Michael, Your understanding is mostly correct. For 2, the follower will issue another fetch request as soon as it finishes processing the response of the previous fetch (by adding data, if any, to its local log).
The latency of ack=-1 shouldn't be affected by the max wait time in the follower. However, when you mentioned that the remote time is 500ms, I was thinking if there may be a bug in our purgatory code. To rule out any purgatory bug, could you check the replica fetch time after each produce request is received by the leader? The replica fetch time should be much less than replica.fetch.wait.max.ms, since it should complete as soon as the produce request is received. Your best result shows close to a factor of 2 difference btw ack=-1 and ack=1, which is actually reasonable. Thanks, Jun On Mon, Feb 3, 2014 at 1:12 PM, Michael Popov <mpo...@microsoft.com> wrote: > Hi Jun, > > I ran a set of tests to check throughput with different values for a > configuration setting replica.fetch.wait.max.ms. In all tests > replica.fetch.min.bytes=1. > > > ---------------------------------+------------------------------------+----------------------------------+ > replica.fetch.wait.max.ms | throughput for acks=-1 | > throughput for acks=1 | > > ---------------------------------+------------------------------------+----------------------------------+ > 50 > 1311 3112 > 100 > 1446 3244 > 200 > 1272 3348 > 500 > 1223 3289 > > There is no a clear winner for both cases. > > I am trying to understand how replication works in Kafka with acks=-1. As > far as I can get it the sequence is like this: > 1. Producer sends a request with data for some topics/partitions to > the leader broker. > This request is processed by the leader broker, i.e. data is > written locally, and this request is placed into waiting queue until > (a) all replicas for topics/partitions in the > request report that data is committed > (leader broker's "committed data offset" > is less or equal to the follower broker's "committed data offset") > (b) if no confirmation arrived for any > topic/partition, the timeout is fired out > After (a) and (b) happened, a leader broker sends response to the > producer. > > 2. A follower broker periodically fetches data from topic/partition on > the leader broker. > Content of these requests from the follower broker is used to > establish "committed data offset" for the replica. > > Did I get it right? > Does it mean that a period of requests from follower brokers and timeout > settings will impact the throughput of a synchronized producer when > acksNum=-1? > > Thank you, > Michael Popov > > -----Original Message----- > From: Jun Rao [mailto:jun...@gmail.com] > Sent: Monday, February 3, 2014 10:09 AM > To: users@kafka.apache.org > Subject: Re: Kafka performance test: "--request-num-acks -1" kills > throughput > > What's replica.fetch.min.bytes? Is it the default value of 1? If not, > could you try to set it to 1? If yes, could you try using a smaller value > of replica.fetch.wait.max.ms? You can then re-run the test and report > back the results. > > Thanks, > > Jun > > > On Mon, Feb 3, 2014 at 9:26 AM, Michael Popov <mpo...@microsoft.com> > wrote: > > > Hi Jun, > > > > Based on visual review of the log files, most of the time is spent in > > remoteTime. > > > > Example: > > > grep Completed kafka-request.log.2014-01-30-01 | sed > > 's/^.*totalTime://' | less > > ... > > 502,queueTime:1,localTime:1,remoteTime:499,sendTime:1 > > (kafka.request.logger) > > 502,queueTime:0,localTime:1,remoteTime:500,sendTime:1 > > (kafka.request.logger) > > 502,queueTime:1,localTime:0,remoteTime:500,sendTime:1 > > (kafka.request.logger) > > 502,queueTime:1,localTime:0,remoteTime:500,sendTime:1 > > (kafka.request.logger) > > 502,queueTime:1,localTime:1,remoteTime:500,sendTime:0 > > (kafka.request.logger) > > 503,queueTime:1,localTime:0,remoteTime:501,sendTime:1 > > (kafka.request.logger) > > 503,queueTime:1,localTime:1,remoteTime:501,sendTime:0 > > (kafka.request.logger) > > 502,queueTime:1,localTime:0,remoteTime:501,sendTime:0 > > (kafka.request.logger) > > 502,queueTime:1,localTime:0,remoteTime:501,sendTime:0 > > (kafka.request.logger) > > ... > > > > Some of the tests were executed with recommended configuration > > settings copied from the product webpage. These configuration settings > > included following lines for "replica": > > > > num.replica.fetchers=4 > > replica.fetch.max.bytes=1048576 > > replica.fetch.wait.max.ms=500 > > replica.high.watermark.checkpoint.interval.ms=5000 > > replica.socket.timeout.ms=30000 > > replica.socket.receive.buffer.bytes=65536 > > replica.lag.time.max.ms=10000 > > replica.lag.max.messages=4000 > > > > In most cases the tests were executed with "out-of-box" settings, > > which don't change "replica" configuration. > > > > We are running these tests on very weak machines. If absolute > > throughput numbers are not as high as in other people's tests, that's > understandable. > > The main concern is why throughput drops 4-10 times when a number of > > expected acks is not 1. > > > > Should we wait for newer versions of Kafka and settle with acks=1 for > now? > > > > Thank you, > > Michael Popov > > > > -----Original Message----- > > From: Jun Rao [mailto:jun...@gmail.com] > > Sent: Friday, January 31, 2014 8:57 PM > > To: users@kafka.apache.org > > Subject: Re: Kafka performance test: "--request-num-acks -1" kills > > throughput > > > > The request log shows you the breakdown of the request time. Where is > > most of the time spent? Also, did you change replica.fetch.wait.max.ms > > in the broker? > > > > Thanks, > > > > Jun > > > > > > On Fri, Jan 31, 2014 at 10:14 AM, Michael Popov <mpo...@microsoft.com > > >wrote: > > > > > Hi Jun, > > > > > > The usage output of bin/kafka-producer-perf-test.sh shows > > > > > > --request-timeout-ms <Integer> The produce request timeout in > ms > > > > > > (default: 3000) > > > > > > In my test runs I used the following command-line arguments: > > > > > > bin/kafka-producer-perf-test.sh --broker-list 10.0.0.8:9092, > > > 10.0.0.10:9092 --compression-codec 0 --message-size 1024 > > > --request-num-acks -1 --sync --messages 100000 -threads 1 > > > --show-detailed-stats --reporting-interval 1000 --topics d2111 | > > > grep -v "at " > > > > > > I assume producer uses default timeout of 3000ms in my tests. > > > > > > I ran a few data processing operations on kafka-request.log.* > > > > > > grep Completed kafka-request.log.2014-01-30-17 | sed > 's/^.*totalTime://' > > > | sed 's/,.*$//' > /tmp/times.log > > > cat /tmp/times.log | sort -k 1n | uniq > /tmp/result.log tail > > > /tmp/result.log > > > 5643 > > > 5761 > > > 5834 > > > 6319 > > > 6320 > > > 6535 > > > 8407 > > > 8554 > > > 9422 > > > 9502 > > > > > > I guess that means some of Fetch requests from the leader broker to > > > the follower broker took much longer than 3000ms. That explains > > > expired producer's requests and probably the overall decrease of the > > > throughput in tests with acksNum!=1. > > > > > > What may cause such long responses? Is there a way to address this > issue? > > > > > > Thank you, > > > Michael Popov > > > > > > > > > > > > -----Original Message----- > > > From: Jun Rao [mailto:jun...@gmail.com] > > > Sent: Friday, January 31, 2014 7:52 AM > > > To: users@kafka.apache.org > > > Subject: Re: Kafka performance test: "--request-num-acks -1" kills > > > throughput > > > > > > Michael, > > > > > > Those SocketTimeoutExceptions meas that the producer didn't receive > > > the response from the broker in time. Could you check the request > > > log in the broker and see what the request completion time is and > > > how it compares with the request socket timeout? > > > > > > I did some testing a while back. Latency wise, ack=-1 is about twice > > > that of ack=1 since it requires an extra network round trip. > > > > > > Thanks, > > > > > > Jun > > > > > > > > > On Thu, Jan 30, 2014 at 11:23 AM, Michael Popov > > > <mpo...@microsoft.com > > > >wrote: > > > > > > > Hi Jun, > > > > > > > > I ran a new new tests today. > > > > > > > > Settings: 2 Kafka brokers, 1 Zookeeper node, 4 client machines. > > > > Creating a new topic with 1 partition for each client process. > > > > Commands to create a topic looked like this one: > > > > > > > > bin/kafka-create-topic.sh --partition 1 --replica 2 --zookeeper > > > > 10.0.0.4:2181 --topic d1 > > > > > > > > Commands to run a test producer looked like this: > > > > > > > > bin/kafka-producer-perf-test.sh --broker-list 10.0.0.8:9092, > > > > 10.0.0.10:9092 --compression-codec 0 --message-size 1024 > > > > --request-num-acks 1 --sync --messages 100000 -threads 10 > > > > --show-detailed-stats --reporting-interval 1000 --topics d1 | grep > > > > -v > > > "at " > > > > > > > > Results of 4 test runs: > > > > ====================================== > > > > --request-num-acks -1 > > > > ===================================== > > > > 2014-01-30 17:55:01:576, 2014-01-30 17:59:23:294, 0, 1024, 200, > > > > 97.66, 0.3731, 100000, 382.0906 > > > > 2014-01-30 17:55:01:921, 2014-01-30 18:20:56:952, 0, 1024, 200, > > > > 97.66, 0.0628, 99999, 64.3068 > > > > 2014-01-30 17:55:00:037, 2014-01-30 18:20:57:164, 0, 1024, 200, > > > > 97.66, 0.0627, 100000, 64.2208 > > > > 2014-01-30 17:54:59:723, 2014-01-30 18:20:34:633, 0, 1024, 200, > > > > 97.66, 0.0636, 100000, 65.1504 > > > > > > > > ====================================== > > > > --request-num-acks 1 > > > > ===================================== > > > > 2014-01-30 18:24:54:588, 2014-01-30 18:27:42:408, 0, 1024, 200, > > > > 97.66, 0.5819, 100000, 595.8765 > > > > 2014-01-30 18:24:54:830, 2014-01-30 18:27:24:080, 0, 1024, 200, > > > > 97.66, 0.6543, 100000, 670.0168 > > > > 2014-01-30 18:24:52:855, 2014-01-30 18:25:48:119, 0, 1024, 200, > > > > 97.66, 1.7671, 100000, 1809.4962 > > > > 2014-01-30 18:24:52:763, 2014-01-30 18:27:45:675, 0, 1024, 200, > > > > 97.66, 0.5648, 100000, 578.3289 > > > > > > > > ====================================== > > > > --request-num-acks -1 > > > > ===================================== > > > > 2014-01-30 18:31:08:756, 2014-01-30 18:57:12:123, 0, 1024, 200, > > > > 97.66, 0.0625, 99998, 63.9632 > > > > 2014-01-30 18:31:08:860, 2014-01-30 18:37:18:634, 0, 1024, 200, > > > > 97.66, 0.2641, 100000, 270.4355 > > > > 2014-01-30 18:31:06:968, 2014-01-30 18:57:28:389, 0, 1024, 200, > > > > 97.66, 0.0618, 99999, 63.2336 > > > > 2014-01-30 18:31:06:819, 2014-01-30 18:37:19:703, 0, 1024, 200, > > > > 97.66, 0.2619, 100000, 268.1799 > > > > > > > > ====================================== > > > > --request-num-acks 1 > > > > ===================================== > > > > 2014-01-30 19:02:35:171, 2014-01-30 19:05:21:650, 0, 1024, 200, > > > > 97.66, 0.5866, 100000, 600.6764 > > > > 2014-01-30 19:02:40:264, 2014-01-30 19:05:01:863, 0, 1024, 200, > > > > 97.66, 0.6897, 100000, 706.2197 > > > > 2014-01-30 19:02:38:537, 2014-01-30 19:03:27:687, 0, 1024, 200, > > > > 97.66, 1.9869, 100000, 2034.5880 > > > > 2014-01-30 19:02:33:384, 2014-01-30 19:04:51:134, 0, 1024, 200, > > > > 97.66, 0.7089, 100000, 725.9528 > > > > > > > > > > > > There is something seriously wrong with Kafka throughput in "2acks" > > > tests. > > > > Anything I can do to resolve this issue? > > > > > > > > One more thing. Producers kept printing error messages like this > > > > during the test runs: > > > > [2014-01-30 18:32:39,777] WARN Failed to send producer request > > > > with correlation id 51676 to broker 1 with data for partitions > > > > [d4111,0] > > > > (kafka.producer.async.DefaultEventHandler) > > > > java.net.SocketTimeoutException > > > > [2014-01-30 18:32:53,308] WARN Produce request with correlation id > > > > 57999 failed due to [d4111,0]: > > > > kafka.common.RequestTimedOutException > > > > (kafka.producer.async.DefaultEventHandler) > > > > [2014-01-30 18:33:22,350] WARN Failed to send producer request > > > > with correlation id 72096 to broker 1 with data for partitions > > > > [d4111,0] > > > > (kafka.producer.async.DefaultEventHandler) > > > > java.net.SocketTimeoutException > > > > [2014-01-30 18:33:39,740] WARN Failed to send producer request > > > > with correlation id 81147 to broker 1 with data for partitions > > > > [d4111,0] > > > > (kafka.producer.async.DefaultEventHandler) > > > > java.net.SocketTimeoutException > > > > > > > > Thank you, > > > > Michael Popov > > > > > > > > > > > > -----Original Message----- > > > > From: Jun Rao [mailto:jun...@gmail.com] > > > > Sent: Wednesday, January 29, 2014 9:01 PM > > > > To: users@kafka.apache.org > > > > Subject: Re: Kafka performance test: "--request-num-acks -1" kills > > > > throughput > > > > > > > > Does the result change with just 1 partition? > > > > > > > > Thanks, > > > > > > > > Jun > > > > > > > > > > > > On Wed, Jan 29, 2014 at 4:06 PM, Michael Popov > > > > <mpo...@microsoft.com> > > > > wrote: > > > > > > > > > Hi, > > > > > > > > > > We need a reliable low-latency message queue that can scale. > > > > > Kafka looks like a right system for this role. > > > > > > > > > > I am running performance tests on multiple platforms: Linux and > > > Windows. > > > > > For test purposes I create topics with 2 replicas and multiple > > > > partitions. > > > > > In all deployments running test producers that wait for both > > replicas' > > > > > acks practically kills Kafka throughput. For example, on the > > > > > following deployment on Linux machines: 2 Kafka brokers, 1 > > > > > Zookeeper node, 4 client hosts to create load, 4 topics with 10 > > > > > partitions each and 2 replicas > > > > > > > > > > - running tests with "--request-num-acks 1" produces > ~3,600 > > > > > msgs/sec > > > > > > > > > > - running tests with "--request-num-acks -1" produces ~348 > > > > > msgs/sec > > > > > > > > > > > > > > > Here is output of one of the four concurrent processes: > > > > > > > > > > [User@Client2 kafka_2.8.0-0.8.0]$ > > > > > bin/kafka-producer-perf-test.sh --broker-list > > > > > 10.0.0.8:9092,10.0.0.10:9092 --compression-codec 0 > > > > > --message-size 1024 --request-num-acks -1 --sync --messages > > > > > 100000 -threads > > > > > 10 --show-detailed-stats --reporting-interval 1000 --topics c12 > > > > > | grep -v "at " > > > > > start.time, end.time, compression, message.size, batch.size, > > > > > total.data.sent.in.MB, MB.sec, total.data.sent.in.nMsg, nMsg.sec > > > > > [2014-01-29 23:21:16,720] WARN Property reconnect.interval is > > > > > not valid > > > > > (kafka.utils.VerifiableProperties) > > > > > [2014-01-29 23:21:16,825] WARN Property reconnect.interval is > > > > > not valid > > > > > (kafka.utils.VerifiableProperties) > > > > > [2014-01-29 23:21:16,830] WARN Property reconnect.interval is > > > > > not valid > > > > > (kafka.utils.VerifiableProperties) > > > > > [2014-01-29 23:21:16,831] WARN Property reconnect.interval is > > > > > not valid > > > > > (kafka.utils.VerifiableProperties) > > > > > [2014-01-29 23:21:16,839] WARN Property reconnect.interval is > > > > > not valid > > > > > (kafka.utils.VerifiableProperties) > > > > > [2014-01-29 23:21:16,841] WARN Property reconnect.interval is > > > > > not valid > > > > > (kafka.utils.VerifiableProperties) > > > > > [2014-01-29 23:21:16,847] WARN Property reconnect.interval is > > > > > not valid > > > > > (kafka.utils.VerifiableProperties) > > > > > [2014-01-29 23:21:16,858] WARN Property reconnect.interval is > > > > > not valid > > > > > (kafka.utils.VerifiableProperties) > > > > > [2014-01-29 23:21:16,862] WARN Property reconnect.interval is > > > > > not valid > > > > > (kafka.utils.VerifiableProperties) > > > > > [2014-01-29 23:21:16,867] WARN Property reconnect.interval is > > > > > not valid > > > > > (kafka.utils.VerifiableProperties) > > > > > [2014-01-29 23:32:03,830] WARN Produce request with correlation > > > > > id > > > > > 11467 failed due to [c12,2]: > > > > > kafka.common.RequestTimedOutException > > > > > (kafka.producer.async.DefaultEventHandler) > > > > > [2014-01-29 23:32:03,831] WARN Produce request with correlation > > > > > id > > > > > 11859 failed due to [c12,8]: > > > > > kafka.common.RequestTimedOutException > > > > > (kafka.producer.async.DefaultEventHandler) > > > > > [2014-01-29 23:32:03,831] WARN Failed to send producer request > > > > > with correlation id 11819 to broker 0 with data for partitions > > > > > [c12,8] > > > > > (kafka.producer.async.DefaultEventHandler) > > > > > java.net.SocketTimeoutException > > > > > [2014-01-29 23:32:03,834] WARN Failed to send producer request > > > > > with correlation id 11315 to broker 0 with data for partitions > > > > > [c12,6] > > > > > (kafka.producer.async.DefaultEventHandler) > > > > > java.net.SocketTimeoutException > > > > > [2014-01-29 23:32:03,834] WARN Failed to send producer request > > > > > with correlation id 11191 to broker 0 with data for partitions > > > > > [c12,4] > > > > > (kafka.producer.async.DefaultEventHandler) > > > > > java.net.SocketTimeoutException > > > > > [2014-01-29 23:32:03,834] WARN Failed to send producer request > > > > > with correlation id 11791 to broker 0 with data for partitions > > > > > [c12,4] > > > > > (kafka.producer.async.DefaultEventHandler) > > > > > java.net.SocketTimeoutException > > > > > [2014-01-29 23:32:03,834] WARN Failed to send producer request > > > > > with correlation id 11395 to broker 0 with data for partitions > > > > > [c12,6] > > > > > (kafka.producer.async.DefaultEventHandler) > > > > > java.net.SocketTimeoutException > > > > > [2014-01-29 23:32:03,834] WARN Failed to send producer request > > > > > with correlation id 11631 to broker 0 with data for partitions > > > > > [c12,4] > > > > > (kafka.producer.async.DefaultEventHandler) > > > > > java.net.SocketTimeoutException > > > > > [2014-01-29 23:32:03,834] WARN Failed to send producer request > > > > > with correlation id 10563 to broker 0 with data for partitions > > > > > [c12,0] > > > > > (kafka.producer.async.DefaultEventHandler) > > > > > java.net.SocketTimeoutException > > > > > [2014-01-29 23:32:03,834] WARN Failed to send producer request > > > > > with correlation id 10907 to broker 0 with data for partitions > > > > > [c12,2] > > > > > (kafka.producer.async.DefaultEventHandler) > > > > > java.net.SocketTimeoutException > > > > > 2014-01-29 23:21:16:562, 2014-01-29 23:40:15:886, 0, 1024, 200, > > > > > 97.66, 0.0857, 100000, 87.7713 > > > > > > > > > > The test result is consistent and reproducible in all deployments: > > > > > numbers can vary but changing acks setting consistently reduces > > > > > Kafka throughput > > > > > 4-10 times. > > > > > > > > > > Is it expected system behavior? Any tuning options to resolve > > > > > the > > > > problem? > > > > > > > > > > Thank you, > > > > > Michael Popov > > > > > > > > > > > > > > > > > > > >
