kafka-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "John Roesler (Jira)" <j...@apache.org>
Subject [jira] [Created] (KAFKA-10091) Improve task idling
Date Wed, 03 Jun 2020 20:29:00 GMT
John Roesler created KAFKA-10091:
------------------------------------

             Summary: Improve task idling
                 Key: KAFKA-10091
                 URL: https://issues.apache.org/jira/browse/KAFKA-10091
             Project: Kafka
          Issue Type: Task
          Components: streams
            Reporter: John Roesler


When Streams is processing a task with multiple inputs, each time it is ready to process a
record, it has to choose which input to process next. It always takes from the input for which
the next record has the least timestamp. The result of this is that Streams processes data
in timestamp order. However, if the buffer for one of the inputs is empty, Streams doesn't
know what timestamp the next record for that input will be.

Streams introduced a configuration "max.task.idle.ms" in KIP-353 to address this issue.

[https://cwiki.apache.org/confluence/display/KAFKA/KIP-353%3A+Improve+Kafka+Streams+Timestamp+Synchronization]

The config allows Streams to wait some amount of time for data to arrive on the empty input,
so that it can make a timestamp-ordered decision about which input to pull from next.

However, this config can be hard to use reliably and efficiently, since what we're really
waiting for is the next poll that _would_ return data from the empty input's partition, and
this guarantee is a function of the poll interval, the max poll interval, and the internal
logic that governs when Streams will poll again.

The ideal case is you'd be able to guarantee at a minimum that _any_ amount of idling would
guarantee you poll data from the empty partition if there's data to fetch.



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

Mime
View raw message