I've thrown together a Pig LoadFunc to read data from Kafka, so you
could load data like:
QUERY_LOGS = load 'kafka://localhost:9092/logs.query#8' using
The path part of the uri is the Kafka topic, and the fragment is the
number of partitions. In the implementation I have, it makes one input
split per partition. Offsets are not really dealt with at this point -
it's a rough prototype.
Anyone have thoughts on whether or not this is a good idea? I know
usually the pattern is: kafka -> hdfs -> mapreduce. If I'm only reading
from this data from Kafka once, is there any reason why I can't skip
writing to HDFS?