You are viewing a plain text version of this content. The canonical link for it is here.
Posted to dev@datafu.apache.org by "Eyal Allweil (JIRA)" <ji...@apache.org> on 2016/05/09 08:50:12 UTC
[jira] [Comment Edited] (DATAFU-117) New UDF - CountDistinctUpTo
[ https://issues.apache.org/jira/browse/DATAFU-117?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15258239#comment-15258239 ]
Eyal Allweil edited comment on DATAFU-117 at 5/9/16 8:50 AM:
-------------------------------------------------------------
Ok, I opened a review board for it - It's at https://reviews.apache.org/r/46701/
I think all your previous comments are addressed there, except for the one about "this.set.add(o) && (this.set.size() == maxAmount". I don't think this can exceed the max size, because a single add operation can only increment the set's size by one, and the UDF is executed in a single thread.
I ran a few tests comparing this UDF to a Pig nested foreach with DISTINCT followed by the builtin COUNT. On small inputs they perform about the same - even up to a million records - but if you have a situation with more skew (I checked 10 million records, with about 4 million distincts) then this UDF with a max value of say, 1,000,000, runs in a few minutes, and the nested foreach+distinct+count takes more than an hour - probably because it needs to keep all the distinct records in memory, rather than just reaching the desired threshold.
was (Author: eyal):
Ok, I opened a review board for it - can you see it? It's at https://reviews.apache.org/r/46701/
I think all your previous comments are addressed there, except for the one about "this.set.add(o) && (this.set.size() == maxAmount". I don't think this can exceed the max size, because a single add operation can only increment the set's size by one, and the UDF is executed in a single thread.
I ran a few tests comparing this UDF to a Pig nested foreach with DISTINCT followed by the builtin COUNT. On small inputs they perform about the same - even up to a million records - but if you have a situation with more skew (I checked 10 million records, with about 4 million distincts) then this UDF with a max value of say, 10000, runs in about four minutes, and the nested foreach+distinct+count takes more than an hour - probably because it needs to keep all the distinct records in memory, rather than just reaching the desired threshold.
> New UDF - CountDistinctUpTo
> ---------------------------
>
> Key: DATAFU-117
> URL: https://issues.apache.org/jira/browse/DATAFU-117
> Project: DataFu
> Issue Type: New Feature
> Reporter: Eyal Allweil
> Attachments: DATAFU-117-2.patch, DATAFU-117.patch
>
>
> A UDF that counts distinct tuples within a bag, but only up to a preset limit. If the bag contains more distinct tuples than the limit, the UDF returns the limit.
> This UDF can run reasonably well even on large bags if the limit chosen is small enough though the count is done in memory.
> We use this UDF in PayPal for filtering, when we don't need to use the actual tuples afterward.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)