logstash/docs/static/troubleshoot/ts-kafka.asciidoc
2021-07-14 14:00:49 -04:00

166 lines
7.3 KiB
Text
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

[[ts-kafka]]
==== Kafka issues and solutions
[discrete]
[[ts-kafka-timeout]]
===== Kafka session timeout issues (input)
*Symptoms*
Throughput issues and duplicate event processing {ls} logs warnings:
-----
[2017-10-18T03:37:59,302][WARN][org.apache.kafka.clients.consumer.internals.ConsumerCoordinator]
Auto offset commit failed for group clap_tx1: Commit cannot be completed since
the group has already rebalanced and assigned the partitions to another member.
-----
The time between subsequent calls to `poll()` was longer than the
configured `session.timeout.ms`, which typically implies that the poll loop is
spending too much time processing messages. You can address this by
increasing the session timeout or by reducing the maximum size of batches
returned in `poll()` with `max.poll.records`.
-----
[INFO][org.apache.kafka.clients.consumer.internals.ConsumerCoordinator] Revoking
previously assigned partitions [] for group log-ronline-node09
`[2018-01-29T14:54:06,485][INFO]`[org.apache.kafka.clients.consumer.internals.ConsumerCoordinator]
Setting newly assigned partitions [elk-pmbr-9] for group log-pmbr
-----
*Background*
Kafka tracks the individual consumers in a consumer group (for example, a number
of {ls} instances) and tries to give each consumer one or more specific
partitions of data in the topic theyre consuming. In order to achieve this,
Kafka tracks whether or not a consumer ({ls} Kafka input thread) is making
progress on their assigned partition, and reassigns partitions that have not
made progress in a set timeframe.
When {ls} requests more events from the Kafka Broker than it can process within
the timeout, it triggers reassignment of partitions. Reassignment of partitions
takes time, and can cause duplicate processing of events and significant
throughput problems.
*Possible solutions*
* Reduce the number of records per request that {ls} polls from the Kafka Broker in one request,
* Reduce the number of Kafka input threads, and/or
* Increase the relevant timeouts in the Kafka Consumer configuration.
*Details*
The `max_poll_records` option sets the number of records to be pulled in one request.
If it exceeds the default value of 500, try reducing it.
The `consumer_threads` option sets the number of input threads. If the value exceeds
the number of pipeline workers configured in the `logstash.yml` file, it should
certainly be reduced.
If the value is greater than 4, try reducing it to `4` or less if the client has
the time/resources for it. Try starting with a value of `1`, and then
incrementing from there to find the optimal performance.
The `session_timeout_ms` option sets the relevant timeout. Set it to a value
that ensures that the number of events in `max_poll_records` can be safely
processed within the time limit.
-----
EXAMPLE
Pipeline throughput is `10k/s` and `max_poll_records` is set to 1k =>. The value
must be at least 100ms if `consumer_threads` is set to `1`. If it is set to a
higher value `n`, then the minimum session timeout increases proportionally to
`n * 100ms`.
-----
In practice the value must be set much higher than the theoretical value because
the behavior of the outputs and filters in a pipeline follows a distribution.
The value should also be higher than the maximum time you expect your outputs to
stall. The default setting is `10s == 10000ms`. If you are experiencing
periodic problems with an output that can stall because of load or similar
effects (such as the Elasticsearch output), there is little downside to
increasing this value significantly to say `60s`.
From a performance perspective, decreasing the `max_poll_records` value is preferable
to increasing the timeout value. Increasing the timeout is your only option if the
clients issues are caused by periodically stalling outputs. Check logs for
evidence of stalling outputs, such as `ES output logging status 429`.
[discrete]
[[ts-schema-registry]]
===== Kafka input plugin crashes when using schema registry
By default, the kafka input plugin checks connectivity and validates the schema registry during plugin registration before events are processed.
In some circumstances, this process may fail when it tries to validate an authenticated schema registry, causing the plugin to crash.
The plugin offers a `schema_registry_validation` setting to change the default behavior.
This setting allows the plugin to skip validation during registration, which allows the plugin to continue and events to be processed.
See the <<plugins-inputs-kafka-schema_registry_validation,kafka input plugin documentation>> for more information about the plugin and other configuration options.
NOTE: An incorrectly configured schema registry will still stop the plugin from processing events.
NOTE: The default setting of `auto` is the best option for most circumstances and should not need to be changed.
[discrete]
[[ts-kafka-many-offset-commits]]
===== Large number of offset commits (input)
*Symptoms*
Logstashs Kafka Input is causing a much higher number of commits to
the offset topic than expected. Often the complaint also mentions redundant
offset commits where the same offset is committed repeatedly.
*Solution*
For Kafka Broker versions 0.10.2.1 to 1.0.x: The problem is caused by a bug in
Kafka. https://issues.apache.org/jira/browse/KAFKA-6362 The clients best option
is upgrading their Kafka Brokers to version 1.1 or newer.
For older versions of
Kafka or if the above does not fully resolve the issue: The problem can also be
caused by setting the value for `poll_timeout_ms` too low relative to the rate
at which the Kafka Brokers receive events themselves (or if Brokers periodically
idle between receiving bursts of events). Increasing the value set for
`poll_timeout_ms` proportionally decreases the number of offsets commits in
this scenario. For example, raising it by 10x will lead to 10x fewer offset commits.
[discrete]
[[ts-kafka-codec-errors-input]]
===== Codec Errors in Kafka Input (before Plugin Version 6.3.4 only)
*Symptoms*
Logstash Kafka input randomly logs errors from the configured codec and/or reads
events incorrectly (partial reads, mixing data between multiple events etc.).
-----
Log example: [2018-02-05T13:51:25,773][FATAL][logstash.runner ] An
unexpected error occurred! {:error=>#<TypeError: can't convert nil into String>,
:backtrace=>["org/jruby/RubyArray.java:1892:in `join'",
"org/jruby/RubyArray.java:1898:in `join'",
"/usr/share/logstash/logstash-core/lib/logstash/util/buftok.rb:87:in `extract'",
"/usr/share/logstash/vendor/bundle/jruby/1.9/gems/logstash-codec-line-3.0.8/lib/logstash/codecs/line.rb:38:in
`decode'",
"/usr/share/logstash/vendor/bundle/jruby/1.9/gems/logstash-input-kafka-5.1.11/lib/logstash/inputs/kafka.rb:241:in
`thread_runner'",
"file:/usr/share/logstash/vendor/jruby/lib/jruby.jar!/jruby/java/java_ext/java.lang.rb:12:in
`each'",
"/usr/share/logstash/vendor/bundle/jruby/1.9/gems/logstash-input-kafka-5.1.11/lib/logstash/inputs/kafka.rb:240:in
`thread_runner'"]}
-----
*Background*
There was a bug in the way the Kafka Input plugin was handling codec instances
when running on multiple threads (`consumer_threads` set to > 1).
https://github.com/logstash-plugins/logstash-input-kafka/issues/210
*Solution*
* Upgrade Kafka Input plugin to v. 6.3.4 or later.
* If (and only if) upgrading is not possible, set `consumer_threads` to `1`.