[
https://issues.apache.org/jira/browse/IGNITE-28085?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Oleg Valuyskiy updated IGNITE-28085:
------------------------------------
Description:
h1. Description
In in-memory clusters, when a data region reaches its {*}maxSize{*}, Ignite
starts page-based eviction. At the moment this condition can only be observed
in logs via the following warning:
{code:java}
Page-based evictions started. Consider increasing 'maxSize' on Data Region
configuration{code}
(this message is logged from
{*}IgniteCacheDatabaseSharedManager#warnFirstEvict{*}).
However, there is no metric indicating such conditions, which makes it
difficult to monitor eviction pressure using monitoring systems.
h1. Proposed change
Introduce a new boolean metric that becomes true when page eviction is
triggered due to data region memory pressure.
Unlike the existing warning log, which is currently emitted once per node, the
metric is introduced at the data region level. This allows monitoring systems
to identify exactly which data region has experienced memory pressure and
started page eviction.
Each data region will therefore expose its own *EvictionsStarted* metric that
becomes true once eviction is triggered in that region.
h1. Note on logging behavior
Currently the warning about page eviction start is logged only once per node
because the *firstEvictWarn* flag in *IgniteCacheDatabaseSharedManager* is
global. As a result, if eviction later starts in another data region, the
corresponding warning will not appear in the logs.
This behavior may be misleading in configurations with multiple data regions
where eviction can occur independently.
Improving this logging behavior is addressed separately in
[IGNITE-28099|https://issues.apache.org/jira/browse/IGNITE-28099].
was:
h1. Description
In in-memory clusters, when a data region reaches its {*}maxSize{*}, Ignite
starts page-based eviction. At the moment this condition can only be observed
in logs via the following warning:
{code:java}
Page-based evictions started. Consider increasing 'maxSize' on Data Region
configuration{code}
(this message is logged from
{*}IgniteCacheDatabaseSharedManager#warnFirstEvict{*}).
However, there is no metric indicating such conditions, which makes it
difficult to monitor eviction pressure using monitoring systems.
h1. Proposed change
Introduce a new boolean metric that becomes true when page eviction is
triggered due to data region memory pressure.
> Add metric indicating that page eviction has started due to DataRegion memory
> pressure
> --------------------------------------------------------------------------------------
>
> Key: IGNITE-28085
> URL: https://issues.apache.org/jira/browse/IGNITE-28085
> Project: Ignite
> Issue Type: Improvement
> Reporter: Oleg Valuyskiy
> Assignee: Oleg Valuyskiy
> Priority: Minor
> Labels: ise
> Time Spent: 10m
> Remaining Estimate: 0h
>
> h1. Description
> In in-memory clusters, when a data region reaches its {*}maxSize{*}, Ignite
> starts page-based eviction. At the moment this condition can only be observed
> in logs via the following warning:
> {code:java}
> Page-based evictions started. Consider increasing 'maxSize' on Data Region
> configuration{code}
> (this message is logged from
> {*}IgniteCacheDatabaseSharedManager#warnFirstEvict{*}).
> However, there is no metric indicating such conditions, which makes it
> difficult to monitor eviction pressure using monitoring systems.
> h1. Proposed change
> Introduce a new boolean metric that becomes true when page eviction is
> triggered due to data region memory pressure.
> Unlike the existing warning log, which is currently emitted once per node,
> the metric is introduced at the data region level. This allows monitoring
> systems to identify exactly which data region has experienced memory pressure
> and started page eviction.
> Each data region will therefore expose its own *EvictionsStarted* metric that
> becomes true once eviction is triggered in that region.
> h1. Note on logging behavior
> Currently the warning about page eviction start is logged only once per node
> because the *firstEvictWarn* flag in *IgniteCacheDatabaseSharedManager* is
> global. As a result, if eviction later starts in another data region, the
> corresponding warning will not appear in the logs.
> This behavior may be misleading in configurations with multiple data regions
> where eviction can occur independently.
> Improving this logging behavior is addressed separately in
> [IGNITE-28099|https://issues.apache.org/jira/browse/IGNITE-28099].
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
