Ah, interesting. While working on my new Tungsten shuffle manager, I came up with some nice testing interfaces for allowing me to manually trigger spills in order to deterministically test those code paths without requiring large amounts of data to be shuffled. Maybe I could make similar test interface changes to the existing shuffle code, which might make it easier to reproduce this in an isolated environment.
On Mon, Jun 1, 2015 at 11:41 PM, Igor Berman <igor.ber...@gmail.com> wrote: > Hi, > small mock data doesn't reproduce the problem. IMHO problem is reproduced > when we make shuffle big enough to split data into disk. > We will work on it to understand and reproduce the problem(not first > priority though...) > > > On 1 June 2015 at 23:02, Josh Rosen <rosenvi...@gmail.com> wrote: > >> How much work is to produce a small standalone reproduction? Can you >> create an Avro file with some mock data, maybe 10 or so records, then >> reproduce this locally? >> >> On Mon, Jun 1, 2015 at 12:31 PM, Igor Berman <igor.ber...@gmail.com> >> wrote: >> >>> switching to use simple pojos instead of using avro for spark >>> serialization solved the problem(I mean reading avro from s3 and than >>> mapping each avro object to it's pojo serializable counterpart with same >>> fields, pojo is registered withing kryo) >>> Any thought where to look for a problem/misconfiguration? >>> >>> On 31 May 2015 at 22:48, Igor Berman <igor.ber...@gmail.com> wrote: >>> >>>> Hi >>>> We are using spark 1.3.1 >>>> Avro-chill (tomorrow will check if its important) we register avro >>>> classes from java >>>> Avro 1.7.6 >>>> On May 31, 2015 22:37, "Josh Rosen" <rosenvi...@gmail.com> wrote: >>>> >>>>> Which Spark version are you using? I'd like to understand whether >>>>> this change could be caused by recent Kryo serializer re-use changes in >>>>> master / Spark 1.4. >>>>> >>>>> On Sun, May 31, 2015 at 11:31 AM, igor.berman <igor.ber...@gmail.com> >>>>> wrote: >>>>> >>>>>> after investigation the problem is somehow connected to avro >>>>>> serialization >>>>>> with kryo + chill-avro(mapping avro object to simple scala case class >>>>>> and >>>>>> running reduce on these case class objects solves the problem) >>>>>> >>>>>> >>>>>> >>>>>> >>>>>> -- >>>>>> View this message in context: >>>>>> http://apache-spark-user-list.1001560.n3.nabble.com/union-and-reduceByKey-wrong-shuffle-tp23092p23093.html >>>>>> Sent from the Apache Spark User List mailing list archive at >>>>>> Nabble.com. >>>>>> >>>>>> --------------------------------------------------------------------- >>>>>> To unsubscribe, e-mail: user-unsubscr...@spark.apache.org >>>>>> For additional commands, e-mail: user-h...@spark.apache.org >>>>>> >>>>>> >>>>> >>> >> >
