[
https://issues.apache.org/jira/browse/IGNITE-28882?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Oleg Zinoviev updated IGNITE-28882:
-----------------------------------
Affects Version/s: 2.19
> Window Functions - Thread Starvation & Redundant Row Size Calculation
> ---------------------------------------------------------------------
>
> Key: IGNITE-28882
> URL: https://issues.apache.org/jira/browse/IGNITE-28882
> Project: Ignite
> Issue Type: Improvement
> Affects Versions: 2.19
> Reporter: Oleg Zinoviev
> Assignee: Oleg Zinoviev
> Priority: Major
> Time Spent: 1h 50m
> Remaining Estimate: 0h
>
> The current implementation of window functions in the scope of IGNITE-14777
> does not address the following two issues:
> # *Thread starvation risk with large partitions.* When a partition contains
> a large number of rows, the computation thread may get blocked for an
> extended period, causing other execution nodes to experience resource
> starvation.
> # *Redundant row size calculation.* The row size is computed multiple times
> to control memory consumption, which introduces unnecessary overhead.
> h4. Required changes
> * Implement proper release of the execution thread when processing large
> partitions so that other execution nodes do not experience starvation.
> * Send computed rows directly to the downstream node to avoid intermediate
> buffering and redundant calculations.
> * Ensure that the row size is calculated exactly once per row to eliminate
> redundant computations.
> * In streaming mode for window functions, do not include row size
> calculation as part of the computation logic, since in this mode the Node
> does not store rows from upstream.
>
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
