Month: July 2016

Michael Noll digs into Kafka Streams, showing how to enrich data and collect aggregates:

The stream of user click events is considered to be a record stream, where each data record represents a self-contained datum.  In contrast, the stream of user geo-location updates is interpreted as a changelog stream, where each data record represents an update (i.e. any previous data records having the same record key will be replaced by the latest update).  In Kafka Streams, a record stream is represented via the so-called KStream interface and a changelog stream via the KTable interface.  Going from the high-level view to the technical view, this means that our streaming application will demonstrate how to perform a join operation between a KStream and a KTable, i.e. it is an example of a stateful computation.  This KStream-KTable join also happens to be Kafka Streams’ equivalent of performing a table lookup in a streaming context, where the table is updated continuously and concurrently.  Specifically, for each user click event in the KStream, we will lookup the user’s region (e.g. “europe”) in the KTable in order to subsequently compute the total number of user clicks per region.

Let’s showcase the beginning (input) and the end (expected output) of this data pipeline with some example data.

This article is fairly detailed, but it covers a rather interesting topic in a good way.

Sai Sriparasa shows how to use Sqoop to transfer data from a Hadoop cluster into a relational database:

Here are a few best practices for exporting with Sqoop:

• Options file—As commands with Sqoop export and Sqoop import tend to be bigger in size, I recommend storing the commands in an options file. By keeping it in an options file, you can even make it part of a version control pipeline to monitor changes to the command.

• Field termination—With Sqoop export, I recommend providing field termination metadata using the “–fields-terminated-by” option. Also, other formatting options such as “lines-terminated-by”, “enclosed-by”, “escaped-by”, etc., can be used as required.

• Mapper tuning—When an export job is submitted, Sqoop creates a Java class and submits a MapReduce job based on input splits; then, each mapper connects to the database to export the data. The default number of mappers is 4, so I recommend tuning the number of mappers depending on the availability of processors on the cluster. Too many mappers might cause the load to increase on the database. We recommend that you monitor the number of connections and keep track of processlist on MySQL.

• Staging table—The Sqoop export job is broken down into multiple transactions based on the mappers. Each transaction is therefore atomic and does not have any dependencies on other transactions. I recommend using the “–staging-table” option that acts as the buffer table for the separate transactions. After all transactions have been committed, a single transaction move is made to move the data to the final destination. Use the “–clear-staging-table” option to clean up the staging table after the export job.

There’s a lot in here which is Amazon-specific and there are a couple of things you’d have to change to deploy to SQL Server, but there’s a lot of useful information here.  I like that Sai shows how to use the Hadoop credential API instead of doing something silly like saving your password in plaintext.