Kafka Offset Management With Spark Streaming

Guru Medasana and Jordan Hambleton explain how to perform Kafka offset management when using Spark Streaming:

Enabling Spark Streaming’s checkpoint is the simplest method for storing offsets, as it is readily available within Spark’s framework. Streaming checkpoints are purposely designed to save the state of the application, in our case to HDFS, so that it can be recovered upon failure.

Checkpointing the Kafka Stream will cause the offset ranges to be stored in the checkpoint. If there is a failure, the Spark Streaming application can begin reading the messages from the checkpoint offset ranges. However, Spark Streaming checkpoints are not recoverable across applications or Spark upgrades and hence not very reliable, especially if you are using this mechanism for a critical production application. We do not recommend managing offsets via Spark checkpoints.

The authors give several options, so check it out and pick the one that works best for you.

Updates In Apache Kafka

Yeva Byzek announces that Apache Kafka 0.11.0.0 is shipping soon:

We are very excited for the GA for Kafka release 0.11.0.0 which is just days away. This release is bringing many new features as described in the previous Log Compaction blog post.

The most notable new feature is Exactly Once Semantics (EOS).  Kafka’s EOS capabilities provide more stringent idempotent producer semantics with exactly once, in-order delivery per partition, and stronger transactional guarantees with atomic writes across multiple partitions. Together, these strong semantics make writing applications easier and expand Kafka’s addressable use cases. You can learn more about EOS in the online talk on June 29, 2017.

“Exactly once,” if done right, would be crazy—there’s a reason most brokers are either “at least once” or “best effort.”

Re-Shaping Data Flows

Maneesh Varshney explains some methods to trim the fat out of analytical data flows:

Big data comes in a variety of shapes. The Extract-Transform-Load (ETL) workflows are more or less stripe-shaped (left panel in the figure above) and produce an output of a similar size to the input. Reporting workflows are funnel-shaped (middle panel in the figure above) and progressively reduce the data size by filtering and aggregating.

However, a wide class of problems in analytics, relevance, and graph processing have a rather curious shape of widening in the middle before slimming down (right panel in the figure above). It gets worse before it gets better.

In this article, we take a deeper dive into this exploding middle shape: understanding why it happens, why it’s a problem, and what can we do about it. We share our experiences of real-life workflows from a spectrum of fields, including Analytics (A/B experimentation), Relevance (user-item feature scoring), and Graph (second degree network/friends-of-friends).

The examples relate directly to Hadoop, but are applicable in other data platforms as well.

Spark Streaming Vs Kafka Streams

Mahesh Chand Kandpal contrasts Kafka Streams with Spark Streaming:

The low latency and an easy-to-use event time support also apply to Kafka Streams. It is a rather focused library, and it’s very well-suited for certain types of tasks. That’s also why some of its design can be so optimized for how Kafka works. You don’t need to set up any kind of special Kafka Streams cluster, and there is no cluster manager. And if you need to do a simple Kafka topic-to-topic transformation, count elements by key, enrich a stream with data from another topic, or run an aggregation or only real-time processing — Kafka Streams is for you.

If event time is not relevant and latencies in the seconds range are acceptable, Spark is the first choice. It is stable and almost any type of system can be easily integrated. In addition it comes with every Hadoop distribution. Furthermore, the code used for batch applications can also be used for the streaming applications as the API is the same.

Read on for more analysis.

Comparing Spark Streaming, Flink, And Kafka Streams

Shivangi Gupta contrasts three streaming technologies:

Flink and Spark are in-memory databases that do not persist their data to storage. They can write their data to permanent storage, but the whole point of streaming is to keep it in memory, to analyze current data. All of this lets programmers write big data programs with streaming data. They can take data in whatever format it is in, join different sets, reduce it to key-value pairs (map), and then run calculations on adjacent pairs to produce some final calculated value. They also can plug these data items into machine learning algorithms to make some projection (predictive models) or discover patterns (classification models).

Streaming has become the product-level battleground in the Hadoop ecosystem, and it’s interesting to see the different approaches that different groups have taken.

Security And Zookeeper

Michael Han describes a few methods you can use to tighten up (or rather, introduce) security in ZooKeeper:

Four Letter Words (acronym as 4lw) is a very popular feature of the Apache ZooKeeper project. In a nutshell, 4lw is a set of commands that you can use to interact with a ZooKeeper ensemble through a shell interface. Because it’s simple and easy to use, lots of ZooKeeper monitoring solutions are built on top of 4lw.

The simplicity of 4lw comes at a cost: the design did not originally consider security, there is no built in support for authentication and access control. Any user that has access to the ZooKeeper client port can send commands to the ensemble. The 4lw commands are read only commands: no actions can be performed. However, they can be computing intensive, and sending too many of them would effectively create a DOS attack that prevents the ensemble’s normal operation.

Read on for details.

Why Hadoop BI Projects Fail

Remy Rosenbaum lays out several reasons why he’s seen business intelligence projects on Hadoop fail:

In order to set up and run an effective Big Data Hadoop project that provides reliable BI, your organization will need to adopt a new mindset that addresses not only the technology, but also the organizational EIM. You will need to conduct a comprehensive analysis of your business with the help of analysts, internal domain experts, and strategists to come up with robust and relevant business use cases. You will also need buy-in from management, and take company politics into consideration.

Your Big Data project needs to work with your existing BI tools, along with your security and monitoring systems. Data security needs to be addressed because standard Hadoop implementations have relatively poor security, and many organizations are wary of keeping all their data in one location.

I do agree with these reasons, though I’m a bit surprised that I didn’t see much about “classic” BI problems like the inability of the company to standardize on terminology or definitions (e.g., what the Kimball method describes as conformed dimensions), the desire to tackle too much of the problem at once, rapidly-changing source systems (and how BI team members tend to be the last to know that something has changed), etc.

Lamba Architecture Basics

Michael Walker walks through the basics of the Lambda architecture:

Lambda architecture – developed by Nathan Marz – provides a clear set of architecture principles that allows both batch and real-time or stream data processing to work together while building immutability and recomputation into the system. Batch processes high volumes of data where a group of transactions is collected over a period of time. Data is collected, entered, processed and then batch results produced. Batch processing requires separate programs for input, process and output. An example is payroll and billing systems. In contrast, real-time data processing involves a continual input, process and output of data. Data must be processed in a small time period (or near real-time). Customer services and bank ATMs are examples.

Lambda architecture has three (3) layers:

  • Batch Layer

  • Serving Layer

  • Speed Layer

I haven’t heard much about the Lambda and Kappa architectures lately, so when I saw this, I figured it was time for a refresher.

Kafka For Pythonistas

Matt Howlett has an introduction to Apache Kafka, designed for Python developers:

In the call to the produce method, both the key and value parameters need to be either a byte-like object (in Python 2.x this includes strings), a Unicode object, or None. In Python 3.x, strings are Unicode and will be converted to a sequence of bytes using the UTF-8 encoding. In Python 2.x, objects of type unicode will be encoded using the default encoding. Often, you will want to serialize objects of a particular type before writing them to Kafka. A common pattern for doing this is to subclass Producer and override the produce method with one that performs the required serialization.

The produce method returns immediately without waiting for confirmation that the message has been successfully produced to Kafka (or otherwise). The flush method blocks until all outstanding produce commands have completed, or the optional timeout (specified as a number of seconds) has been exceeded. You can test to see whether all produce commands have completed by checking the value returned by the flush method: if it is greater than zero, there are still produce commands that have yet to complete. Note that you should typically call flush only at application teardown, not during normal flow of execution, as it will prevent requests from being streamlined in a performant manner.

This is a fairly gentle introduction to the topic if you’re already familiar with Python and have a familiarity with message broker systems.

S3 And HDFS Data Migration

Ilya Yalovyy looks at S3DistCp, which allows you efficiently to migrate data back and forth between HDFS and S3:

Raw files often land in S3 or HDFS in an uncompressed text format. This format is suboptimal both for the cost of storage and for running analytics on that data. S3DistCp can help you efficiently store data and compress files on the fly with the --outputCodec option:

$ s3-dist-cp --src s3://my-tables/incoming/hourly_table_filtered --dest s3://my-tables/incoming/hourly_table_gz --outputCodec=gz

The current version of S3DistCp supports the codecs gzip, gz, lzo, lzop, and snappy, and the keywords none and keep (the default). These keywords have the following meaning:

  • none” – Save files uncompressed. If the files are compressed, then S3DistCp decompresses them.

  • keep” – Don’t change the compression of the files but copy them as-is.

This is an important article if you’ve got a Hadoop cluster running on EC2 nodes.

Categories

June 2017
MTWTFSS
« May  
 1234
567891011
12131415161718
19202122232425
2627282930