[
https://issues.apache.org/jira/browse/FLINK-6233?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16093041#comment-16093041
]
Fabian Hueske commented on FLINK-6233:
--------------------------------------
Hi [~xccui],
I agree that it would be nice to support arbitrary joins, however I am not sure
that this should be our current focus. I think theta joins are much less common
than equality join and might also be prohibitively expensive to compute. If I
read the code of your demo correctly, you are implementing a broadcast-forward
strategy to compute theta joins and joining the results in a non-keyed
ProcessFunction.
The problem with non-keyed operators is that they cannot use keyed state but
only operator state. However, operator state is not integrated with Flink's
state backends and is hence kept on the JVM heap. So it cannot use RocksDB to
spill to disk. A non-keyed operator that holds too much state on the heap will
fail. In the context of a broadcast join, this is especially important because
one stream will be replicated to all parallel operator and needs to be stored
as state. I would agree to have a somewhat inefficient join would be good as a
fallback but IMO it also needs to be reliable which cannot be achieved the
operator state (at least for now).
Regarding the timestamp issue, I think you got the point. Let's say we have the
following query:
{code}
// Table left: [a, b, ltime]
// Table right: [c, d, rtime]
SELECT *
FROM left, right
WHERE a = c AND ltime BETWEEN rtime - INTERVAL '1' HOUR AND rtime + INTERVAL
'1' HOUR
{code}
{{left}} and {{right}} are streaming tables and their rows are timestamped with
{{ltime}} and {{rtime}} respectively.
When we do {{SELECT *}} it is not clear which of both timestamps should be
assigned to the result of the join. There can only be one row timestamp, the
other timestamp can be materialized though.
Even if we know which timestamp we want to materialize, we have to hold back
the watermarks such that we do not emit late data. Holding back watermarks is a
difficult business though and can not be done with a ProcessFunction but only
in a custom operator.
> 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)
