Optimizing aggregate summation of a sparse vector in Spark (and saving to a parquet)

Question

Please excuse the Pyspark NOOB question.

My final stage in producing a Spark dataframe in PySpark is the following:

indexer = StringIndexer(inputCol="kpID", outputCol="KPindex")
inputs = [indexer.getOutputCol()]
encoder = OneHotEncoderEstimator(inputCols=inputs, outputCols=["KPvec"])
pipeline = Pipeline(stages=[indexer, encoder])
df_bayes = pipeline.fit(df_bayes).transform(df_bayes)

def sparse_to_array(v):
    v = DenseVector(v)
    new_array = list([float(x) for x in v])
    return new_array

sparse_to_array_udf = F.udf(sparse_to_array, ArrayType(FloatType()))
df_bayes = df_bayes.select('id',sparse_to_array_udf(col('KPvec')).alias('KPvec'))
df_bayes = df_bayes.repartition(15000,col('id'))
df_bayes = df_bayes.select('id','KPvec').groupby('id').agg(F.array(*[F.sum(F.col('KPvec')[i]) for i in range(len(kids))]).alias("KPvec")).cache()

I am trying to aggregate-sum a sparse vector which represents a one hot encoded categorical variable.

On my EMR cluster, this takes 188s to complete. And the resulting dataframe has ~50M rows. I then try to write this dataframe to a parquet.

I have tried:

df_bayes.write.format("parquet") \
.partitionBy("id") \
.bucketBy(500,"KPvec") \
.option("path", "s3://..."+"output.parquet") \
.saveAsTable("output")

And:

df_bayes.write.repartition(1500,col('id')).parquet("s3://..."+"output.parquet")

And without re-partitioning.

In each case, the job takes a very long time and eventually fails with an ExecutorLostFailure (which is due to the EMR running with many spot instances).

In spite of caching before, I suspect that many of these steps actually don't relate to the parquet writing but rather to the computation steps I have requested.

I suspect that this is the case because if I try to calculate the dimensions of the dataframe, I see that the DAG Visualization is:

The repeating steps and the ~6GB of shuffle writes before the job fails, indicate to me that there is a big inefficiency in the way I perform the calculation.

Also, when I run explain I get the following:

== Physical Plan ==
InMemoryTableScan [id#1, KPvec#52167]
   +- InMemoryRelation [id#1, KPvec#52167], StorageLevel(disk, memory, deserialized, 1 replicas)
         +- HashAggregate(keys=[id#1], functions=[sum(cast(KPvec#27[0] as double)), sum(cast(KPvec#27[1] as double)), sum(cast(KPvec#27[2] as double)), sum(cast(KPvec#27[3] as double)), sum(cast(KPvec#27[4] as double)), sum(cast(KPvec#27[5] as double)), sum(cast(KPvec#27[6] as double)), sum(cast(KPvec#27[7] as double)), sum(cast(KPvec#27[8] as double)), sum(cast(KPvec#27[9] as double)), sum(cast(KPvec#27[10] as double)), sum(cast(KPvec#27[11] as double)), sum(cast(KPvec#27[12] as double)), sum(cast(KPvec#27[13] as double)), sum(cast(KPvec#27[14] as double)), sum(cast(KPvec#27[15] as double)), sum(cast(KPvec#27[16] as double)), sum(cast(KPvec#27[17] as double)), sum(cast(KPvec#27[18] as double)), sum(cast(KPvec#27[19] as double)), sum(cast(KPvec#27[20] as double)), sum(cast(KPvec#27[21] as double)), sum(cast(KPvec#27[22] as double)), sum(cast(KPvec#27[23] as double)), ... 52109 more fields])
            +- HashAggregate(keys=[id#1], functions=[partial_sum(cast(KPvec#27[0] as double)), partial_sum(cast(KPvec#27[1] as double)), partial_sum(cast(KPvec#27[2] as double)), partial_sum(cast(KPvec#27[3] as double)), partial_sum(cast(KPvec#27[4] as double)), partial_sum(cast(KPvec#27[5] as double)), partial_sum(cast(KPvec#27[6] as double)), partial_sum(cast(KPvec#27[7] as double)), partial_sum(cast(KPvec#27[8] as double)), partial_sum(cast(KPvec#27[9] as double)), partial_sum(cast(KPvec#27[10] as double)), partial_sum(cast(KPvec#27[11] as double)), partial_sum(cast(KPvec#27[12] as double)), partial_sum(cast(KPvec#27[13] as double)), partial_sum(cast(KPvec#27[14] as double)), partial_sum(cast(KPvec#27[15] as double)), partial_sum(cast(KPvec#27[16] as double)), partial_sum(cast(KPvec#27[17] as double)), partial_sum(cast(KPvec#27[18] as double)), partial_sum(cast(KPvec#27[19] as double)), partial_sum(cast(KPvec#27[20] as double)), partial_sum(cast(KPvec#27[21] as double)), partial_sum(cast(KPvec#27[22] as double)), partial_sum(cast(KPvec#27[23] as double)), ... 52109 more fields])
               +- Exchange hashpartitioning(id#1, 15000)
                  +- *(2) Project [id#1, pythonUDF0#52170 AS KPvec#27]
                     +- BatchEvalPython [sparse_to_array(KPvec#3)], [KPvec#3, id#1, pythonUDF0#52170]
                        +- *(1) Project [KPvec#3, id#1]
                           +- *(1) FileScan parquet [id#1,KPvec#3] Batched: false, Format: Parquet, Location: InMemoryFileIndex[s3://...bayesnetw..., PartitionFilters: [], PushedFilters: [], ReadSchema: struct<id:string,KPvec:struct<type:tinyint,size:int,indices:array<int>,values:array<double>>>

Can anyone point me in the direction of what I am doing wrong here?

Thank you in advance.

user1860389 user1860389 · Accepted Answer · 2020-03-06T08:10:26

So to answer my own question in case it helps someone else, the solution is to collect_set the features you want to one-hot encode and then use CountVectorizor instead of Spark ML's OneHotEncoder.

df.select('id','feature').groupby('id').agg(F.collect_set('feature').alias('feature'))

countModel = CountVectorizer().setInputCol("feature").setOutputCol("feature_vec").fit(df)
df = countModel.transform(df).select('id','KPvec')

Then you can simply save it to a parquet. For me, this is quite fast.

Optimizing aggregate summation of a sparse vector in Spark (and saving to a parquet)

1 Answers