Simple scalding example

1 minute read

Scalding makes hadoop job writing much smaller and compact. It comes at a price, the price is that you are no longer in easy to use java mature world but enter the wilderness of scala (for good and bad).

In scalding you must start thinking in pipes (best known in unix world as shell pipes), take input do operations produce output and pass on to the next pipe. A pipe is basically a stream, the basic workflow, is that you create a stream of inputs (be it a book lines) then you operate on it, create an output pipe then have another operation work on that pipe and another one, until at the end of this process you write your results to HDFS.

Note, to see the full sample source code have a look at scalding-counters-example

Before we dig in the source the flow you are going to see is similar to the following pseudo code:

MyJob extends Job Read input into a pipe do work on pipe return new pipe do work on pipe return new pipe do work on pipe return new pipe write results to hdfs

And the actual scala code:

class ScaldingCounterExampleJob(args : Args) extends Job(args) { val stat = Stat(“alice-counter”) TypedPipe.from(“is alice really alice ?”.split(“ “)) .flatMap { line => { stat.inc line.split(“”“\s+”””) } } .groupBy { word => word } .size .write(TypedTsv(args(“output”)))

}

  1. We create an input pipe from a bunch of words by using TypedPipe.from("bunch of words")
  2. Now that we have a pipe we can start doing operations on it that’s the flatMap, groupBy
  3. At the end we write the output to HDFS with .write

Caveats with this naive implementation (which people tend to stick with) is that its difficult to unit test. Try to test the internal line.split or flatMap, practically impossible. For this we will use the scalding externalized operation design pattern (I really don’t understand why after all those years of good programming practices we had to go backwards into untestable code only to invent new design patterns to solve this naive problem of unit testing our code, there should be no design pattern for that, this is how we should write our code originally, in anyway we will presnet this pattern in future posts.

Categories: , , , , , ,

Updated:

Leave a Comment