Measuring our elephants’ appetite

April 17, 2009

We recently wanted to have a rough idea of how much data our Hadoop clusters process on a daily basis. Here’s the Dumbo program I used to obtain this information:

from datetime import date

class Mapper:
    def __init__(self):
        from re import compile
        self.numbers = []
        self.numbers.append(compile("HDFS bytes read:([0-9]+)"))
        self.numbers.append(compile("Local bytes read:([0-9]+)"))
        self.finish = compile('FINISH_TIME="([^"]+)"')
    def __call__(self, key, value):
        if value.startswith("Job") and "COUNTERS" in value:
            gb = 0  # gigabytes processed
            for number in self.numbers:
                mo =
                if mo: gb += int(round(float( / 2**30))
            ts = float( / 1000
            datetuple = date.fromtimestamp(ts).timetuple()[:3]
            yield datetuple, gb

if __name__ == "__main__":
    from dumbo import run, sumreducer
    run(Mapper, sumreducer, combiner=sumreducer)

Running this on the job logs for one of our clusters (which are gathered by the shell script discussed in this previous post) led to the following graph:

Bytes processed daily by one of our Hadoop clusters

This graph clearly shows why some of us get annoyed sometimes when they want to explore data on this cluster on certain days of the week or month…