[
https://issues.apache.org/jira/browse/FLINK-15160?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17272685#comment-17272685
]
Dawid Wysakowicz commented on FLINK-15160:
------------------------------------------
An idea for a fix is we could add a method to the {{TimerService}} for
registering timer for timeout:
{code}
public interface TimerService {
/**
* Current processing time as returned from {@link
org.apache.flink.streaming.api.TimerService}.
*/
long currentProcessingTime();
void registerTimeoutTimer(long timeoutTime);
}
{code}
and then in the {{NFA#computeNextState}} use this method in to register a timer
for timeout when putting event for a start state:
{code}
if (isStartState(computationState)) {
startTimestamp = event.getTimestamp();
startEventId = event.getEventId();
// register timer for timeout in case no more events
come for that key
timerService.registerTimeoutTimer(startTimestamp +
windowTime);
} else {
startTimestamp = computationState.getStartTimestamp();
startEventId = computationState.getStartEventID();
}
{code}
> Clean up is not applied if there are no incoming events for a key.
> ------------------------------------------------------------------
>
> Key: FLINK-15160
> URL: https://issues.apache.org/jira/browse/FLINK-15160
> Project: Flink
> Issue Type: Bug
> Components: Library / CEP
> Affects Versions: 1.8.2, 1.9.1, 1.10.0
> Reporter: Dawid Wysakowicz
> Priority: Major
>
> In CepOperator the pruning of timed out partial matches happens along with
> feeding events into the NFA. Either when unbuffering on Watermark or
> according to the processing time.
> 1. Processing time
> The state is pruned only with the timestamps of incoming events. If there are
> no incoming events no pruning happens
> 2. Event time
> It is slightly more complicated, but the outcome is similar. We register
> timers that pop events from the buffer, but we do not register any timers for
> when the timeout of a partial match could happen. Therefore if there will be
> no more events we will never prune matches.
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
