Comparing Performance: HBase1 vs HBase2

Surbhi Kochhar takes us through performance improvements between HBase version 1 and HBase version 2:

We are loading the YCSB dataset with 1000,000,000 records with each record 1KB in size, creating total 1TB of data. After loading, we wait for all compaction operations to finish before starting workload test. Each workload tested was run 3 times for 15min each and the throughput* measured. The average number is taken from 3 tests to produce the final number. 

The post argues that there’s an improvement, but when a majority of cases end up worse (even if just a little bit), I’m not sure it’s much of an improvement.

The Transaction Log in Delta Tables

Burak Yavuz, et al, explain how the transaction log works with Delta Tables in Apache Spark:

When a user creates a Delta Lake table, that table’s transaction log is automatically created in the _delta_log subdirectory. As he or she makes changes to that table, those changes are recorded as ordered, atomic commits in the transaction log. Each commit is written out as a JSON file, starting with 000000.json. Additional changes to the table generate subsequent JSON files in ascending numerical order so that the next commit is written out as 000001.json, the following as 000002.json, and so on.

It’s interesting that they chose JSON instead of a binary transaction log like relational databases use.

Spark for .NET Developers

Ed Elliott has a long-form post covering spark-dotnet:

The .NET driver is made up of two parts, and the first part is a Java JAR file which is loaded by Spark and then runs the .NET application. The second part of the .NET driver runs in the process and acts as a proxy between the .NET code and .NET Java classes (from the JAR file) which then translate the requests into Java requests in the Java VM which hosts Spark.

The .NET driver is added to a .NET program using NuGet and ships both the .NET library as well as two Java jars. One jar is for Spark 2.3 and one for Spark 2.4, and you do need to use the correct one on your installed version of Scala.

As much as I’ve enjoyed his series, getting it in a single-post format is great.

Debugging Spark Applications in Visual Studio

Ed Elliott continues a series on spark-dotnet:

There are two approaches, one I have used for years with dotnet when I want to debug something that is challenging to get a debugger attached – think apps which spawn other processes and they fail in the startup routine. You can add a Debugger.Launch() to your program then when spark executes it, a prompt will be displayed and you can attach Visual Studio to your program. (as an aside I used to do this manually a lot by writing an __asm int 3 into an app to get it to break at an appropriate point, great memories but we don’t need to do that anymore luckily :).

The second approach is to start the spark-dotnet driver in debug mode which instead of launching your app, it starts and listens for incoming requests – you can then run your program as normal (f5), set a breakpoint and your breakpoint will be hit.

Read on to see how it’s done, as well as a possibly-accidental benefit to this.

Performance Tuning Neural Network Training

Sean Owen takes us through a few techniques for speeding up neural network model training:

Step #2: Use Early Stopping
Keras (and other frameworks) have built-in support for stopping when further training appears to be making the model worse. In Keras, it’s the EarlyStopping callback. Using it means passing the validation data to the training process for evaluation on every epoch. Training will stop after several epochs have passed with no improvement. restore_best_weights=True ensures that the final model’s weights are from its best epoch, not just the last one. This should be your default.

Sean focuses here on Keras + TensorFlow on Spark, but several of the tips are cross-product and generally applicable.

Machine Learning and Delta Lake

Brenner Heintz and Denny Lee walk us through solving data engineering problems with Delta Lake:

As a result, companies tend to have a lot of raw, unstructured data that they’ve collected from various sources sitting stagnant in data lakes. Without a way to reliably combine historical data with real-time streaming data, and add structure to the data so that it can be fed into machine learning models, these data lakes can quickly become convoluted, unorganized messes that have given rise to the term “data swamps.”

Before a single data point has been transformed or analyzed, data engineers have already run into their first dilemma: how to bring together processing of historical (“batch”) data, and real-time streaming data. Traditionally, one might use a lambda architecture to bridge this gap, but that presents problems of its own stemming from lambda’s complexity, as well as its tendency to cause data loss or corruption.

Read the whole thing.

Cloudera Stream Processing

Dinesh Chandrasekhar announces the new iteration of Cloudera’s streaming data processor:

Cloudera Stream Processing (CSP) is a product within the Cloudera DataFlow platform that packs Kafka along with some key streaming components that empower enterprises to handle some of the most complex and sophisticated streaming use cases. CSP provides advanced messaging, real-time processing and analytics on real-time streaming data using Apache Kafka. CSP also supports key management and monitoring capabilities powered by Cloudera Streams Management (CSM).

Sounds like they’re taking the Kafka route over Spark Streaming, Flink, Airflow, etc.

Kafka 2.3 and Kafka Connect Improvements

Robin Moffatt goes over improvements in Kafka Connect with the release of Apache Kafka 2.3:

A Kafka Connect cluster is made up of one or more worker processes, and the cluster distributes the work of connectors as tasks. When a connector or worker is added or removed, Kafka Connect will attempt to rebalance these tasks. Before version 2.3 of Kafka, the cluster stopped all tasks, recomputed where to run all tasks, and then started everything again. Each rebalance halted all ingest and egress work for usually short periods of time, but also sometimes for a not insignificant duration of time.

Now with KIP-415, Apache Kafka 2.3 instead uses incremental cooperative rebalancing, which rebalances only those tasks that need to be started, stopped, or moved. For more details, there are available resources that you can readlisten, and watch, or you can hear the lead engineer on the work, Konstantine Karantasis, talk about it in person at the upcoming Kafka Summit.

Looks like some nice improvements here.

The Databricks File System

Brad Llewellyn takes us through the Azure Databricks File System:

Today, we’re going to talk about the Databricks File System (DBFS) in Azure Databricks.  If you haven’t read the previous posts in this series, IntroductionCluster Creation and Notebooks, they may provide some useful context.  You can find the files from this post in our GitHub Repository.  Let’s move on to the core of this post, DBFS.

As we mentioned in the previous post, there are three major concepts for us to understand about Azure Databricks, Clusters, Code and Data.  For this post, we’re going to talk about the storage layer underneath Azure Databricks, DBFS.  Since Azure Databricks manages Spark clusters, it requires an underlying Hadoop Distributed File System (HDFS).  This is exactly what DBFS is.  Basically, HDFS is the low cost, fault-tolerant, distributed file system that makes the entire Hadoop ecosystem work.  We may dig deeper into HDFS in a later post.  For now, you can read more about HDFS here and here.

Click through for more detail on DBFS.

An Apache Flume Overview

Daniel Berman takes us through an overview of Apache Flume:

Apache Flume was developed by Cloudera to provide a way to quickly and reliably stream large volumes of log files generated by web servers into Hadoop. There, applications can perform further analysis on data in a distributed environment. Initially, Apache Flume was developed to handle only log data. Later, it was equipped to handle event data as well.

Click through to get a code-free, high-level understanding of Flume and where it can work for you.

Categories

August 2019
MTWTFSS
« Jul  
 1234
567891011
12131415161718
19202122232425
262728293031