spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Aaron Olson <>
Subject Pyspark Memory Woes
Date Tue, 11 Mar 2014 17:11:18 GMT
Dear Sparkians,

We are working on a system to do relational modeling on top of Spark, all
done in pyspark. While we've been learning a lot about Spark internals so
far, we're currently running into memory issues and wondering how best to
profile to fix them. Here are our symptoms:

   - We're operating on data sets up to 80G in size of uncompressed JSON,
   66 million records in the largest one.
   - Sometimes we're joining those large data sets, but cardinality never
   exceeds 66 million (unless we've got a bug somewhere).
   - We're seeing various OOM problems: sometimes python takes all
   available mem, sometimes we OOM with no heap space left, and occasionally
   OOM with GC overhead limit exceeded.
   - Sometimes we also see what looks like a single huge message sent over
   the wire that exceeds the wire format limitations.
   - Java heap space is 1.5G per worker, 24 or 32 cores across 46 nodes. It
   seems like we should have more than enough to do this comfortably.
   - We're trying to isolate specific steps now, but every time it errors,
   we're partitioning (i.e. partitionByKey is involved somewhere).

We've been instrumneting according to the monitoring and tuning docs, but a
bit at a loss for where we're going wrong. We suspect poor/wrong
partitioning on our part somehow. With that in mind, some questions:

   - How exactly is partitioning information propagated? It looks like
   within a pipelined RDD the parent partitioning is preserved throughout
   unless we either specifically repartition or go through a reduce. We're
   splitting as much as we can on maps and letting reduces happen normally. Is
   that good practice?
   - When doing e.g. partitionByKey, does an entire partition get sent to
   one worker process?
   - When does Spark stream data? Are there easy ways to sabotage the
   streaming? Are there any knobs for us to twiddle here?
   - Is there any way to specify the number of shuffles for a given reduce
   - How can we get better insight into what our workers are doing,
   specifically around moving data in and out of python land?

I realise it's hard to troubleshoot in the absence of code but any test
case we have would be contrived. We're collecting more metrics and trying
to reason about what might be happening, but any guidance at this point
would be most helpful.


Aaron Olson
Data Engineer, Shopify

View raw message