Getting Started

HDFS Connector


The HDFS connector allows your Samza jobs to read data stored in HDFS files. Likewise, you can write processed results to HDFS. To interact with HDFS, Samza requires your job to run on the same YARN cluster.

Consuming from HDFS

Input Partitioning

Partitioning works at the level of individual directories and files. Each directory is treated as its own stream and each of its files is treated as a partition. For example, Samza creates 5 partitions when it’s reading from a directory containing 5 files. There is no way to parallelize the consumption when reading from a single file - you can only have one container to process the file.

Input Event format

Samza supports avro natively, and it’s easy to extend to other serialization formats. Each avro record read from HDFS is wrapped into a message-envelope. The envelope contains these 3 fields:

  • The key, which is empty

  • The value, which is set to the avro GenericRecord

  • The partition, which is set to the name of the HDFS file

To support non-avro input formats, you can implement the SingleFileHdfsReader interface.


While streaming sources like Kafka are unbounded, files on HDFS have finite data and have a notion of EOF. When reading from HDFS, your Samza job automatically exits after consuming all the data. You can implement EndOfStreamListenerTask to get a callback once EOF has been reached.

Defining streams

Samza uses the notion of a system to describe any I/O source it interacts with. To consume from HDFS, you should create a new system that points to - HdfsSystemFactory. You can then associate multiple streams with this system. Each stream should have a physical name, which should be set to the name of the directory on HDFS.



The above example defines a stream called hdfs-clickstream that reads data from the /data/clickstream/2016/09/11 directory.

Whitelists & Blacklists

If you only want to consume from files that match a certain pattern, you can configure a whitelist. Likewise, you can also blacklist consuming from certain files. When both are specified, the whitelist selects the files to be filtered and the blacklist is later applied on its results.


Producing to HDFS

Output format

Samza allows writing your output results to HDFS in AVRO format. You can either use avro’s GenericRecords or have Samza automatically infer the schema for your object using reflection.

# set the SystemFactory implementation to instantiate HdfsSystemProducer aliased to 'hdfs'

If your output is non-avro, you can describe its format by implementing your own serializer.


Output directory structure

Samza allows you to control the base HDFS directory to write your output. You can also organize the output into sub-directories depending on the time your application ran, by configuring a date-formatter.


You can configure the maximum size of each file or the maximum number of records per-file. Once either limits have been reached, Samza will create a new file.



You can access Kerberos-enabled HDFS clusters by providing your principal and the path to your key-tab file. Samza takes care of automatically creating and renewing your Kerberos tokens periodically.

# Kerberos principal

# Path of the keytab file (local path)