[
https://issues.apache.org/jira/browse/APEXMALHAR-2244?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
David Yan updated APEXMALHAR-2244:
----------------------------------
Description:
The spillable data structures currently does not make any assumption about the
key that is used in Managed State, and as a result, it uses ManagedStateImpl to
interface with Managed State and uses time buckets that are based on the apex
window id. But for WindowedStorage used by WindowedOperator, the key to the
storage is a window, which is event time based. Using the default
ManagedStateImpl would be very inefficient for event time based keys, since it
would write data that would belong to the same window to different time buckets.
On a high level, the below summarizes roughly what needs to be done:
1. a way to tell the spillable data structures to use the
ManagedTimeUnifiedStateImpl
2. a way to tell the spillable data structures how to extract the timestamp
from the key. Note that in the case of WindowedOperator, the timestamp should
be the end timestamp of the window (beginTimeMillis + durationMillis), not the
begin timestamp.
3. a way to tell the spillable data structures how to assign the time bucket
given that timestamp
4. with point 3, the spillable implementations of WindowedStorage will need to
take a config parameter that says how much time (in millis) is each time bucket
5. only purge a time bucket when all keys that belong to that time bucket are
removed and the apex window id of the first window in which the keys are all
removed has been committed
was:
The spillable data structures currently does not make any assumption about the
key that is used in Managed State, and as a result, it uses ManagedStateImpl to
interface with Managed State and uses time buckets that are based on the apex
window id. But for WindowedStorage used by WindowedOperator, the key to the
storage is a window, which is event time based. Using the default
ManagedStateImpl would be very inefficient for event time based keys, since it
would write data that would belong to the same window to different time buckets.
On a high level, the below summarizes roughly what needs to be done:
1. a way to tell the spillable data structures to use the
ManagedTimeUnifiedStateImpl
2. a way to tell the spillable data structures how to extract the timestamp
from the key. Note that in the case of WindowedOperator, the timestamp should
be the end timestamp of the window (beginTimeMillis + durationMillis), not the
begin timestamp.
3. a way to tell the spillable data structures how to assign the time bucket
given that timestamp
4. only purge a time bucket when all keys that belong to that time bucket are
removed and the apex window id of the first window in which the keys are all
removed has been committed
> Optimize WindowedStorage and Spillable data structures for time series
> ----------------------------------------------------------------------
>
> Key: APEXMALHAR-2244
> URL: https://issues.apache.org/jira/browse/APEXMALHAR-2244
> Project: Apache Apex Malhar
> Issue Type: Sub-task
> Reporter: David Yan
> Assignee: Siyuan Hua
>
> The spillable data structures currently does not make any assumption about
> the key that is used in Managed State, and as a result, it uses
> ManagedStateImpl to interface with Managed State and uses time buckets that
> are based on the apex window id. But for WindowedStorage used by
> WindowedOperator, the key to the storage is a window, which is event time
> based. Using the default ManagedStateImpl would be very inefficient for event
> time based keys, since it would write data that would belong to the same
> window to different time buckets.
> On a high level, the below summarizes roughly what needs to be done:
> 1. a way to tell the spillable data structures to use the
> ManagedTimeUnifiedStateImpl
> 2. a way to tell the spillable data structures how to extract the timestamp
> from the key. Note that in the case of WindowedOperator, the timestamp should
> be the end timestamp of the window (beginTimeMillis + durationMillis), not
> the begin timestamp.
> 3. a way to tell the spillable data structures how to assign the time bucket
> given that timestamp
> 4. with point 3, the spillable implementations of WindowedStorage will need
> to take a config parameter that says how much time (in millis) is each time
> bucket
> 5. only purge a time bucket when all keys that belong to that time bucket are
> removed and the apex window id of the first window in which the keys are all
> removed has been committed
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
