flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Gábor Gévay <gga...@gmail.com>
Subject Re: How to get top N elements in a DataSet?
Date Tue, 24 Jan 2017 10:49:38 GMT
Hello,

Btw. there is a Jira about this:
https://issues.apache.org/jira/browse/FLINK-2549
Note that the discussion there suggests a more efficient approach,
which doesn't involve sorting the entire partitions.

And if I remember correctly, this question comes up from time to time
on the mailing list.

Best,
Gábor



2017-01-24 11:35 GMT+01:00 Fabian Hueske <fhueske@gmail.com>:
> Hi Ivan,
>
> I think you can use MapPartition for that.
> So basically:
>
> dataset // assuming some partitioning that can be reused to avoid a shuffle
>   .sortPartition(1, Order.DESCENDING)
>   .mapPartition(new ReturnFirstTen())
>   .sortPartition(1, Order.DESCENDING).parallelism(1)
>   .mapPartition(new ReturnFirstTen())
>
> Best, Fabian
>
>
> 2017-01-24 10:10 GMT+01:00 Ivan Mushketyk <ivan.mushketik@gmail.com>:
>>
>> Hi,
>>
>> I have a dataset of tuples with two fields ids and ratings and I need to
>> find 10 elements with the highest rating in this dataset. I found a
>> solution, but I think it's suboptimal and I think there should be a better
>> way to do it.
>>
>> The best thing that I came up with is to partition dataset by rating, sort
>> locally and write the partitioned dataset to disk:
>>
>> dataset
>> .partitionCustom(new Partitioner<Double>() {
>>   @Override
>>   public int partition(Double key, int numPartitions) {
>>     return key.intValue() % numPartitions;
>>   }
>> }, 1) . // partition by rating
>> .setParallelism(5)
>> .sortPartition(1, Order.DESCENDING) // locally sort by rating
>> .writeAsText("..."); // write the partitioned dataset to disk
>>
>> This will store tuples in sorted files with names 5, 4, 3, ... that
>> contain ratings in ranges (5, 4], (4, 3], and so on. Then I can read sorted
>> data from disk and and N elements with the highest rating.
>> Is there a way to do the same but without writing a partitioned dataset to
>> a disk?
>>
>> I tried to use "first(10)" but it seems to give top 10 items from a random
>> partition. Is there a way to get top N elements from every partition? Then I
>> could locally sort top values from every partition and find top 10 global
>> values.
>>
>> Best regards,
>> Ivan.
>>
>>
>

Mime
View raw message