flink-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Till Rohrmann <trohrm...@apache.org>
Subject Re: RichMapPartitionFunction - problems with collect
Date Tue, 26 Apr 2016 12:25:07 GMT
Hi Sergio,

sorry for the late reply. I figured out your problem. The reason why you
see apparently inconsistent results is that you execute your job multiple
times. Each collect call triggers an eager execution of your Flink job.
Since the intermediate results are not stored the whole topology has to be
re-executed for every collect call. Since the input split assignment of
your libSVM file happens lazily, it can happen that the different sub tasks
get different splits of the input file assigned. Therefore, it happens that
you see different lengths for different features of the same partition.

If you replace the last 6 lines of your program with:

transposed.filter(_._1._1 == (nFeatures - 1)).map(t => t._1 ->
t._2.size).reduceGroup(_.mkString(",")).output(new
PrintingOutputFormat())
val b = transposed.filter(_._1._1 == 10).map(t => t._1 ->
t._2.size).reduceGroup(_.mkString(",")).output(new
PrintingOutputFormat())
val c = transposed.filter(_._1._1 == 12).map(t => t._1 ->
t._2.size).reduceGroup(_.mkString(",")).output(new
PrintingOutputFormat())

env.execute()

you should see the correct results. If you need a deterministic input split
assignment to the different sources, then you would have to implement your
own InputFormat which returns a special InputSplitAssigner which does the
deterministic input split assignment. Or simply try to avoid collect, count
and print which trigger the eager execution of a Flink job.

Cheers,
Till
​

On Wed, Apr 13, 2016 at 5:47 PM, Sergio Ramírez <sramirez@correo.ugr.es>
wrote:

> Hello again:
>
> Any news about this problem with enriched MapPartition function?
>
> Thank you
>
>
> On 06/04/16 17:01, Sergio Ramírez wrote:
>
>> Hello,
>>
>> Ok, please find enclosed the test code and the input data.
>>
>> Cheers
>>
>> On 31/03/16 10:07, Till Rohrmann wrote:
>>
>>> Hi Sergio,
>>>
>>> could you please provide a complete example (including input data) to
>>> reproduce your problem. It is hard to tell what's going wrong when one
>>> only
>>> sees a fraction of the program.
>>>
>>> Cheers,
>>> Till
>>>
>>> On Tue, Mar 29, 2016 at 5:58 PM, Sergio Ramírez <sramirez@correo.ugr.es>
>>> wrote:
>>>
>>> Hi again,
>>>>
>>>> I've not been able to solve the problem with the instruction you gave
>>>> me.
>>>> I've tried with static variables (matrices) also unsuccessfully. I've
>>>> also
>>>> tried this simpler code:
>>>>
>>>>
>>>> def mapPartition(it: java.lang.Iterable[LabeledVector], out:
>>>> Collector[((Int, Int), Int)]): Unit = {
>>>>            val index = getRuntimeContext().getIndexOfThisSubtask() //
>>>> Partition index
>>>>            var ninst = 0
>>>>            for(reg <- it.asScala) {
>>>>              requireByteValues(reg.vector)
>>>>              ninst += 1
>>>>            }
>>>>            for(i <- 0 until nFeatures) out.collect((i, index) -> ninst)
>>>>          }
>>>>
>>>> The result is as follows:
>>>>
>>>> Attribute 10, first seven partitions:
>>>> ((10,0),201),((10,1),200),((10,2),201),((10,3),200),((10,4),200),((10,5),201),((10,6),201),((10,7),201)
>>>>
>>>> Attribute 12, first seven partitions:
>>>> ((12,0),201),((12,1),201),((12,2),201),((12,3),200),((12,4),201),((12,5),200),((12,6),200),((12,7),201)
>>>>
>>>>
>>>> As you can see, for example, for block 6 different number of instances
>>>> are
>>>> shown, but  it's impossible.
>>>>
>>>>
>>>> On 24/03/16 22:39, Chesnay Schepler wrote:
>>>>
>>>> Haven't looked to deeply into this, but this sounds like object reuse is
>>>>> enabled, at which point buffering values effectively causes you to
>>>>> store
>>>>> the same value multiple times.
>>>>>
>>>>> can you try disabling objectReuse using
>>>>> env.getConfig().disableObjectReuse() ?
>>>>>
>>>>> On 22.03.2016 16:53, Sergio Ramírez wrote:
>>>>>
>>>>> Hi all,
>>>>>>
>>>>>> I've been having some problems with RichMapPartitionFunction.
>>>>>> Firstly, I
>>>>>> tried to convert the iterable into an array unsuccessfully. Then,
I
>>>>>> have
>>>>>> used some buffers to store the values per column. I am trying to
>>>>>> transpose
>>>>>> the local matrix of LabeledVectors that I have in each partition.
>>>>>>
>>>>>> None of these solutions have worked. For example, for partition 7
and
>>>>>> feature 10, the vector is empty, whereas for the same partition and
>>>>>> feature
>>>>>> 11, the vectors contains 200 elements. And this change on each
>>>>>> execution,
>>>>>> different partitions and features.
>>>>>>
>>>>>> I think there is a problem with using the collect method out of the
>>>>>> iterable loop.
>>>>>>
>>>>>> new RichMapPartitionFunction[LabeledVector, ((Int, Int),
>>>>>> Array[Byte])]()
>>>>>> {
>>>>>>          def mapPartition(it: java.lang.Iterable[LabeledVector],
out:
>>>>>> Collector[((Int, Int), Array[Byte])]): Unit = {
>>>>>>            val index = getRuntimeContext().getIndexOfThisSubtask()
>>>>>>            val mat = for (i <- 0 until nFeatures) yield new
>>>>>> scala.collection.mutable.ListBuffer[Byte]
>>>>>>            for(reg <- it.asScala) {
>>>>>>              for (i <- 0 until (nFeatures - 1)) mat(i) +=
>>>>>> reg.vector(i).toByte
>>>>>>              mat(nFeatures - 1) += classMap(reg.label)
>>>>>>            }
>>>>>>            for(i <- 0 until nFeatures) out.collect((i, index)
->
>>>>>> mat(i).toArray) // numPartitions
>>>>>>          }
>>>>>>   }
>>>>>>
>>>>>> Regards
>>>>>>
>>>>>>
>>>>>>
>>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message