Hi Leo, I'm not sure if this is the issue you're encountering, but this is what we found when we went from 0.8.1.1 to 0.8.2.1.
Snappy compression didn't work as expected. Something in the library broke compressing bundles of messages and each message was compressed individually (which for us caused a lot of overhead). Disk usage went way up and CPU usage went incrementally up (still under 1%). I didn't monitor latency, it was well within the tolerances of our system. We resolved this issue by switching our compression to gzip. This issue is supposedly fixed in 0.9.0.0 but we haven't verified it yet. Cliff On Thu, Jan 21, 2016 at 4:04 AM, Clelio De Souza <cleli...@gmail.com> wrote: > Hi all, > > > We are using Kafka in production and we have been facing some performance > degradation of the cluster, apparently after the cluster is a bit "old". > > > We have our production cluster which is up and running since 31/12/2015 and > performance tests on our application measuring a full round trip of TCP > packets and Kafka producing/consumption of data (3 hops in total for every > single TCP packet being sent, persisted and consumed in the other end). The > results for the production cluster shows a latency of ~ 130ms to 200ms. > > > In our Test environment we have the very same software and specification in > AWS instances, i.e. Test environment as being a mirror of Prod. The Kafka > cluster has been running in Test since 18/12/2015 and the same performance > tests (as described above) shows a increase of latency to ~ 800ms to > 1000ms. > > > We have just recently setup a new fresh Kafka cluster (on 18/01/2016) > trying to get to the bottom of this performance degradation problem and in > the new Kafka cluster deployed in Test in replacement of the original Test > Kafka cluster we found a very small latency of ~ 10ms to 15ms. > > > We are using Kafka 0.8.2.1 version for all those environment mentioned > above. And the same cluster configuration has been setup on all of them, as > 3 brokers as m3.xlarge AWS instances. The amount of data and Kafka topics > are roughly the same among those environments, therefore the performance > degradation seems to be not directly related to the amount of data in the > cluster. We suspect that something running inside of the Kafka cluster, > such as repartitioning or log rentention (even though our topics are to > setup to last for ~ 2 years and it has not elapsed this time at all). > > > The Kafka broker config can be found as below. If anyone could shed some > lights on what it may be causing the performance degradation for our Kafka > cluster, it would be great and very much appreciate it. > > > Thanks, > Leo > > -------------------- > > > # Licensed to the Apache Software Foundation (ASF) under one or more > # contributor license agreements. See the NOTICE file distributed with > # this work for additional information regarding copyright ownership. > # The ASF licenses this file to You under the Apache License, Version 2.0 > # (the "License"); you may not use this file except in compliance with > # the License. You may obtain a copy of the License at > # > # http://www.apache.org/licenses/LICENSE-2.0 > # > # Unless required by applicable law or agreed to in writing, software > # distributed under the License is distributed on an "AS IS" BASIS, > # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. > # See the License for the specific language governing permissions and > # limitations under the License. > # see kafka.server.KafkaConfig for additional details and defaults > > ############################# Server Basics ############################# > > # The id of the broker. This must be set to a unique integer for each > broker. > broker.id=<broker_num> > > ############################# Socket Server Settings > ############################# > > # The port the socket server listens on > port=9092 > > # Hostname the broker will bind to. If not set, the server will bind to all > interfaces > #host.name=localhost > > # Hostname the broker will advertise to producers and consumers. If not > set, it uses the > # value for "host.name" if configured. Otherwise, it will use the value > returned from > # java.net.InetAddress.getCanonicalHostName(). > #advertised.host.name=<hostname routable by clients> > > # The port to publish to ZooKeeper for clients to use. If this is not set, > # it will publish the same port that the broker binds to. > #advertised.port=<port accessible by clients> > > # The number of threads handling network requests > num.network.threads=3 > > # The number of threads doing disk I/O > num.io.threads=8 > > # The send buffer (SO_SNDBUF) used by the socket server > socket.send.buffer.bytes=102400 > > # The receive buffer (SO_RCVBUF) used by the socket server > socket.receive.buffer.bytes=102400 > > # The maximum size of a request that the socket server will accept > (protection against OOM) > socket.request.max.bytes=104857600 > > ############################# Log Basics ############################# > > # A comma seperated list of directories under which to store log files > log.dirs=/data/kafka/logs > > # The default number of log partitions per topic. More partitions allow > greater > # parallelism for consumption, but this will also result in more files > across > # the brokers. > num.partitions=8 > > # The number of threads per data directory to be used for log recovery at > startup and flushing at shutdown. > # This value is recommended to be increased for installations with data > dirs located in RAID array. > num.recovery.threads.per.data.dir=1 > > ############################# Log Flush Policy > ############################# > > # Messages are immediately written to the filesystem but by default we only > fsync() to sync > # the OS cache lazily. The following configurations control the flush of > data to disk. > # There are a few important trade-offs here: > # 1. Durability: Unflushed data may be lost if you are not using > replication. > # 2. Latency: Very large flush intervals may lead to latency spikes when > the flush does occur as there will be a lot of data to flush. > # 3. Throughput: The flush is generally the most expensive operation, > and a small flush interval may lead to exceessive seeks. > # The settings below allow one to configure the flush policy to flush data > after a period of time or > # every N messages (or both). This can be done globally and overridden on a > per-topic basis. > > # The number of messages to accept before forcing a flush of data to disk > #log.flush.interval.messages=10000 > > # The maximum amount of time a message can sit in a log before we force a > flush > #log.flush.interval.ms=1000 > > ############################# Log Retention Policy > ############################# > > # The following configurations control the disposal of log segments. The > policy can > # be set to delete segments after a period of time, or after a given size > has accumulated. > # A segment will be deleted whenever *either* of these criteria are met. > Deletion always happens > # from the end of the log. > > # The minimum age of a log file to be eligible for deletion > # Failsafe is we don't lose any messages for 20+ years, topics should > # be configured individually > log.retention.hours=200000 > > # A size-based retention policy for logs. Segments are pruned from the log > as long as the remaining > # segments don't drop below log.retention.bytes. > #log.retention.bytes=1073741824 > > # The maximum size of a log segment file. When this size is reached a new > log segment will be created. > log.segment.bytes=1073741824 > > # The interval at which log segments are checked to see if they can be > deleted according > # to the retention policies > log.retention.check.interval.ms=300000 > > # By default the log cleaner is disabled and the log retention policy will > default to just delete segments after their retention expires. > # If log.cleaner.enable=true is set the cleaner will be enabled and > individual logs can then be marked for log compaction. > log.cleaner.enable=false > > default.replication.factor=3 > > auto.create.topics.enable=true > > controlled.shutdown.enable=true > > delete.topic.enable=true > > ############################# Zookeeper ############################# > > # Zookeeper connection string (see zookeeper docs for details). > # This is a comma separated host:port pairs, each corresponding to a zk > # server. e.g. "127.0.0.1:3000,127.0.0.1:3001,127.0.0.1:3002". > # You can also append an optional chroot string to the urls to specify the > # root directory for all kafka znodes. > zookeeper.connect=<zk1-address>:2181,<zk2-address>:2181,<zk3-address>:2181 > > # Timeout in ms for connecting to zookeeper > zookeeper.connection.timeout.ms=6000 > -- Cliff Rhyne Software Engineering Lead e: crh...@signal.co signal.co ________________________ Cut Through the Noise This e-mail and any files transmitted with it are for the sole use of the intended recipient(s) and may contain confidential and privileged information. Any unauthorized use of this email is strictly prohibited. ©2015 Signal. All rights reserved.
