[
https://issues.apache.org/jira/browse/PHOENIX-3131?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15405078#comment-15405078
]
James Taylor commented on PHOENIX-3131:
---------------------------------------
Those two queries are different, [~tuyuri]. The first one:
{code}
select url,sum(pageview) as pv FROM pageview_site where dt > to_date
('2016-06-01') group by url limit 100 offset 2
{code}
will push the limit to the server and only look at the first 100 groups it
finds.
The second one is more of a topN query, but post aggregation:
{code}
select url,sum(pageview) as pv FROM pageview_site where dt > to_date
('2016-06-01') group by url order by pv desc limit 100 offset 2
{code}
will not be able to push the limit to the server in the same way because the
ordering is done on the client (as it orders through a merge sort done on the
client). So, in other words, it needs to go through all groups in order to find
the top 100.
You could try adding a secondary index on url, pv (asc or desc). Might need to
order by url, pv desc. Would have to play around with it a bit.
> improve "order by " performance with aggregated query
> ------------------------------------------------------
>
> Key: PHOENIX-3131
> URL: https://issues.apache.org/jira/browse/PHOENIX-3131
> Project: Phoenix
> Issue Type: Improvement
> Affects Versions: 4.8.0
> Reporter: tu nguyen khac
> Priority: Critical
>
> I created a table in phoenix with query : ( 4 node , ram 8gb, 4 cores / node
> )
> CREATE TABLE pageview_site (
> url varchar(255) not null,
> pageview bigint,
> dt date not null,
> CONSTRAINT PK PRIMARY KEY (url, dt ROW_TIMESTAMP)
> ) SALT_BUCKETS = 4;
> After that :
> 1. I tried to upsert about : 13 milions rows to this table .
> 2. Run 2 queries :
> a. select url,sum(pageview) as pv FROM pageview_site where dt > to_date
> ('2016-06-01') group by url limit 100 offset 2;
> the duration this query in about : 0.5 second
> b. select url,sum(pageview) as pv FROM pageview_site where dt > to_date
> ('2016-06-01') group by ur order by pv descl limit 100 offset 2;
> the duration this query in about : 9.5 seconds
> what happens with 2nd query ?? I think we should improve performance for
> "order by " command
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
