[
https://issues.apache.org/jira/browse/KAFKA-3101?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15366418#comment-15366418
]
Guozhang Wang commented on KAFKA-3101:
--------------------------------------
[~bbejeck] Regarding KAFKA-3776, there is still one open question about
caching, whether we should store objects or bytes, which is related to memory
management. You can find the details about it in the linked discussion thread:
https://cwiki.apache.org/confluence/display/KAFKA/Discussion%3A+Memory+Management+in+Kafka+Streams#Discussion:MemoryManagementinKafkaStreams-OpenQuestions
So it will be great if you could first do some experiment on performance
comparison with object size estimation (e.g. https://github.com/jbellis/jamm)
v.s. serde cost to in order to make this decision first.
> Optimize Aggregation Outputs
> ----------------------------
>
> Key: KAFKA-3101
> URL: https://issues.apache.org/jira/browse/KAFKA-3101
> Project: Kafka
> Issue Type: Improvement
> Components: streams
> Reporter: Guozhang Wang
> Labels: architecture
> Fix For: 0.10.1.0
>
>
> Today we emit one output record for each incoming message for Table /
> Windowed Stream Aggregations. For example, say we have a sequence of
> aggregate outputs computed from the input stream (assuming there is no agg
> value for this key before):
> V1, V2, V3, V4, V5
> Then the aggregator will output the following sequence of Change<newValue,
> oldValue>:
> <V1, null>, <V2, V1>, <V3, V2>, <V4, V3>, <V5, V4>
> where could cost a lot of CPU overhead computing the intermediate results.
> Instead if we can let the underlying state store to "remember" the last
> emitted old value, we can reduce the number of emits based on some configs.
> More specifically, we can add one more field in the KV store engine storing
> the last emitted old value, which only get updated when we emit to the
> downstream processor. For example:
> At Beginning:
> Store: key => empty (no agg values yet)
> V1 computed:
> Update Both in Store: key => (V1, V1), Emit <V1, null>
> V2 computed:
> Update NewValue in Store: key => (V2, V1), No Emit
> V3 computed:
> Update NewValue in Store: key => (V3, V1), No Emit
> V4 computed:
> Update Both in Store: key => (V4, V4), Emit <V4, V1>
> V5 computed:
> Update NewValue in Store: key => (V5, V4), No Emit
> One more thing to consider is that, we need a "closing" time control on the
> not-yet-emitted keys; when some time has elapsed (or the window is to be
> closed), we need to check for any key if their current materialized pairs
> have not been emitted (for example <V5, V4> in the above example).
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
