Status

Current state: Accepted

Discussion thread: DISCUSS KIP-1052 

JIRA: KAFKA-17645

Motivation

When analyzing performance of long-running applications like Kafka, it is useful to separate the application's lifecycle into phases such as initialization, steady-state, transient-states due to changes in system load, and shut-down. Of these states, steady-state and transient-response-states are of special interest since these represent the normal operating states of the application. To measure the steady state performance of Kafka, one would typically set up a cluster, then create topics and proceed to measure the performance of various load points on their cluster. 
 
A problem arises, however, when using the producer performance test to measure steady-state performance. In our observations, the producer performance test has high latency during the initialization phase as producers are coming online and brokers are beginning to exchange replicated messages. Currently, all measurements with the producer performance test will include measurements of the high latency that occurs during startup in any evaluation of steady-state performance, especially when test durations are not sufficiently long. 

For example, if we observe high startup-related latency that lasts for 30 seconds into the producer performance test (about 6 prints of the 5-second window), we would need to collect at least 3000 seconds (about 50 minutes) of data for the startup latency to represent less than 1% of the latency (p99) and not affect our analysis of the steady-state latency.  This 3000-second test is still problematic in the evaluation of steady-state, however, since we cannot guarantee that high startup latency will be completely resolved by the 30-second (<1%) mark, causing even these long tests to have unstable results.

Consider, however, an example of testing that seeks to understand steady state performance. One test runs for 5 minutes compared to a test with the same configuration that runs for 50 minutes. We would observe the 5-minute test has a p99 latency higher than the 50-minute test even though the steady state latency is the same. Thus, for a user to accurately measure Kafka steady-state performance, they must gather very long tests to get a “clean” p99, uncontaminated by startup latency. 
 
When analyzing steady-state latency in Kafka, it is desirable to not include startup-latency in your measurements.  Therefore, we propose adding a mechanism to optionally separate the statistics for data gathered on the first N records during warmup from those measurements collected during steady-state measurements.

Public Interfaces

The producer performance test will add a new command line argument, --warmup-records, which will describe the number of total records from num-records, that should be treated as warmup and removed from the steady-state set of records.

Proposed Changes

We propose a modification of the producer performance test to add a command line argument for "warmup records". The total number of records for the test must be strictly greater than the number of warmup records or the test should throw an error. Negative values for warmup records should also throw an error. A value of "0" for warmup records will have the test behave identically to not specifying the parameter. Messages sent during warmup would be accumulated in a separate Stats object from the post-warmup records that make up the steady state portion of the test.

For reporting, when the warmup-records option is invoked, an additional summary line will be printed at the end of the run. The first line printed would look identical to the existing summary print, with statistics that describe the performance of the whole test including both warmup and steady-state. This first print represents the performance of the test without a warmup and would enable relatability of new experiments to older tests that did not separate out steady state performance. The second summary line printed when warmup is invoked is the summary of the post-warmup, steady-state portion of the test. This steady-state performance could then be used to understand steady-state behavior in the absence of cluster startup costs. 

Some examples of command lines and their resulting outputs follow.  All data is fabricated but are representative of differences observed between warmup and steady state.

When run with no warmup records, the printed summary would look identical to previous versions:  

$ bin/kafka-producer-perf-test.sh --payload-file /opt/kafka/payload.txt --num-records 1000000 --throughput 50000
Reading payloads from: /opt/kafka/payload.txt
Number of messages read: 5000
<snip producer perf test time-series log, summary stats lines follow>
1000000 records sent, 50000.123456 records/sec (49.44 MB/sec), 5.50 ms avg latency, 600.00 ms max latency, 5 ms 50th, 10 ms 95th, 50 ms 99th, 200 ms 99.9th.

When run with warmup-records, the summary line printed first would be the whole test, similar to current behavior, followed by a second steady-state-only summary line:

$ bin/kafka-producer-perf-test.sh --payload-file /opt/kafka/payload.txt --warmup-records 100000 --num-records 1000000 --throughput 50000
Reading payloads from: /opt/kafka/payload.txt
Number of messages read: 5000
Warmup first 100000 records. Steady-state results will print after the complete-test summary.
<snip producer perf test time-series log, summary stats lines follow>
1000000 records sent, 50000.123456 records/sec (49.44 MB/sec), 5.50 ms avg latency, 600.00 ms max latency, 5 ms 50th, 10 ms 95th, 50 ms 99th, 200 ms 99.9th.
900000 steady state records sent, 50000.012345 records/sec (49.32 MB/sec), 5.00 ms avg latency, 95.00 ms max latency, 5 ms 50th, 8 ms 95th, 21 ms 99th, 55 ms 99.9th.

Regarding the process of choosing a warmup duration, the warmup records should comprise the time required for producers to startup, allocate resources, connect to the brokers, then reach steady state performance. Although a one-minute warmup is common in Java performance analysis, a simple process for more accurately determining the warmup time of a given topology would be to run the test without a warmup first. The resulting time-series of the producer performance test output would typically have high latency during the first few windows of output (at 5 s per window). The minimum warmup records could then be chosen to be the number of records sent once producer latency reaches an acceptable steady state. Although the producer performance test output should provide sufficient information to set a warmup, additional data sources such as JMX:producer:record-queue-time, JMX:producer:request-latency-avg, and the various broker queue-time metrics could be consulted to help precisely determine when steady state has been achieved. Since the warmup data will be reported separately from the desired steady state results, and running longer warmups should not impact steady state performance, it's always good practice to run a warmup for slightly longer than is deemed necessary to guarantee records from the warmup have been processed and ensure steady state operation.

As of this writing, the producer performance test seeks to meet its target throughput by evaluating the total time elapsed and total records sent since the producer started up. This could lead to higher steady-state producer throughput than expected if the producer is sending slowly during the warmup phase. This behavior may be a desired artifact of the producer's operation in some test cases, but could cause confusion, particularly in cases where warmups and test durations are short. In general, once the producer has reached its intended steady state throughput and does not vary too much between print windows of the test, that can be considered a sufficient warmup.

Compatibility, Deprecation, and Migration Plan

The proposed modification to the producer performance test would not interfere with the previous methods and command lines used to measure producer performance. Those mechanisms would not be deprecated, but the users would have a choice to either include Kafka warmup in their results or avoid the warmup latency and focus on steady-state performance.

One option will be added to the producer performance test tool. The option, --warmup-records, will be added to the producer performance test to request that the initial records sent in a test be gathered into a separate Stats object from the steady-state records to follow. The default value for warmup records would be 0 to ensure the continued normal operation of the producer perf test. Only the analysis and reporting of the producer performance test may change but will require the user to opt-in and add the warmup-records parameter. The default behavior of the producer performance test would be unchanged to maintain functionality for those depending on the old behavior.

$ bin/kafka-producer-perf-test.sh -h | grep -A1 "-warmup"
--warmup-records WARMUP-RECORDS
    The number of records to treat as warmup at the beginning of the test. If the desired test condition is one of steady state, designating warmup records will enable a second print line to describe statistics of the steady state portion. (default: 0)

Additional work should be described in a subsequent KIP to implement a warmup phase in the consumer performance test. 

Rejected Alternatives

An additional feature for producer warmup would be for the producer to automatically detect when its performance has stabilized then switch modes into steady state. The producer could compare variance between windows then choose to switch to steady state once the between-window variance had reached a threshold at a specified minimum. Mechanisms for automatic detection of producer stability at end of warmup include simple checks such as time-based warmup or latency-based warmup, but monitored performance indicators could also include DNS queries, broker SSL connections, Producer Throttler actions, response intervals from brokers, and even JVM compilation activity. While this could dramatically simplify the use of the producer performance tool, automatic determination of the correct limits at which the producer should switch modes is a complex tuning problem, since that would effectively be defining steady state for all deployments. Due to this additional complexity, we believe that auto-detection of the producer’s warmup phase should be discussed in a subsequent KIP.

Another rejected alternative would be for an additional option that controls whether to separate the warmup portion of the test from the steady-state portion of the test in the printed lines. This option is rejected because, when studying steady-state the performance of the warmup portion of the test is, essentially, nonsensical and does not hold analytical value. The warmup option already causes a steady-state summary to be printed following the whole-test summary print. In contrast, when studying warmup, users can already isolate warmup behavior by running short tests that focus only on the first minute of test warmup.