You are viewing an old version of this page. View the current version.

Compare with Current View Page History

« Previous Version 7 Next »

This page is meant as a template for writing a KIP. To create a KIP choose Tools->Copy on this page and modify with your content and replace the heading with the next KIP number and a description of your issue. Replace anything in italics with your own description.

Status

Current state: Draft

Discussion thread: here [Change the link from the KIP proposal email archive to your own email thread]

JIRA: KAFKA-7722

Please keep the discussion on the mailing list rather than commenting on the wiki (wiki discussions get unwieldy fast).

Motivation

The ProducerPerformance tool is the standard for benchmarking Kafka producer send performance. The tool uses sampling and Java’s System.currentTimeMillis() to record latencies and provide percentiles at the end of each run. This method works well in most cases but results in a lack of precision.

This is a problem in low-latency environments in which sending a message takes at most a couple milliseconds. For example, if a user wants to run a benchmark test comparing two different Kafka cluster setups, it is hard to compare results if low latency values are observed.

Example Scenario:

Benchmark #1: average latency = 2 ms

Benchmark #2: average latency = 3 ms

Since System.currentTimeMillis() gives values with millisecond precision:

  • A result of 2 ms can have a true value between [1.5, 2.5) before rounding
  • A result of 3 ms can have a true value between [2.5, 3.5) before rounding

If we want to calculate the increase in latency observed, then we have a big spread of either:

  • (1) Smallest spread: 2.5 - (2.5) or ~0, resulting in approximately ~0% increase in latency
  • (2) Largest spread: (3.5) - 1.5 or ~2, resulting in approximately 2/2 or ~100% increase in latency

This means we cannot effectively compare these two benchmark results as the latency in #2 is anywhere from 0 to 100% higher than in #1, which is a big spread and results in very different implications.

To achieve a higher precision to handle low-latency environments better, we can use Java’s System.nanoTime().

Public Interfaces


Proposed Changes

This KIP proposes using nanoTime() to record latencies to provide higher precision.

Latency will be recorded in microseconds by truncating the result of calling nanoTime() by 3 digits.

long startTimeUs = System.nanoTime() / 1000;     // value in microseconds
// Send message…
// Message sent and callback called…
long latency = System.nanoTime() / 1000 - startTimeUs;     // value in microseconds


The code change required in this KIP is rather small, but this small change has a few considerations to discuss and address.

Considerations

Consideration #1: How Much Precision?



Compatibility, Deprecation, and Migration Plan

  • What impact (if any) will there be on existing users?
  • If we are changing behavior how will we phase out the older behavior?
  • If we need special migration tools, describe them here.
  • When will we remove the existing behavior?

Rejected Alternatives

If there are alternative ways of accomplishing the same thing, what were they? The purpose of this section is to motivate why the design is the way it is and not some other way.

  • No labels