[ 
https://issues.apache.org/jira/browse/IGNITE-29032?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Dmitry Werner updated IGNITE-29032:
-----------------------------------
    Description: 
With page eviction enabled, an in-memory data region can still overflow with 
IgniteOutOfMemoryException when a large entry or a putAll batch is inserted.

Before an insert, eviction frees pages only until a fixed number of empty pages 
is available (DataRegionConfiguration.emptyPagesPoolSize, 100 by default), 
regardless of how much space the insert needs. An entry or batch that needs 
more pages than that exhausts the region. The only workaround is to raise 
emptyPagesPoolSize to cover the largest entry.

Eviction should take the size of the inserted data into account: before an 
insert, as many pages as the entry or batch needs are reserved, and old entries 
are evicted until that many pages are free. An insert that cannot fit in the 
region at all is rejected with IgniteOutOfMemoryException.

  was:
Improving the guarantees of the page eviction mechanism on in-memory data 
regions to eliminate the possibility of data region overflow. For example, when 
inserting large records, or during batch inserts via putAll.

The mechanism should take into account the volume of data being inserted and, 
accordingly, must ensure that old records are evicted in the same volume.

Currently, there is a possible scenario in which an in-memory data region with 
the pageEviction mechanism enabled may overflow. For example, when inserting 
large records, or during batch inserts via putAll, the data region can overflow.

The likelihood of such an event increases when the page eviction and expiry 
policy mechanisms operate simultaneously.


> Improvement of the page eviction mechanism for in-memory
> --------------------------------------------------------
>
>                 Key: IGNITE-29032
>                 URL: https://issues.apache.org/jira/browse/IGNITE-29032
>             Project: Ignite
>          Issue Type: Improvement
>            Reporter: Dmitry Werner
>            Assignee: Dmitry Werner
>            Priority: Major
>             Fix For: 2.20
>
>          Time Spent: 7h
>  Remaining Estimate: 0h
>
> With page eviction enabled, an in-memory data region can still overflow with 
> IgniteOutOfMemoryException when a large entry or a putAll batch is inserted.
> Before an insert, eviction frees pages only until a fixed number of empty 
> pages is available (DataRegionConfiguration.emptyPagesPoolSize, 100 by 
> default), regardless of how much space the insert needs. An entry or batch 
> that needs more pages than that exhausts the region. The only workaround is 
> to raise emptyPagesPoolSize to cover the largest entry.
> Eviction should take the size of the inserted data into account: before an 
> insert, as many pages as the entry or batch needs are reserved, and old 
> entries are evicted until that many pages are free. An insert that cannot fit 
> in the region at all is rejected with IgniteOutOfMemoryException.



--
This message was sent by Atlassian Jira
(v10.3.26#10030026)

Reply via email to