flink-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Fabian Hueske <fhue...@gmail.com>
Subject Re: dataset sort
Date Fri, 09 Feb 2018 15:49:36 GMT
The reason why this isn't working in Flink are that

* a file can only be written by a single process
* Flink does not support merging of sorted network partitions but reads
round-robin from incoming network channels.

I think if you sort the historic data in parallel (without range
partitioning, i.e., randomly partitioned) and write it out in multiple
files, you could implement a source function that reads all files in
parallel and generates ascending watermarks.
It would be important that you have as many parallel source tasks as you
have files to ensure that watermarks are properly generated. Apart from
that, this should result in a nicely sorted stream.
The watermark handling of the DataStream API will take care to "merge" the
sorted files.

Best, Fabian

2018-02-09 16:23 GMT+01:00 david westwood <david.d.westwood@gmail.com>:

> Thanks.
> I have to stream in the historical data and its out-of-boundedness >>
> real-time data. I thought there was some elegant way using mapPartition
> that I wasn't seeing.
> On Fri, Feb 9, 2018 at 5:10 AM, Fabian Hueske <fhueske@gmail.com> wrote:
>> You can also partition by range and sort and write each partition. Once
>> all partitions have been written to files, you can concatenate the files.
>> As Till said it is not possible to sort in parallel and write in order to
>> a single file.
>> Best, Fabian
>> 2018-02-09 10:35 GMT+01:00 Till Rohrmann <trohrmann@apache.org>:
>>> Hi David,
>>> Flink only supports sorting within partitions. Thus, if you want to
>>> write out a globally sorted dataset you should set the parallelism to 1
>>> which effectively results in a single partition. Decreasing the
>>> parallelism of an operator will cause the individual partitions to lose its
>>> sort order because the individual partitions are read in a non
>>> deterministic order.
>>> Cheers,
>>> Till
>>> On Thu, Feb 8, 2018 at 8:07 PM, david westwood <
>>> david.d.westwood@gmail.com> wrote:
>>>> Hi:
>>>> I would like to sort historical data using the dataset api.
>>>> env.setParallelism(10)
>>>> val dataset = [(Long, String)] ..
>>>> .paritionByRange(_._1)
>>>> .sortPartition(_._1, Order.ASCEDING)
>>>> .writeAsCsv("mydata.csv").setParallelism(1)
>>>> the data is out of order (in local order)
>>>> but
>>>> .print()
>>>> prints the data in to correct order. I have run a small toy sample
>>>> multiple times.
>>>> Is there a way to sort the entire dataset with parallelism > 1 and
>>>> write it to a single file in ascending order?

View raw message