Press "Enter" to skip to content

Category: Hadoop

Build Versus Buy For Hadoop

Published 2017-10-24 by Kevin Feasel

Tom Phelan walks through some thoughts on whether to build versus buy when using big data platforms:

This means you absolutely must sweat the details up front. Big Data project failures are more often than not predicated by the statement: “We will do this bit now, and figure the rest out later”. But you need to begin with the end in mind.

You need to know the performance that you’ll be able to deliver and what your requirements are. You need to know how to integrate with your corporate Active Directory, and LDAP, and Kerberos services. You need to know your network topology and security requirements as well as the required user roles and responsibilities breakdown. You need to know how you’ll handle high availability, QoS, and multi-tenancy. You need to know how you’ll manage upgrades to the latest versions of your Hadoop distribution or other big data tools, and how you’ll respond to requests for new big data frameworks and new data science tools. If not, you’re just asking for trouble.

The motif in his post is building your own car, which makes sense as an extended metaphor.

Comments closed

Page Ranking With Kafka Streams

Published 2017-10-20 by Kevin Feasel

Hunter Kelly walks through a page ranking algorithm:

Once you have the adjacency matrix, you perform some straightforward matrix calculations to calculate a vector of Hub scores and a vector of Authority scores as follows:

  • Sum across the columns and normalize, this becomes your Hub vector
  • Multiply the Hub vector element-wise across the adjacency matrix
  • Sum down the rows and normalize, this becomes your Authority vector
  • Multiply the Authority vector element-wise down the the adjacency matrix
  • Repeat

An important thing to note is that the algorithm is iterative: you perform the steps above until  eventually you reach convergence—that is, the vectors stop changing—and you’re done. For our purposes, we just pick a set number of iterations, execute them, and then accept the results from that point.  We’re mostly interested in the top entries, and those tend to stabilize pretty quickly.

This is an architectural-level post, so there’s no code but there is a useful discussion of the algorithm.

Comments closed

Stateful Processing In Spark Streaming

Published 2017-10-20 by Kevin Feasel

Bill Chambers and Jules Damji look at a couple of stateful scenarios within Spark Streaming:

No streaming events are free of duplicate entries. Dropping duplicate entries in record-at-a-time systems is imperative—and often a cumbersome operation for a couple of reasons. First, you’ll have to process small or large batches of records at time to discard them. Second, some events, because of network high latencies, may arrive out-of-order or late, which may force you to reiterate or repeat the process. How do you account for that?

Structured Streaming, which ensures exactly once-semantics, can drop duplicate messages as they come in based on arbitrary keys. To deduplicate data, Spark will maintain a number of user-specified keys and ensure that duplicates, when encountered, are discarded.

Just as other stateful processing APIs in Structured Streaming are bounded by declaring watermarking for late data semantics, so is dropping duplicates. Without watermarking, the maintained state can grow infinitely over the course of your stream.

In this scenario, you would still want some sort of de-duplication code at the far end of your process if you can never have duplicates come in across the lifetime of the application.  This sounds like it’s more about preventing bursty duplicates from sensors.

Comments closed

Optimizing Apache Flink

Published 2017-10-19 by Kevin Feasel

Ivan Mushketyk has a few tips for speeding up programs using Apache Flink:

One more way to optimize your Flink application is to provide some information about what your user-defined functions are doing with input data. Since Flink can’t parse and understand code, you can provide crucial information that will help to build a more efficient execution plan. There are three annotations that we can use:

  1. @ForwardedFields: Specifies what fields in an input value were left unchanged and are used in an output value.

  2. @NotForwardedFields: Specifies fields that were not preserved in the same positions in the output.

  3. @ReadFields: Specifies what fields were used to compute a result value. You should only specify fields that were used in computations and not merely copied to the output.

Click through for his four tips.

Comments closed

Anomaly Detection With Kafka Streams

Published 2017-10-17 by Kevin Feasel

Ajmal Karuthakantakath shows us an application which performs fairly simple anomaly detection using Kafka Streams:

The problem is in the banking loan payment domain, where customers have taken a loan and they need to make monthly payments to repay the loan amount.

Assume there are millions of customers in the system and all these customers need to make monthly payments to their account. Each customer may have a different monthly due date depending on their monthly loan due date.

Each customer payment will appear as a PaymentScheduleEvent event. Customers can make more than one PaymentScheduleEvent per month. Each monthly due date for a customer will appear as a PaymentDueEvent.

An arbitrarily chosen anomaly condition for this example is that if the amount due is more than $150 for any customer at any point in time, this generates an anomaly.

Click through for instructions, the application, and further resources.  If you want to learn Kafka Streams, this should keep you busy for a little while.

Comments closed

Crossing The Streams With Kafka

Published 2017-10-16 by Kevin Feasel

Himani Arora shows how to join two Kafka streams together:

KStream-KStream Join

It is a sliding window join, that means, all tuples close to each other with regard to time are joined. Time here is the difference up to size of the window.

These joins are always windowed joins because otherwise, the size of the internal state store used to perform the join would grow indefinitely.

In the following example, we perform an inner join between two streams. The output the joined stream will be of type KStream<K, ...>

Read on to learn more about two additional join types.

Comments closed

Benchmarking Streaming Systems

Published 2017-10-13 by Kevin Feasel

Burak Yavuz shares a benchmark of Spark Streaming versus Flink and Kafka Streams:

At Databricks, we used Databricks Notebooks and cluster management to set up a reproducible benchmarking harness that compares the performance of Apache Spark’s Structured Streaming, running on Databricks Unified Analytics Platform, against other open source streaming systems such as Apache Kafka Streams and Apache Flink. In particular, we used the following systems and versions in our benchmarks:

The Yahoo Streaming Benchmark is a well-known benchmark used in industry to evaluate streaming systems. When setting up our benchmark, we wanted to push each streaming system to its absolute limits, yet keep the business logic the same as in the Yahoo Streaming Benchmark. We shared some of the results we achieved from these benchmarks during Spark Summit West 2017 keynote showing that Spark can reach 5x or higher throughput over other popular streaming systems. In this blog, we discuss in more detail about how we performed this benchmark, and how you can reproduce the results yourselves.

Standard vendor-based metric warnings aside, you can get the benchmark details at their GitHub repo.

Comments closed

Testing Azure Data Lake Store Performance

Published 2017-10-12 by Kevin Feasel

Zhen Zeng and Govind Kamat stress test Azure Data Lake Store:

Now that we know the read and write throughput characteristics of a single Data Node, we would like to see how per-node performance scales when the number of Data Nodes in a cluster is increased.

The tool we use for scale testing is the Tera* suite that comes packaged with Hadoop.  This is a benchmark that combines performance testing of the HDFS and MapReduce layers of a Hadoop cluster.  The suite is comprised of three tools that are typically executed in sequence:

  • TeraGen, that tool that generates the input data.  We use it to test the write performance of HDFS and ADLS.

  • TeraSort, which sorts the input data in a distributed fashion.  This test is CPU bound and we don’t really use it to characterize the I/O performance or HDFS and ADLS, but it is included for completeness.

  • TeraValidate, the test that reads and validates the sorted data from the previous stage.  We use it to test the read performance of HDFS and ADLS.

It’s an interesting look at how well ADLS scales.  In general, my reading of this is fairly positive for Azure Data Lake Store.

Comments closed

Predicting Advertising Budgets With Kafka Streams

Published 2017-10-11 by Kevin Feasel

Boyang Chen explains how Pinterest uses Kafka Streams to reduce advertising overdelivery:

Overdelivery occurs when free ads are shown for out-of-budget advertisers. This reduces opportunities for advertisers with available budget to have their products and services discovered by potential customers.

Overdelivery is a difficult problem to solve for two reason:

  1. Real-time spend data: Information about ad impressions needs to be fed back into the system within seconds in order to shut down out-of-budget campaigns.

  2. Predictive spend: Fast, historical spend data isn’t enough. The system needs to be able to predict spend that might occur in the future and slow down campaigns close to reaching their budget. That’s because an inserted ad could remain available to be acted on by a user. This makes the spend information difficult to accurately measure in a short timeframe. Such a natural delay is inevitable, and the only thing we can be sure of is the ad insertion event.

This is a very interesting architectural overview.

1 Comment

Polybase And HDInsight

Published 2017-10-11 by Kevin Feasel

I have a post up on trying to integrate Polybase with HDInsight:

But now we run into a problem:  there are certain ports which need to be open for Polybase to work.  This includes port 50010 on each of the data nodes against which we want to run MapReduce jobs.  This goes back to the issue we see with spinning up data nodes in Docker:  ports are not available.  If you’ve put your HDInsight cluster into an Azure VNet and monkey around with ports, you might be able to open all of the ports necessary to get this working, but that’s a lot more than I’d want to mess with, as somebody who hasn’t taken the time to learn much about cloud networking.

As I mention in the post, I’d much rather build my own Hadoop cluster; I don’t think you save much maintenance time in the long run going with HDInsight.

Comments closed