Home | About | Sematext search-lucene.com search-hadoop.com
 Search Hadoop and all its subprojects:

Switch to Threaded View
Pig, mail # user - Loader partitioning on field


Copy link to this message
-
Re: Loader partitioning on field
Jeff Yuan 2013-03-14, 22:56
Well, I have a UDF called "SCHEDULED_TIME()" that returns the time
when a pig query is scheduled to run in the system. This time is
passed in by the system to pig when the job is launched. Since I
partition files by time field, a user could filter based on the result
of this UDF.

On Thu, Mar 14, 2013 at 3:15 PM, Jonathan Coveney <[EMAIL PROTECTED]> wrote:
> No, it is not. But if it knew that, how would that filter be meaningful?
> What do you have in mind?
>
>
> 2013/3/14 Jeff Yuan <[EMAIL PROTECTED]>
>
>> Rohini, I see your point.
>>
>> One followup question: it's possible for the result of a UDF to be
>> constant and not dependent on the tuples of each record, right? Is Pig
>> able to make such a determination in this case and push the pushdown
>> such UDF results to load?
>>
>> Thanks,
>> Jeff
>>
>> On Thu, Mar 14, 2013 at 2:30 PM, Rohini Palaniswamy
>> <[EMAIL PROTECTED]> wrote:
>> > The filter push down to LoadFunc happens on the front end before the job
>> > launch and the UDF is still not evaluated then. So you need to have
>> > constants in your filter condition.
>> >
>> > Logical plan is internal to pig and will never be exposed. Refer
>> > https://issues.apache.org/jira/browse/PIG-3199
>> >
>> > Regards,
>> > Rohini
>> >
>> >
>> > On Thu, Mar 14, 2013 at 2:00 PM, Jeff Yuan <[EMAIL PROTECTED]>
>> wrote:
>> >
>> >> Thanks! Regarding 1), where there is a UDF in the filter step on a
>> >> partition field. The UDF is not first evaluated before and then the
>> >> result passed to the load function?
>> >>
>> >> A separate question: In a LoadFunc, is there a way to get a reference
>> >> to the logical query plan?
>> >>
>> >> Thanks again.
>> >>
>> >> On Thu, Mar 14, 2013 at 1:51 PM, Rohini Palaniswamy
>> >> <[EMAIL PROTECTED]> wrote:
>> >> > Jeff,
>> >> >
>> >> > 1) It should not. If it does push, then it is a bug in pig.
>> >> >
>> >> > 2) I think it should be fine.
>> >> >
>> >> > 3) Look at PColFilterExtractor and PartitionFilterOptimizer
>> >> >
>> >> > Regards,
>> >> >
>> >> > Rohini
>> >> >
>> >> >
>> >> > On Thu, Mar 14, 2013 at 1:31 PM, Jeff Yuan <[EMAIL PROTECTED]>
>> >> wrote:
>> >> >
>> >> >> I am writing a loader for a storage format, which partitions by a
>> >> >> particular field in the record. So I would like to implement
>> something
>> >> >> which can push down filters on the partitioned field so that the
>> >> >> record reader does not need to read files that are outside the
>> >> >> filtered range. In the interface "LoadMetadata", the
>> >> >> "getPartitionKeys" and "setPartitionFilter" functions seem to support
>> >> >> what I need (where Pig should pass the filtering expression on the
>> >> >> declared partition keys to "setPartitionFilter", but I have a couple
>> >> >> of questions. I'm going to reference the following example, where
>> >> >> timestamp is the partition key.
>> >> >>
>> >> >> a = load 'stored_data' using CustomLoader();
>> >> >> b = filter a by timestamp = CUSTOM_UDF(date, month);
>> >> >>
>> >> >> 1. Would partitioning work in this case where the partition key
>> filter
>> >> >> includes a UDF?
>> >> >>
>> >> >> 2. Does the partition statement need to be directly after the load
>> >> >> statement? What I mean is, if I declare a variable c between a and b
>> >> >> which does some other operation on a, will Pig pass the filter
>> >> >> expression of b when loading a?
>> >> >>
>> >> >> 3. Can you point out roughly where this "setPartitionFilter" function
>> >> >> is called in Pig code during the load process? I couldn't seem to
>> find
>> >> >> it through a search of the Pig source.
>> >> >>
>> >> >> Thanks a lot!
>> >> >>
>> >>
>>