[jira] [Commented] (KAFKA-4608) RocksDBWindowStore.fetch() is inefficient for large ranges

["Guozhang Wang (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/KAFKA-4608?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16363350#comment-16363350 ] 

Guozhang Wang commented on KAFKA-4608:
--------------------------------------

I have filed https://issues.apache.org/jira/browse/KAFKA-6560 to tackle on this issue, it aims to only use point queries for window stores than range queries.

> RocksDBWindowStore.fetch() is inefficient for large ranges
> ----------------------------------------------------------
>
>                 Key: KAFKA-4608
>                 URL: https://issues.apache.org/jira/browse/KAFKA-4608
>             Project: Kafka
>          Issue Type: Improvement
>          Components: streams
>    Affects Versions: 0.10.1.1
>            Reporter: Elias Levy
>            Priority: Major
>
> It is not unreasonable for a user to call {{RocksDBWindowStore.fetch}} to scan for a key across a large time range.  For instance, someone may call it with a {{timeFrom}} of zero or a {{timeTo}} of max long in an attempt to fetch keys matching across all time forwards or backwards.  
> But if you do so, {{fetch}} will peg the CPU, as it attempts to iterate over every single segment id in the range. That is obviously very inefficient.  
> {{fetch}} should trim the {{timeFrom}}/{{timeTo}} range based on the available time range in the {{segments}} hash map, so that it only iterates over the available time range.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)