Here is some information on actually running Kafka as a production system. This is meant as a page for people to record their operational and monitoring practices to help people gather knowledge about successfully running Kafka is production. Feel free to add a section for your configuration if you have anything you want to share. There is nothing magically about most of these configurations, you may be able to improve on them, but they may serve as a helpful starting place.
Hardware
We are using dual quad-core Intel Xeon machines with 24GB of memory. In general this should not matter too much, we only see pretty low CPU usage at peak even with GZIP compression enabled and a number of clients that don't batch requests. The memory is probably more than is needed for caching the active segments of the log.
The disk throughput is important. We have 8x7200 rpm SATA drives in a RAID 10 array. In general this is the performance bottleneck, and more disks is more better. Depending on how you configure flush behavior you may or may not benefit from more expensive disks (if you flush often then higher RPM SAS drives may be better).
OS Settings
We use linux. Ext4 is the filesystem and we run using software RAID 10. We haven't benchmarked filesystems so other filesystems may be superior.
We have added two tuning changes: (1) we upped the number of file descriptors since we have lots of topics and lots of connections, and (2) we upped the max socket buffer size to enable high-performance data transfer between data centers (described here).
Java
$ java -version java version "1.6.0_21" Java(TM) SE Runtime Environment (build 1.6.0_21-b06) Java HotSpot(TM) 64-Bit Server VM (build 17.0-b16, mixed mode)
Here are our command line options:
java -server -Xms3072m -Xmx3072m -XX:NewSize=256m -XX:MaxNewSize=256m -XX:+UseConcMarkSweepGC -XX:CMSInitiatingOccupancyFraction=70 -XX:+PrintGCDetails -XX:+PrintGCDateStamps -XX:+PrintTenuringDistribution -Xloggc:logs/gc.log -Djava.awt.headless=true -Dcom.sun.management.jmxremote -classpath <long list of jars>
Kafka
We are running Kafka 0.7 right now but may move to trunk as we fix bugs.
The most important server configurations for performance are those that control the disk flush rate. The more often data is flushed to disk, the more "seek-bound" Kafka will be and the lower the throughput. However we don't hand out data until is is sync'd to disk, so delaying flush also adds some latency. The flush behavior can be set by either giving a timeout (flush at most every 30 seconds, say) or a number of messages (flush every 1000 messages). This can be overridden at the topic level, if desired. The setting always applies to each partition.
The most important client configurations for performance are (1) compression, (2) sync vs async production, (3) batch size for async production, (4) fetch size.
Here is our server configuration:
kafka.log.default.flush.interval.ms=10000 kafka.log.file.size=1073741824 kafka.log.default.flush.scheduler.interval.ms=2000 kafka.log.flush.interval=3000kafka.socket.send.buffer=2097152 kafka.socket.receive.buffer=2097152 kafka.monitoring.period.secs=30 kafka.num.threads=8 kafka.log.cleanup.interval.mins=30 kafka.log.retention.hours=168 kafka.zookeeper.sessiontimeoutms=6000 kafka.zookeeper.connection.timeout=2000 kafka.num.partitions=1
Client configuration varies a fair amount.
Monitoring
Our monitoring is done though a centralized monitoring system custom to LinkedIn, but it keys off the JMX stats exposed from kafka. To see what is available the easiest thing is just to start a kafka broker and/or client and fire up JConsole and take a look.
Server Stats
- bean name: kafka:type=kafka.SocketServerStats
def getProduceRequestsPerSecond: Double def getFetchRequestsPerSecond: Double def getAvgProduceRequestMs: Double def getMaxProduceRequestMs: Double def getAvgFetchRequestMs: Double def getMaxFetchRequestMs: Double def getBytesReadPerSecond: Double def getBytesWrittenPerSecond: Double def getNumFetchRequests: Long def getNumProduceRequests: Long def getTotalBytesRead: Long def getTotalBytesWritten: Long def getTotalFetchRequestMs: Long def getTotalProduceRequestMs: Long
bean name: kafka:type=kafka.BrokerAllTopicStat kafka:type=kafka.BrokerAllTopicStat.[topic]
def getMessagesIn: Long def getBytesIn: Long def getBytesOut: Long def getFailedProduceRequest: Long def getFailedFetchRequest: Long
- bean name: kafka:type=kafka.LogFlushStats
def getFlushesPerSecond: Double def getAvgFlushMs: Double def getTotalFlushMs: Long def getMaxFlushMs: Double def getNumFlushes: Long
Producer stats
- bean name: kafka:type=kafka.KafkaProducerStats
def getProduceRequestsPerSecond: Double def getAvgProduceRequestMs: Double def getMaxProduceRequestMs: Double def getNumProduceRequests: Long
- bean name: kafka.producer.Producer:type=AsyncProducerStats
def getAsyncProducerEvents: Int def getAsyncProducerDroppedEvents: Int
Consumer stats
- bean name: kafka:type=kafka.ConsumerStats
def getPartOwnerStats: String def getConsumerGroup: String def getOffsetLag(topic: String, brokerId: Int, partitionId: Int): Long def getConsumedOffset(topic: String, brokerId: Int, partitionId: Int): Long def getLatestOffset(topic: String, brokerId: Int, partitionId: Int): Long
bean name: kafka:type=kafka.ConsumerAllTopicStat kafka:type=kafka.ConsumerTopicStat.[topic]
def getMessagesPerTopic: Long def getBytesPerTopic: Long
- bean name: kafka:type=kafka.SimpleConsumerStats
def getFetchRequestsPerSecond: Double def getAvgFetchRequestMs: Double def getMaxFetchRequestMs: Double def getNumFetchRequests: Long def getConsumerThroughput: Double
Audit
The final alerting we do is on the correctness of the data delivery. We audit that every message that is sent is consumed by all consumers and measure the lag for this to occur. For important topics we alert if a certain completeness is not achieved in a certain time period. The details of this are discussed in KAFKA-260.
Zookeeper
Zookeeper is essential for the correct operation of Kafka. There are a number of things that must be done to keep zookeeper running happily as we have learned the hard way, hopefully Dave and Neha will add this since I don't know what we did.