Logback appender for Apache Kafka
This appender lets your application publish its application logs directly to Apache Kafka.
Due to a breaking change in the Logback Encoder API you need to use at least logback version 1.2.
Add
logback-kafka-appenderand
logback-classicas library dependencies to your project.
[maven pom.xml] com.github.danielwegener logback-kafka-appender 0.2.0 runtime ch.qos.logback logback-classic 1.2.3 runtime
// [build.sbt] libraryDependencies += "com.github.danielwegener" % "logback-kafka-appender" % "0.2.0" libraryDependencies += "ch.qos.logback" % "logback-classic" % "1.2.3"
This is an example
logback.xmlthat uses a common
PatternLayoutto encode a log message as a string.
[src/main/resources/logback.xml]<appender name="STDOUT" class="ch.qos.logback.core.ConsoleAppender"> <encoder> <pattern>%d{HH:mm:ss.SSS} [%thread] %-5level %logger{36} - %msg%n</pattern> </encoder> </appender> <!-- This is the kafkaAppender --> <appender name="kafkaAppender" class="com.github.danielwegener.logback.kafka.KafkaAppender"> <encoder> <pattern>%d{HH:mm:ss.SSS} [%thread] %-5level %logger{36} - %msg%n</pattern> </encoder> <topic>logs</topic> <keyingstrategy class="com.github.danielwegener.logback.kafka.keying.NoKeyKeyingStrategy"></keyingstrategy> <deliverystrategy class="com.github.danielwegener.logback.kafka.delivery.AsynchronousDeliveryStrategy"></deliverystrategy> <!-- Optional parameter to use a fixed partition --> <!-- <partition>0</partition> --> <!-- Optional parameter to include log timestamps into the kafka message --> <!-- <appendTimestamp>true</appendTimestamp> --> <!-- each <producerConfig> translates to regular kafka-client config (format: key=value) --> <!-- producer configs are documented here: https://kafka.apache.org/documentation.html#newproducerconfigs --> <!-- bootstrap.servers is the only mandatory producerConfig --> <producerconfig>bootstrap.servers=localhost:9092</producerconfig> <!-- this is the fallback appender if kafka is not available. --> <appender-ref ref="STDOUT"></appender-ref> </appender> <root level="info"> <appender-ref ref="kafkaAppender"></appender-ref> </root>
You may also look at the complete configuration examples
logback-kafka-appender depends on
org.apache.kafka:kafka-clients:1.0.0:jar. It can append logs to a kafka broker with version 0.9.0.0 or higher.
The dependency to kafka-clients is not shadowed and may be upgraded to a higher, api compatible, version through dependency overrides.
Direct logging over the network is not a trivial thing because it might be much less reliable than the local file system and has a much bigger impact on the application performance if the transport has hiccups.
You need make a essential decision: Is it more important to deliver all logs to the remote Kafka or is it more important to keep the application running smoothly? Either of this decisions allows you to tune this appender for throughput.
| Strategy | Description | |---|---| |
AsynchronousDeliveryStrategy| Dispatches each log message to the
Kafka Producer. If the delivery fails for some reasons, the message is dispatched to the fallback appenders. However, this DeliveryStrategy does block if the producers send buffer is full (this can happen if the connection to the broker gets lost). To avoid even this blocking, enable the producerConfig
block.on.buffer.full=false. All log messages that cannot be delivered fast enough will then immediately go to the fallback appenders. | |
BlockingDeliveryStrategy| Blocks each calling thread until the log message is actually delivered. Normally this strategy is discouraged because it has a huge negative impact on throughput. Warning: This strategy should not be used together with the producerConfig
linger.ms|
The
AsynchronousDeliveryStrategydoes not prevent you from being blocked by the Kafka metadata exchange. That means: If all brokers are not reachable when the logging context starts, or all brokers become unreachable for a longer time period (>
metadata.max.age.ms), your appender will eventually block. This behavior is undesirable in general and can be migitated with kafka-clients 0.9 (see #16). Until then, you can wrap the KafkaAppender with logback's own AsyncAppender.
An example configuration could look like this:
<!-- This is the kafkaAppender --> <appender name="kafkaAppender" class="com.github.danielwegener.logback.kafka.KafkaAppender"> <!-- Kafka Appender configuration --> </appender> <appender name="ASYNC" class="ch.qos.logback.classic.AsyncAppender"> <appender-ref ref="kafkaAppender"></appender-ref> </appender> <root level="info"> <appender-ref ref="ASYNC"></appender-ref> </root>
You may also roll your own delivery strategy. Just extend
com.github.danielwegener.logback.kafka.delivery.DeliveryStrategy.
If, for whatever reason, the kafka-producer decides that it cannot publish a log message, the message could still be logged to a fallback appender (a
ConsoleAppenderon STDOUT or STDERR would be a reasonable choice for that).
Just add your fallback appender(s) as logback
appender-refto the
KafkaAppendersection in your
logback.xml. Every message that cannot be delivered to kafka will be written to all defined
appender-ref's.
Example:
whileSTDOUTis an defined appender.
Note that the
AsynchronousDeliveryStrategywill reuse the kafka producers io thread to write the message to the fallback appenders. Thus all fallback appenders should be reasonable fast so they do not slow down or break the kafka producer.
This appender uses the kafka producer introduced in kafka-0.8.2. It uses the producer default configuration.
You may override any known kafka producer config with an
Name=Valueblock (note that the
boostrap.serversconfig is mandatory). This allows a lot of fine tuning potential (eg. with
batch.size,
compression.typeand
linger.ms).
This module supports any
ch.qos.logback.core.encoder.Encoder. This allows you to use any encoder that is capable of encoding an
ILoggingEventor
IAccessEventlike the well-known logback
PatternLayoutEncoderor for example the logstash-logback-encoder's
LogstashEncoxer.
If you want to write something different than string on your kafka logging topic, you may roll your encoding mechanism. A use case would be to to smaller message sizes and/or better serialization/deserialization performance on the producing or consuming side. Useful formats could be BSON, Avro or others.
To roll your own implementation please refer to the logback documentation. Note that logback-kafka-appender will never call the
headerBytes()or
footerBytes()method.
Your encoder should be type-parameterized for any subtype of the type of event you want to support (typically
ILoggingEvent) like in
public class MyEncoder extends ch.qos.logback.core.encoder.Encoder {/*..*/}
Kafka's scalability and ordering guarantees heavily rely on the concepts of partitions (more details here). For application logging this means that we need to decide how we want to distribute our log messages over multiple kafka topic partitions. One implication of this decision is how messages are ordered when they are consumed from a arbitrary multi-partition consumer since kafka only provides a guaranteed read order only on each single partition. Another implication is how evenly our log messages are distributed across all available partitions and therefore balanced between multiple brokers.
The order of log messages may or may not be important, depending on the intended consumer-audience (e.g. a logstash indexer will reorder all message by its timestamp anyway).
You can provide a fixed partition for the kafka appender using the
partitionproperty or let the producer use the message key to partition a message. Thus
logback-kafka-appendersupports the following keying strategies strategies:
| Strategy | Description | |---|---| |
NoKeyKeyingStrategy(default) | Does not generate a message key. Results in round robin distribution across partition if no fixed partition is provided. | |
HostNameKeyingStrategy| This strategy uses the HOSTNAME as message key. This is useful because it ensures that all log messages issued by this host will remain in the correct order for any consumer. But this strategy can lead to uneven log distribution for a small number of hosts (compared to the number of partitions). | |
ContextNameKeyingStrategy| This strategy uses logback's CONTEXT_NAME as message key. This is ensures that all log messages logged by the same logging context will remain in the correct order for any consumer. But this strategy can lead to uneven log distribution for a small number of hosts (compared to the number of partitions). This strategy only works for
ILoggingEvents. | |
ThreadNameKeyingStrategy| This strategy uses the calling threads name as message key. This ensures that all messages logged by the same thread will remain in the correct order for any consumer. But this strategy can lead to uneven log distribution for a small number of thread(-names) (compared to the number of partitions). This strategy only works for
ILoggingEvents. | |
LoggerNameKeyingStrategy| * This strategy uses the logger name as message key. This ensures that all messages logged by the same logger will remain in the correct order for any consumer. But this strategy can lead to uneven log distribution for a small number of distinct loggers (compared to the number of partitions). This strategy only works for
ILoggingEvents. |
If none of the above keying strategies satisfies your requirements, you can easily implement your own by implementing a custom
KeyingStrategy:
package foo; import com.github.danielwegener.logback.kafka.keying.KeyingStrategy;/* This is a valid example but does not really make much sense */ public class LevelKeyingStrategy implements KeyingStrategy { @Override public byte[] createKey(ILoggingEvent e) { return ByteBuffer.allocate(4).putInt(e.getLevel()).array(); } }
As most custom logback component, your custom partitioning strategy may also implement the
ch.qos.logback.core.spi.ContextAwareand
ch.qos.logback.core.spi.LifeCycleinterfaces.
A custom keying strategy may especially become handy when you want to use kafka's log compactation facility.
This project is licensed under the Apache License Version 2.0.