Another issue is that hadooprdd (which sc.textfile uses) might split input
files and even if it doesn't split, it doesn't guarantee that part files
numbers go to the corresponding partition number in the rdd. Eg part-00000
could go to partition 27
On Apr 24, 2015 7:41 AM, "Michal Michalski" <michal.michal...@boxever.com>
wrote:
> Of course after you do it, you probably want to call
> repartition(somevalue) on your RDD to "get your paralellism back".
>
> Kind regards,
> Michał Michalski,
> michal.michal...@boxever.com
>
> On 24 April 2015 at 15:28, Michal Michalski <michal.michal...@boxever.com>
> wrote:
>
>> I did a quick test as I was curious about it too. I created a file with
>> numbers from 0 to 999, in order, line by line. Then I did:
>>
>> scala> val numbers = sc.textFile("./numbers.txt")
>> scala> val zipped = numbers.zipWithUniqueId
>> scala> zipped.foreach(i => println(i))
>>
>> Expected result if the order was preserved would be something like: (0,
>> 0), (1, 1) etc.
>> Unfortunately, the output looks like this:
>>
>> (126,1)
>> (223,2)
>> (320,3)
>> (1,0)
>> (127,11)
>> (2,10)
>> (...)
>>
>> The workaround I found that works for me for my specific use case
>> (relatively small input files) is setting explicitly the number of
>> partitions to 1 when reading a single *text* file:
>>
>> scala> val numbers_sp = sc.textFile("./numbers.txt", 1)
>>
>> Than the output is exactly as I would expect.
>>
>> I didn't dive into the code too much, but I took a very quick look at it
>> and figured out - correct me if I missed something, it's Friday afternoon!
>> ;-) - that this workaround will work fine for all the input formats
>> inheriting from org.apache.hadoop.mapred.FileInputFormat including
>> TextInputFormat, of course - see the implementation of getSplits() method
>> there (
>> http://grepcode.com/file/repo1.maven.org/maven2/org.jvnet.hudson.hadoop/hadoop-core/0.19.1-hudson-2/org/apache/hadoop/mapred/FileInputFormat.java#FileInputFormat.getSplits%28org.apache.hadoop.mapred.JobConf%2Cint%29
>> ).
>> The numSplits variable passed there is exactly the same value as you
>> provide as a second argument to textFile, which is minPartitions. However,
>> while *min* suggests that we can only define a minimal number of
>> partitions, while we have no control over the max, from what I can see in
>> the code, that value specifies the *exact* number of partitions per the
>> FileInputFormat.getSplits implementation. Of course it can differ for other
>> input formats, but in this case it should work just fine.
>>
>>
>> Kind regards,
>> Michał Michalski,
>> michal.michal...@boxever.com
>>
>> On 24 April 2015 at 14:05, Spico Florin <spicoflo...@gmail.com> wrote:
>>
>>> Hello!
>>> I know that HadoopRDD partitions are built based on the number of
>>> splits in HDFS. I'm wondering if these partitions preserve the initial
>>> order of data in file.
>>> As an example, if I have an HDFS (myTextFile) file that has these splits:
>>>
>>> split 0-> line 1, ..., line k
>>> split 1->line k+1,..., line k+n
>>> splt 2->line k+n, line k+n+m
>>>
>>> and the code
>>> val lines=sc.textFile("hdfs://mytextFile")
>>> lines.zipWithIndex()
>>>
>>> will the order of lines preserved?
>>> (line 1, zipIndex 1) , .. (line k, zipIndex k), and so one.
>>>
>>> I found this question on stackoverflow (
>>> http://stackoverflow.com/questions/26046410/how-can-i-obtain-an-element-position-in-sparks-rdd)
>>> whose answer intrigued me:
>>> "Essentially, RDD's zipWithIndex() method seems to do this, but it won't
>>> preserve the original ordering of the data the RDD was created from"
>>>
>>> Can you please confirm that is this the correct answer?
>>>
>>> Thanks.
>>> Florin
>>>
>>>
>>>
>>>
>>>
>>>
>>
>
