Category: Hadoop

An Apache Sqoop Tutorial

Published 2017-11-22 by Kevin Feasel

Subham Sinha has an introductory-level tutorial on Apache Sqoop:

For Hadoop developer, the actual game starts after the data is being loaded in HDFS. They play around this data in order to gain various insights hidden in the data stored in HDFS.

So, for this analysis the data residing in the relational database management systems need to be transferred to HDFS. The task of writing MapReduce code for importing and exporting data from relational database to HDFS is uninteresting & tedious. This is where Apache Sqoop comes to rescue and removes their pain. It automates the process of importing & exporting the data.

Sqoop makes the life of developers easy by providing CLI for importing and exporting data. They just have to provide basic information like database authentication, source, destination, operations etc. It takes care of remaining part.

Sqoop internally converts the command into MapReduce tasks, which are then executed over HDFS. It uses YARN framework to import and export the data, which provides fault tolerance on top of parallelism.

In my experience, Sqoop does two things really well: first, it lets you move data from a relational database into HDFS (or Hive). Second, it lets you move data from HDFS (or Hive) into a staging table on a relational database. That can make Sqoop a useful part of an ETL process.

Comments closed

Getting Started With Zeppelin

Published 2017-11-16 by Kevin Feasel

Sangeeta Gulia shows us how to get started building notebooks with Apache Zeppelin on top of Spark:

There are 3 interpreter modes available in Zeppelin.

1) Shared Mode

In Shared mode, a SparkContext and a Scala REPL is being shared among all interpreters in the group. So every Note will be sharing single SparkContext and single Scala REPL. In this mode, if NoteA defines variable ‘a’ then NoteB not only able to read variable ‘a’ but also able to override the variable.

2) Scoped Mode

In Scoped mode, each Note has its own Scala REPL. So variable defined in a Note can not be read or overridden in another Note. However, still single SparkContext serves all the Interpreter Groups. And all the jobs are submitted to this SparkContext and fair scheduler schedules the job. This could be useful when user does not want to share Scala session, but want to keep single Spark application and leverage its fair scheduler.

3) Isolated Mode

In Isolated mode, each Note has its own SparkContext and Scala REPL.

The default mode of %spark interpreter is ‘Globally Shared’.

This is mostly a step-by-step on installing Zeppelin, but does go into some detail on how Zeppelin works.

Comments closed

How Per-Second AWS Billing Helps With Data Processing

Published 2017-11-07 by Kevin Feasel

Prakash Chockalingam explains how AWS per-second billing can make resource allocation easier:

Because of the hourly increments in billing, users spend a lot of time playing a giant game of Tetris with their big data workloads — figuring out how to pack jobs to use every minute of the compute hour. Examples:

If a job could be run on 10 nodes and finished in 20 minutes, it was better to run it on fewer nodes so that it takes around 50 minutes. As a result, you would pay less.

Running two 10 node jobs that took 20 minutes in parallel would cost two times more than running them sequentially.

The above problem got compounded if there were many such jobs to be run. To handle this challenge, many organizations turned to a resource scheduler like YARN. Organizations were following the traditional on-premises model of setting up one or more big multi-tenant cluster on the cloud and running YARN to bin-pack the different jobs.

Read on to see how this has changed as a result of per-second billing.

Comments closed

Tracking Kafka Consumer Lag

Published 2017-11-07 by Kevin Feasel

Simarpreet Kaur Monga has a Scala-based example showing how to calculate Kafka offset lag for consumers:

The Consumer can subscribe to multiple topics, you need to pass the list of topics you want to consume from. For the sake of simplicity, I have just passed a single topic to consume from.

Now that the consumer has subscribed to the topic, it can consume from that topic.

The consumer maintains an offset to keep the track of the next record it needs to read.

Now, let us see how we can find the consumer offsets.

The Consumer offsets can be found using the method offset of class ConsumerRecord. This offset points to the record in a Kafka partition. The consumer consumes the records from the topic in the form of an object of class ConsumerRecord. The class ConsumerRecord also consists of a topic name and a partition number from which the record is being received, and a timestamp as marked by the corresponding ProducerRecord (the record sent by the producer).

Click through for the rest of the story.

Comments closed

Kafka 1.0 Released

Published 2017-11-03 by Kevin Feasel

Neha Narkhede announces Apache Kafka 1.0:

And Kafka 1.0.0 is no mere bump of the version number. The Apache Kafka Project Management Committee and the broader Kafka community has packed a number of valuable enhancements into the release. Let me summarize a few of them:

Since its introduction in version 0.10, the Streams API has become hugely popular among Kafka users, including the likes of Pinterest, Rabobank, Zalando, and The New York Times. In 1.0, the API continues to evolve at a healthy pace. To begin with, the builder API has been improved (KIP-120). A new API has been added to expose the state of active tasks at runtime (KIP-130). The new cogroup API makes it much easier to deal with partitioned aggregates with fewer StateStores and fewer moving parts in your code (KIP-150). Debuggability gets easier with enhancements to the print() and writeAsText() methods (KIP-160). And if that’s not enough, check out KIP-138 and KIP-161 too. For more on streams, check out the Apache Kafka Streams documentation, including some helpful new tutorial videos.
Operating Kafka at scale requires that the system remain observable, and to make that easier, we’ve made a number of improvements to metrics. These are too many to summarize without becoming tedious, but Connect metrics have been significantly improved (KIP-196), a litany of new health check metrics are now exposed (KIP-188), and we now have a global topic and partition count (KIP-168). (That last one sounds so simple, but you’ve wanted it in the past, haven’t you?) Check out KIP-164 and KIP-187 for even more.
We now support Java 9, leading, significantly faster TLS and CRC32C implementations. Over-the-wire encryption will be faster now, which will keep Kafka fast and compute costs low when encryption is enabled.

And there are more where that came from. Congratulations to the Kafka team for hitting this big milestone.

Comments closed

Hortonworks Data Platform 2.6.3 Released

Published 2017-11-02 by Kevin Feasel

Roni Fontaine, et al, announce Hortonworks Data Platform 2.6.3:

First, in this release, we made huge strides with regards to Data Science. This includes:

Apache Spark 2.2 is GA. This enhances the stability of Spark for structured streaming.
Apache Zeppelin 0.7.3 is GA. This provides support for notebooks stored in HDFS, and improves the scalability and multi-user support.
Apache Livy 0.4.0 is GA. This version supports multiple programming languages in the same Spark context.

This is a point release, so most of it is around adding updated versions of different products.

Comments closed

Vectorized UDFs For PySpark

Published 2017-11-01 by Kevin Feasel

Li Jin talks about a performance optimization coming in Apache Spark 2.3:

To enable data scientists to leverage the value of big data, Spark added a Python API in version 0.7, with support for user-defined functions. These user-defined functions operate one-row-at-a-time, and thus suffer from high serialization and invocation overhead. As a result, many data pipelines define UDFs in Java and Scala, and then invoke them from Python.

Vectorized UDFs built on top of Apache Arrow bring you the best of both worlds—the ability to define low-overhead, high performance UDFs entirely in Python.

This looks like a good performance improvement coming to PySpark, bringing it closer to Scala/Java performance with respect to UDFs.

Comments closed

Hive Without Hadoop

Published 2017-11-01 by Kevin Feasel

Anubhav Tarar notes that you can run Hive queries against data even without a Hadoop installation:

Starting with release 0.7, Hive also supports a mode to run map-reduce jobs in local-mode automatically

you just have to do two things first create your warehouse in local system and give the default fs name to local put these properties inside your hive-site.xml

This is a fairly short post; click through to see the changes you’d make to hive-site.xml.

Comments closed

Avro And Streaming Data

Published 2017-10-30 by Kevin Feasel

Pat Patterson shows how to get the advantages of the Avro file format while streaming individual records:

Avro is a very efficient way of storing data in files, since the schema is written just once, at the beginning of the file, followed by any number of records (contrast this with JSON or XML, where each data element is tagged with metadata). Similarly, Avro is well suited to connection-oriented protocols, where participants can exchange schema data at the start of a session and exchange serialized records from that point on. Avro works less well in a message-oriented scenario since producers and consumers are loosely coupled and may read or write any number of records at a time. To ensure that the consumer has the correct schema, it must either be exchanged “out of band” or accompany every message. Unfortunately, sending the schema with every message imposes significant overhead — in many cases, the schema is as big as the data or even bigger!

Read on to see how the Confluent Schema Registry can solve this problem.

Comments closed

Tips For Running Kafka Streams On AWS

Published 2017-10-26 by Kevin Feasel

Ian Duffy and Nina Hanzlikova have some advice if you’re looking to spin up some EC2 instances to run Kafka Streams:

With upgrades in the underlying Kafka Streams library, the Kafka community introduced many improvements to the underlying stream configuration defaults. Where in previous, more unstable iterations of the client library we spent a lot of time tweaking config values such as session.timeout.ms, max.poll.interval.ms, and request.timeout.ms to achieve some level of stability.

With new releases we found ourselves discarding these custom values and achieving better results. However, some timeout issues persisted on some of our services, where a service would frequently get stuck in a rebalancing state. We noticed that reducing the max.poll.records value for the stream configs would sometimes alleviate issues experienced by these services. From partition lag profiles we also saw that the consuming issue seemed to be confined to only a few partitions, while the others would continue processing normally between re-balances. Ultimately we realised that the processing time for a record in these services could be very long (up to minutes) in some edge cases. Kafka has a fairly large maximum offset commit time before a stream consumer is considered dead (5 minutes) but with larger message batches of data this timeout was still being exceeded. By the time the processing of the record was finished the stream was already marked as failed and so the offset could not be committed. On rebalance, this same record would once again be fetched from Kafka, would fail to process in a timely manner and the situation would repeat. Therefore for any of the affected applications we introduced a processing timeout, ensuring there was an upper bound on the time taken by any of our edge cases.

There are some interesting tidbits in here.

Comments closed