[
https://issues.apache.org/jira/browse/CALCITE-3181?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16881121#comment-16881121
]
Haisheng Yuan commented on CALCITE-3181:
----------------------------------------
e.g. compute the total page views for the top 10 hot-sale products in each
category:
{code:java}
select category, product, pv,
sum(pv) over(partition by category order by sale_count desc limit 10) from
prod;
{code}
Instead of sort-based window implementation, we can use hash table based
implementation for window with limit. Use the partition key as the hash key,
each distinct key has an entry that contains a heap if there is order by
clause, with the heap size be the limit value. In distributed system, we can
have 2-phase window (global and local), execute local window before shuffling
tuples by the window partition key, which might reduce the shuffle cost
dramatically.
But the side effect is also obvious: this will break the assumption that window
operator keeps the cardinality unchanged.
> Support limit per group in Window
> ---------------------------------
>
> Key: CALCITE-3181
> URL: https://issues.apache.org/jira/browse/CALCITE-3181
> Project: Calcite
> Issue Type: Improvement
> Components: core
> Reporter: Haisheng Yuan
> Priority: Major
>
> We have a lot of queries like the following to retrieve top N tuples per
> group:
> {code:java}
> SELECT x, y FROM
> (SELECT x, y, ROW_NUMBER() OVER (PARTITION BY x ORDER BY y)
> AS rn FROM t1) t2 WHERE rn <= 3;
> {code}
> The performance is not good if each group has a lot more tuples than wanted,
> because we will retrieve and sort all the tuples, instead of just doing a
> top-N heap sort.
> In order to do optimization for this kind of query, we need to extend window
> to support limit, if and only if there is only 1 window function, and it is
> {{row_number()}}. We also need a substitute rule to push the limit into
> window. Of course, we also need to modify executor to support this
> optimization (can be later).
> {code:java}
> Filter (rn <= 3)
> +- Window (window#0={Partition by x order by y ROW_NUMBER()})
> {code}
> to
> {code:java}
> Filter (rn <= 3)
> +- Window (window#0={Partition by x order by y limit 3 ROW_NUMBER()})
> {code}
> Thoughts? Objections?
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
