[ https://issues.apache.org/jira/browse/FLINK-6233?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16125182#comment-16125182 ]
Xingcan Cui edited comment on FLINK-6233 at 8/14/17 8:52 AM: ------------------------------------------------------------- Hi [~fhueske], thanks for the previous work by you and yuhong, the implementation process is going well. Nevertheless, I've got some minor questions/ideas. # Considering that the core logics of the rowtime inner join and proctime inner join are almost the same. Can I extract an abstract {{TimeWindowInnerJoin}} class and let the {{ProcTimeWindowInnerJoin}} and {{RowTimeWindowInnerJoin}} extend it? # The clean up process for the cached data is triggered by ProcessingTimeTimers in the {{ProcTimeWindowInnerJoin}}. For {{RowTimeWindowInnerJoin}}, I think this process could be directly triggered by the watermarks without registering the EventTimeTimer, right? # Since the collections provided by the state backend are simple, it may be inefficient to search for the out-of-dated records. I think the current "short-circuit" codes (as shown below) can not clean all the expired data. {code:java} while (keyIter.hasNext && !validTimestamp) { val recordTime = keyIter.next if (recordTime < expiredTime) { removeList.add(recordTime) } else { // we found a timestamp that is still valid validTimestamp = true } } {code} To cope with that, I plan to split the "cache window" into continuous static-panes, and casting one to expired as a whole. By doing like that, we may store some extra records, whose time interval is equal to the static span of the panes, but can remove the expired data efficiently. # I'd like to introduce an extra {{allowLateness}} parameter (which can be set in the {{StreamQueryConfig}}) to the join function. But for now, I'll give it a default {{0L}} value. There's an extra problem. As I mentioned before, the two streams in the {{CoProcessOperator}} share the same {{InternalTimeServiceManager}}, which means their watermarks are forcibly synchronized to the lower ones. I know why it is designed so, but still think we should provide separate time services for the two input streams. After all, we can not imagine the rowtimes of the two streams are naturally synchronized. However, I can provide an implementation based on the current mechanism and do further optimizations in the future. What do you think? was (Author: xccui): Hi [~fhueske], thanks for the previous work by you and yuhong, the implementation process is going well. Nevertheless, I've got some minor questions/ideas. # Considering that the core logics of the rowtime inner join and proctime inner join are almost the same. Can I extract an abstract {{TimeWindowInnerJoin}} class and let the {{ProcTimeWindowInnerJoin}} and {{RowTimeWindowInnerJoin}} extend it? # The clean up process for the cached data is triggered by ProcessingTimeTimers in the {{ProcTimeWindowInnerJoin}}. For {{RowTimeWindowInnerJoin}}, I think this process could be directly triggered by the watermarks without registering the EventTimeTimer, right? # Since the collections provided by the state backend are simple, it may be inefficient to search for the out-of-dated records. I think the current "short-circuit" codes (as shown below) can not clean all the expired data. {code:java} while (keyIter.hasNext && !validTimestamp) { val recordTime = keyIter.next if (recordTime < expiredTime) { removeList.add(recordTime) } else { // we found a timestamp that is still valid validTimestamp = true } } {code} To cope with that, I plan to split the "cache window" into continuous static-panes, and casting one to expired as a whole. By doing like that, we may store some extra records, whose time interval is equal to the static span of the panes, but can remove the expired data efficiently. # I'd like to introduce an extra {{allowLateness}} parameter (which can be set in the {{StreamQueryConfig}}) to the join function. But for now, I'll give it a default {{0L}} value. > Support rowtime inner equi-join between two streams in the SQL API > ------------------------------------------------------------------ > > Key: FLINK-6233 > URL: https://issues.apache.org/jira/browse/FLINK-6233 > Project: Flink > Issue Type: Sub-task > Components: Table API & SQL > Reporter: hongyuhong > Assignee: Xingcan Cui > > The goal of this issue is to add support for inner equi-join on proc time > streams to the SQL interface. > Queries similar to the following should be supported: > {code} > SELECT o.rowtime , o.productId, o.orderId, s.rowtime AS shipTime > FROM Orders AS o > JOIN Shipments AS s > ON o.orderId = s.orderId > AND o.rowtime BETWEEN s.rowtime AND s.rowtime + INTERVAL '1' HOUR; > {code} > The following restrictions should initially apply: > * The join hint only support inner join > * The ON clause should include equi-join condition > * The time-condition {{o.rowtime BETWEEN s.rowtime AND s.rowtime + INTERVAL > '1' HOUR}} only can use rowtime that is a system attribute, the time > condition only support bounded time range like {{o.rowtime BETWEEN s.rowtime > - INTERVAL '1' HOUR AND s.rowtime + INTERVAL '1' HOUR}}, not support > unbounded like {{o.rowtime < s.rowtime}} , and should include both two > stream's rowtime attribute, {{o.rowtime between rowtime () and rowtime () + > 1}} should also not be supported. > An row-time streams join will not be able to handle late data, because this > would mean in insert a row into a sorted order shift all other computations. > This would be too expensive to maintain. Therefore, we will throw an error if > a user tries to use an row-time stream join with late data handling. > This issue includes: > * Design of the DataStream operator to deal with stream join > * Translation from Calcite's RelNode representation (LogicalJoin). -- This message was sent by Atlassian JIRA (v6.4.14#64029)