and processing guarantees.
Configuration properties¶
The following tables provide several important configuration properties for Java and librdkafka clients. For a complete listing of configuration properties, see the following documentation:
Before you modify properties¶
Before you start modifying client configuration properties to find out if you can tweak client performance, be sure to complete the following steps.
- Verify your client is using default configuration properties. Someone may have changed configuration properties from their default settings.
- Update your client to the latest supported version available. Default configuration property settings are optimized in later clients. For more information, see Client versions and support.
Important
When modifying configuration properties, monitor the impact on your system and ensure it behaves as expected. Always test any changes in a staging or pre-production environment before rolling them out to production.
Common properties¶
The following table provides several common configuration properties for Producers and Consumers that you can review for potential modification.
| Configuration property | Java default | librdkafka default | Notes |
|---|---|---|---|
client.id |
empty string | rdkafka | You should set the client.id to something meaningful in your
application, especially if you are running multiple clients or want to
easily trace logs or activities to specific client instances. |
connections.max.idle.ms |
540000 ms (9 min) | See librdkafka socket.timeout.ms |
You can change this when an intermediate load balancer disconnects idle connections after inactivity. For example: AWS 350 seconds, Azure 4 minutes, Google Cloud 10 minutes. |
sasl.kerberos.service.name |
null | kafka | Changing the default service name will cause issues for those who don’t have it configured. |
socket.connection.setup.timeout.max.ms |
30000 ms (30 sec) | not available | librdkafka doesn’t have exponential backoff for this timeout. |
socket.connection.setup.timeout.ms |
10000 ms (10 sec) | 30000 ms (30 sec) | librdkafka doesn’t have exponential backoff for this timeout. |
metadata.max.age.ms |
300000 ms (5 min) | 900000 ms (15 min) | librdkafka has the topic.metadata.refresh.interval.ms property that
defaults to 300000 milliseconds (5 minutes). |
reconnect.backoff.max.ms |
1000 ms (1 sec) | 10000 ms (10 sec) | |
reconnect.backoff.ms |
50 ms | 100 ms | |
max.in.flight.requests.per.connection |
5 | 1000000 | librdkafka produces to a single partition per batch, setting it to 5 limits producing to 5 partitions per broker. |
Producer properties¶
The following table provides a few configuration properties for Producers that you can review for potential modification.
| Configuration property | Java default | librdkafka default | Notes |
|---|---|---|---|
batch.size |
16384 | 1000000 | |
delivery.timeout.ms |
120000 ms (2 min) | 300000 ms (5 min) | |
linger.ms |
0 ms | 5 ms | librdkafka linger.ms reduces the number of in-flight Produce requests
and increases batching (see max.in.flight.requests.per.connection) |
enable.idempotence |
true | false | Enabling idempotence sets max.in.flight.requests.per.connection to
5 (see max.in.flight.requests.per.connection) |
partitioner |
murmur2_random (default Kafka partitioner) | consistent_random | Changing the default partitioner causes the client to send keyed messages to different partitions. If both a librdkafka-based and a Java Client are producing to the same topic, change this property to murmur2_random for the librdkafka client so that messages with the same key are sent to the same partition. |
Consumer properties¶
The following table provides a few configuration properties for Consumers that you can review for potential modification.
| Configuration property | Java default | librdkafka default | Notes |
|---|---|---|---|
allow.auto.create.topics |
true | false | |
isolation.level |
read_uncommitted | read_committed | |
partition.assignment.strategy |
RangeAssignor, CooperativeStickyAssignor | range, roundrobin | Online upgrade from eager to cooperative assignor is not supported in librdkafka. |
check.crcs |
true | false | Record checksum validation comes at slightly increased CPU usage. Checksum is also present at the IPv4 and TCP layers. Other types of checks could be available at disk sector (ECC) or file system level (not in ext4 by default). |
OpenId Connect (OIDC) and token retry behavior¶
The OIDC retry behavior handles operations such as retrieving a new authentication tokens or refreshing existing tokens. Before modifying configuration properties related to retry behavior, read this section to understand how it works for your client.
Java Client¶
The token refresh process begins when a credential’s lifetime reaches a
specified percentage. This percentage is 80% by default, but you can configure a
different value by implementing the
org.apache.kafka.common.security.oauthbearer.OAuthBearerToken interface in a
custom class and specifying the token lifetime.
If an authentication error occurs during the token refresh process, the client waits 10 seconds before retrying the token refresh. You cannot configure the wait time before the refresh retry.
There are no maximum number of tries for token refresh. The token refresh process continues to retry until it succeeds or the application is closed. The underlying HTTP request to fetch the token from the identity provider (IdP) uses an exponential retry delay mechanism (backoff). This retry delay mechanism doubles after each failed attempt.
By default, the retry backoff starts at 100 ms and increase up to a maximum of 10000 ms.
You can configure alternatives values if you like by configuring
sasl.oauthbearer.jwks.endpoint.retry.backoff.ms and
sasl.oauthbearer.jwks.endpoint.retry.backoff.max.ms.
Schema Registry Java Client¶
The retry behavior for this client is identical to the Java Client. By default,
the retry starts at 100 ms and increases up to a maximum of 10000ms. You can
configure alternative values if you like by configuring
sasl.oauthbearer.jwks.endpoint.retry.backoff.ms and
sasl.oauthbearer.jwks.endpoint.retry.backoff.max.ms.
JavaScript Client for Kafka¶
The token refresh process begins when a credential’s lifetime reaches 80%. The token refresh can fail due to one of these error codes:
| Code | Description |
|---|---|
408 |
Request timeout |
425 |
Too early |
429 |
Too many requests |
500 |
Internal server error |
502 |
Bad gateway |
503 |
Service unavailable |
504 |
Gateway timeout |
To see the errors in the source go here. If any of these errors occur, the retry process attempts four retries with backoff (5s, 10s, 15s, 20s), with no error logging.
If the process returns a different error code, or if all the retries fail, then the process logs an error and repeats the same process with a linear backoff (1+4) tries after 10s. This process continues until retry succeeds or the token expires. If all attempts to refresh a token fail and the token expires, produce and fetch operations begin failing also.
If you would like to add custom behavior, for example, jitter, or different timeouts, you can write an application-side token refresh callback. It replaces the process of fetching the credentials and the 1+4 tries. The library triggers the custom callback at 80% of token expiry duration, and in case the callback signals failure, it is re-triggered after 10s.
Schema Registry JavaScript Client¶
The token refresh process begins at 30 minutes before the token expires. If token
retrieval fails due to a 429 rate limitation error, the process retries
maxRetries, from a retriesWaitMs minimum to a maximum retriesMaxWaitMs delay.
You can configure these retry properties on your client, for example:
const clientConfig = {
baseURLs: ['http://my-schema-registry:8081'],
maxRetries: 5,
retriesWaitMs: 1000,
retriesMaxWaitMs: 8000,
};
After each retry, the delay value before the next retry increases exponentially by multiples of two, with full jitter. Full jitter means the retry process multiplies the delay value by a random number from zero to one to determine the actual delay. Any other non-retriable error code causes the process to throw an error.
Token retrieval at startup follows the same retry pattern.
librdkafka derived (non-Java) clients¶
For these clients, the token refresh process is identical to the JavaScript Client for Kafka.