Category: Streaming

David Brinegar discusses consumer groups and lag in Apache Kafka:

While the Consumer Group uses the broker APIs, it is more of an application pattern or a set of behaviors embedded into your application.  The Kafka brokers are an important part of the puzzle but do not provide the Consumer Group behavior directly.  A Consumer Group based application may run on several nodes, and when they start up they coordinate with each other in order to split up the work.  This is slightly imperfect because the work, in this case, is a set of partitions defined by the Producer.  Each Consumer node can read a partition and one can split up the partitions to match the number of consumer nodes as needed.  If the number of Consumer Group nodes is more than the number of partitions, the excess nodes remain idle. This might be desirable to handle failover.  If there are more partitions than Consumer Group nodes, then some nodes will be reading more than one partition.

Read the whole thing.  It’s part one of a series.

Timothy Spann ties together a bunch of interesting things with Apache Nifi, including integrations with Twitter, Slack, Tensorflow, and Zeppelin:

First up, I used GetTwitter to read tweets and filtered on these terms:

strata, stratahadoop, strataconf, NIFI, FutureOfData, ApacheNiFi, Hortonworks, Hadoop, ApacheHive, HBase, ApacheSpark, ApacheTez, MachineLearning, ApachePhoenix, ApacheCalcite,ApacheStorm, ApacheAtlas, ApacheKnox, Apache Ranger, HDFS, Apache Pig, Accumulo, Apache Flume, Sqoop, Apache Falcon

Input:

InvokeHttp: I used this to download the first image URL from tweets.

It’s interesting to see this all tie together relatively easily.

Chris Marshall shows how to perform anomaly detection using AWS Kinesis Analytics:

The RANDOM_CUT_FOREST function greatly simplifies the programming required for anomaly detection.  However, understanding your data domain is paramount when performing data analytics.  The RANDOM_CUT_FOREST function is a tool for data scientists, not a replacement for them.  Knowing whether your data is logarithmic, circadian rhythmic, linear, etc. will provide the insights necessary to select the right parameters for RANDOM_CUT_FOREST.  For more information about parameters, see the RANDOM_CUT_FOREST Function.

Fortunately, the default values work in a wide variety of cases. In this case, use the default values for all but the subSampleSize parameter.  Typically, you would use a larger sample size to increase the pool of random samples used to calculate the anomaly score; for this post, use 12 samples so as to start evaluating the anomaly scores sooner.

Your SQL query outputs one record every ten seconds from the tumbling window so you’ll have enough evaluation values after two minutes to start calculating the anomaly score.  You are also using a cutoff value where records are only output to “DESTINATION_SQL_STREAM” if the anomaly score from the function is greater than 2 using the WHERE clause. To help visualize the cutoff point, here are the data points from a few runs through the pipeline using the sample Python script:

This kind of scenario is pretty cool—you could also do things like detecting service outages in streams (fewer than X events in a window, where X is some very small number relative to your overall data) or changes in advertising campaigns.

Neha Narkhede and Stephan Ewen compare Apache Flink versus Kafka Streams:

Before Flink, users of stream processing frameworks had to make hard choices and trade off either latency, throughput, or result accuracy. Flink was the first open source framework (and still the only one), that has been demonstrated to deliver (1) throughput in the order oftens of millions of events per second in moderate clusters, (2) sub-second latency that can be as low as few 10s of milliseconds, (3) guaranteed exactly once semantics for application state, as well as exactly once end-to-end delivery with supported sources and sinks (e.g., pipelines from Kafka to Flink to HDFS or Cassandra), and (4) accurate results in the presence of out of order data arrival through its support for event time. Flink is based on a cluster architecture with master and worker nodes. Flink clusters are highly available, and can be deployed standalone or with resource managers such as YARN and Mesos. This architecture is what allows Flink to use a lightweight checkpointing mechanism to guarantee exactly-once results in the case of failures, as well allow easy and correct re-processing via savepoints without sacrificing latency or throughput. Finally, Flink is also a full-fledged batch processing framework, and, in addition to its DataStream and DataSet APIs (for stream and batch processing respectively), offers a variety of higher-level APIs and libraries, such as CEP (for Complex Event Processing), SQL and Table (for structured streams and tables), FlinkML (for Machine Learning), and Gelly (for graph processing). Flink has been proven to run very robustly in production at very large scale by several companies, powering applications that are used every day by end customers.

The upshot is that the two products don’t do exactly the same thing, and there might be room in your organization for the two of them.

Kartik Paramasivam discusses data access problems and solutions within a streaming architecture:

Using a remote store: This is the traditional model for building applications. Here, when an application needs to process an event, it makes a remote call to a separate SQL or No-SQL database. In this model, write operations are always remote calls, but reads can be performed on a local cache in certain scenarios. There are a large number of applications at LinkedIn that fall into this category.

Another pattern is to use a remote cache (e.g., Couchbase) that is fronting a remote database (e.g., Oracle). If the remote cache is used primarily for reading adjunct data, then applications use an Oracle change capture stream (using Databus) to populate the remote cache.

This is a must-read if you’re looking at implementing a streaming architecture and need to do any kind of data enrichment.

Matthias J Sax shows how to reprocess input data using Kafka Streams:

In this blog post we describe how to tell a Kafka Streams application to reprocess its input data from scratch. This is actually a very common situation when you are implementing stream processing applications in practice, and it might be required for a number of reasons, including but not limited to: during development and testing, when addressing bugs in production, when doing A/B testing of algorithms and campaigns, when giving demos to customers or internal stakeholders, and so on.

The quick answer is you can do this either manually (cumbersome and error-prone) or you can use the new application reset tool for Kafka Streams, which is an easy-to-use solution for the problem. The application reset tool is available starting with upcoming Confluent Platform 3.0.1 and Apache Kafka 0.10.0.1.

In the first part of this post we explain how to use the new application reset tool. In the second part we discuss what is required for a proper (manual) reset of a Kafka Streams application. This parts includes a deep dive into relevant Kafka Streams internals, namely internal topics, operator state, and offset commits. As you will see, these details make manually resetting an application a bit complex, hence the motivation to create an easy to use application reset tool.

Being able to reprocess streams is a critical part of the Kappa architecture, and this article is a nice overview of how to do that if you’re using Kafka Streams.