You are viewing a plain text version of this content. The canonical link for it is here.
Posted to jira@kafka.apache.org by "John Roesler (Jira)" <ji...@apache.org> on 2020/02/14 17:57:00 UTC

[jira] [Created] (KAFKA-9557) Thread-level "process" metrics are computed incorrectly

John Roesler created KAFKA-9557:
-----------------------------------

             Summary: Thread-level "process" metrics are computed incorrectly
                 Key: KAFKA-9557
                 URL: https://issues.apache.org/jira/browse/KAFKA-9557
             Project: Kafka
          Issue Type: Bug
          Components: streams
            Reporter: John Roesler
            Assignee: John Roesler


Among others, Streams reports the following two thread-level "process" metrics:

"process-rate": The average number of process calls per second.
"process-total": The total number of process calls across all tasks.

See the docs: https://kafka.apache.org/documentation/#kafka_streams_thread_monitoring

There's some surprising ambiguity in these definitions that has led to Streams actually reporting something different than what most people would probably expect. Specifically, it's not defined what a "process call" is.

A reasonable definition of a "process call" is processing a record or processing a task (both of which are publicly facing concepts, and both of which are the same, since tasks process records one at a time). However, we currently measure number of invocations to a private, internal `process()` method, which would actually process more than one record at a time. Thus, the current metric is under-counting the throughput, in an esoteric and confusing way.

Instead, we should simply change the rate and total metrics to measure the (rate and total) of _record_ processing.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)