CODEWITHSUNDEEP
javaapache-sparkrdd
lochi
lochi

Reputation: 880

Spark data processing with grouping

I need to group a set of csv lines by a certain column and do some processing on each group. 

    JavaRDD<String> lines = sc.textFile
                        ("somefile.csv");
                JavaPairRDD<String, String> pairRDD = lines.mapToPair(new SomeParser());
                List<String> keys = pairRDD.keys().distinct().collect();
                for (String key : keys)
                {
                List<String> rows = pairRDD.lookup(key);

            noOfVisits = rows.size();
            country = COMMA.split(rows.get(0))[6];
            accessDuration = getAccessDuration(rows,timeFormat);
            Map<String,Integer> counts = getCounts(rows);
            whitepapers = counts.get("whitepapers");
            tutorials = counts.get("tutorials");
            workshops = counts.get("workshops");
            casestudies = counts.get("casestudies");
            productPages = counts.get("productpages");        
            }

    private static long dateParser(String dateString) throws ParseException {
        SimpleDateFormat format = new SimpleDateFormat("MMM dd yyyy HH:mma");
        Date date = format.parse(dateString);
        return date.getTime();
    }
dateParser is called for each row. Then min and max for the group is calculated to get the access duration. Others are string matches.

pairRDD.lookup is extremely slow.. Is there a better way to do this with spark.

Upvotes: 1

Views: 3095

Answers (1)

maasg
maasg

Reputation: 37435

I think you could simply use that column as key and do a groupByKey. There's no mention on the operation on those rows. If it's a function that combines those rows somehow, you could even use reduceByKey.

Something like:

import org.apache.spark.SparkContext._  // implicit pair functions
val pairs = lines.map(parser _)
val grouped = pairs.groupByKey
// here grouped is of the form: (key, Iterator[String])

* EDIT * After looking at the process, I think it would be more efficient to map each row into the data it contributes and then use aggregateByKey to reduce them all to a total. aggregateByKey takes 2 functions and a zero: 

def aggregateByKey[U: ClassTag](zeroValue: U)(seqOp: (U, V) => U,
      combOp: (U, U) => U): RDD[(K, U)]

The first function is a partition aggregator and will efficiently run through the local partitions, creating local aggregated partials per partition. The combineOperation will take those partial aggregations and combine them together to obtain a final result.

Something like this:

val lines = sc.textFile("somefile.csv")
// parse returns a key and a decomposed Record of values tracked:(key, Record("country", timestamp,"whitepaper",...)) 

val records = lines.map(parse(_))

val totals = records.aggregateByKey((0,Set[String].empty,Long.MaxValue, Long.MinValue, Map[String,Int].empty),
(record, (count, countrySet, minTime, maxTime, counterMap )) => (count+1,countrySet + record.country, math.min(minTime,record.timestamp), math.max(maxTime, record.timestamp), ...)
(cumm1, cumm2) => ???  // add each field of the cummulator
)

This is the most efficient method in Spark to do key-based aggregations.

Upvotes: 3

Related Questions