[
https://issues.apache.org/jira/browse/GEODE-9150?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Darrel Schneider updated GEODE-9150:
------------------------------------
Description:
Partition region expiration usually drives all the expiration from the primary
(the unusual case is when the pr also has evict destroy in which case
expiration is done on primary and secondaries). But the secondaries still have
all the expiration tasks (one for each entry which consumes about 112 bytes of
memory per entry). When they fire on the secondary they go into a different
pendingExpiry map owned by each bucket. They hang around in that map until the
primary expiration happens or a secondary becomes primary. The pendingExpiry
map also eats up some memory.
What we could do instead is no expiration scheduling on secondaries. All
expiration would be driven from the primary and use the normal distribution to
the secondaries. When a secondary becomes primary then it could kick off a
background thread scheduling an expiry task for each entry. Currently when it
becomes primary it processes the pending expiry map so that could be replaced
with scheduling expiry. LocalRegion already has the code that supports
rescheduling expiry since it is something users can dynamically change on an
existing region. This would not only reduce the memory usage on secondaries but
also simplify our expiration implementation. PR eviction currently complicates
our pr expiration implementation but if our pr eviction destroy was always
distributed when the region had expiration, then we would not need any special
expiration code to deal with evict destroy.
was:
Partition region expiration usually drives all the expiration from the primary
(the unusual case is when the pr also has evict destroy in which case
expiration is done on primary and secondaries). But the secondaries still have
all the expiration tasks (one for each entry which consumes about 112 bytes of
memory per entry). When they fire on the secondary they go into a different
pendingExpiry map owned by each bucket. They hang around in that map until the
primary expiration happens or a secondary becomes primary. The pendingExpiry
map also eats up some memory.
What we could do instead is no expiration scheduling on secondaries. When a
secondary becomes primary then it could kick off a background thread scheduling
an expiry task for each entry. Currently when it becomes primary it processes
the pending expiry map so that could be replaced with scheduling expiry.
LocalRegion already has the code that supports rescheduling expiry since it is
something users can dynamically change on an existing region. This would not
only reduce the memory usage on secondaries but also simplify our expiration
implementation. PR eviction currently complicates our pr expiration
implementation but if our pr eviction destroy was always distributed when the
region had expiration, then we would not need any special expiration code to
deal with evict destroy.
> partition region expiration uses extra memory on secondaries
> ------------------------------------------------------------
>
> Key: GEODE-9150
> URL: https://issues.apache.org/jira/browse/GEODE-9150
> Project: Geode
> Issue Type: Improvement
> Components: expiration
> Reporter: Darrel Schneider
> Priority: Major
> Labels: GeodeOperationAPI
>
> Partition region expiration usually drives all the expiration from the
> primary (the unusual case is when the pr also has evict destroy in which case
> expiration is done on primary and secondaries). But the secondaries still
> have all the expiration tasks (one for each entry which consumes about 112
> bytes of memory per entry). When they fire on the secondary they go into a
> different pendingExpiry map owned by each bucket. They hang around in that
> map until the primary expiration happens or a secondary becomes primary. The
> pendingExpiry map also eats up some memory.
> What we could do instead is no expiration scheduling on secondaries. All
> expiration would be driven from the primary and use the normal distribution
> to the secondaries. When a secondary becomes primary then it could kick off a
> background thread scheduling an expiry task for each entry. Currently when it
> becomes primary it processes the pending expiry map so that could be replaced
> with scheduling expiry. LocalRegion already has the code that supports
> rescheduling expiry since it is something users can dynamically change on an
> existing region. This would not only reduce the memory usage on secondaries
> but also simplify our expiration implementation. PR eviction currently
> complicates our pr expiration implementation but if our pr eviction destroy
> was always distributed when the region had expiration, then we would not need
> any special expiration code to deal with evict destroy.
>
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
