Implementing Real-Time Trending Topics in Storm

http://www.michael-noll.com/blog/2013/01/18/implementing-real-time-trending-topics-in-storm/

by skadamat • 13 years ago

73 6 13 years ago

samspenc• 13 years ago

Michael Noll has some really good introductory articles to Big Data, and I highly recommend his posts for anyone starting in this field.

His post on writing Python streaming programs in Hadoop - http://www.michael-noll.com/tutorials/writing-an-hadoop-mapr... - got me started in this area two years ago, and that's what I'm still doing these days! :)

dxbydt• 13 years ago

I read this one a few months back & felt, like most programmers, that it was an awesome article, but I would forget all of it unless I rewrote everything from scratch all by myself :) So here's my rewrite in Scala - its not realtime, but still takes only 4 minutes on my laptop for a typical weekly dataset, so you can put it on a cron schedule & get a trending topics with a 5 min delay. Ofcourse, no storm, no hadoop etc. just a plain old scala program anybody can run on their laptop. http://bit.ly/J82szS

pinhead• 13 years ago

Maybe check out Spark [1], it primarily supports Scala and they have great streaming functionality that seems to beat out Storm and others in terms of performance for some workloads [2].

[1] http://spark.incubator.apache.org/

[2] http://www.cs.berkeley.edu/~matei/papers/2013/sosp_spark_str...

adambard• 13 years ago

Storm (and Trident) are just fantastic. If you're looking for a streaming complement (or replacement!) for Hadoop, you can't go wrong here.

capkutay• 13 years ago

Sorry for the shameless trolling, but if you are a fan of storm and see potential in the real-time side of big data, consider applying to work on the Platform team at WebAction.

We're a well-funded company (raised an $11m Series B from a $15billion private equity firm) real-time big data analytics company for the enterprise. We're based in Palo Alto. If interested, feel free to shoot us a message!

jobs@webaction.com

scott_s• 13 years ago

I work on IBM InfoSphere Streams (https://www.ibmdw.net/streamsdev/), and the language our system uses is SPL (Streams Processing Language). I was curious to see what this problem would look like implemented in SPL, so came up this this:

  composite Main {
  graph
    @parallel(width=2)
    stream<rstring word> Words = Custom() {
      logic onProcess: {
        while (true) {
          // a "word" looks like "0", "1", ..., "99"
          submit({word=(rstring)((int32)(random() * 100))}, Words);
        }
      }
    }

    @parallel(width=3, partitionBy=[{port=Words, attributes=[word]}])
    stream<rstring word, int32 rolling_count> Counts = Aggregate(Words) {
      window Words: sliding, time(9), time(3), partitioned;
      param partitionBy: word;
      output Counts: word=Any(), rolling_count=Count();
    }

    stream<rstring word, int32 rolling_count> Sorted = Sort(Words) {
      window Words: sliding, time(2), time(2);
      param sortBy: rolling_count;
    }

    stream<list<rstring word, int32 rolling_count> hist> Rankings = Aggregate(Sorted) {
      window Sorted: tumbling, punctuation;
      output Rankings: hist=Collect();
    }
  }

I did not parallelize the reduce part. I'd need to do some more thinking on how to do that, as I split it into two operators. I think it shows the flavor, though.

Two caveats: One, I have not compiled and tested this, so errors may abound. I was just curious what the "flavor" of it would look like in our system. Two, standard disclaimer of I speak in an unofficial capacity, what I say does not represent IBM or Streams.

Implementing Real-Time Trending Topics in Storm

Related Stories