Hello geeks, we have discussed how to start programming with Spark in Scala. In this blog, we will discuss how we can use Hive with Spark 2.0.
When you start to work with Hive, you need HiveContext (inherits SqlContext), core-site.xml,hdfs-site.xml, and hive-site.xml for Spark. In case you don’t configure hive-site.xml then the context automatically creates metastore_db in the current directory and creates warehousedirectory indicated by HiveConf(which defaults user/hive/warehouse).
Rahul has made his demo code available on GitHub.
Several people have approached these letters as a cryptographic cipher. The odd circumstances of death do sound like something out of a John Le Carré spy novel. Some of the best attempts, however, fail to produce anything but truly convoluted parsings.
Another possibility may already have occurred to you: Are they the first letters of words in a sentence (aninitialism)? Some suspect this death was a suicide, and that the message is merely some form of final note. With this morbid scenario in mind, it’s easy to imagine many phrases, like “My Life Is All But Over,” that fit the letters because indeed their frequency seems to match that of English text.
This lead has been picked up a few times. These writeups (example) present indications that the message is indeed an initialism. However, they don’t apply what is arguably the clear statistical tool for this job. And they don’t take advantage of big data. So, let’s do both.
Read on for Chi Square testing and book parsing examples using Spark. Spoiler alert: Sean doesn’t solve the mystery, but it’s still a fun read.
Debugging at full scale can be slow, challenging, and resource intensive. We started off by converting the most resource intensive part of the Hive-based pipeline: stage two. We started with a sample of 50 GB of compressed input, then gradually scaled up to 300 GB, 1 TB, and then 20 TB. At each size increment, we resolved performance and stability issues, but experimenting with 20 TB is where we found our largest opportunity for improvement.
While running on 20 TB of input, we discovered that we were generating too many output files (each sized around 100 MB) due to the large number of tasks. Three out of 10 hours of job runtime were spent moving files from the staging directory to the final directory in HDFS. Initially, we considered two options: Either improve batch renaming in HDFS to support our use case, or configure Spark to generate fewer output files (difficult due to the large number of tasks — 70,000 — in this stage). We stepped back from the problem and considered a third alternative. Since the tmp_table2 table we generate in step two of the pipeline is temporary and used only to store the pipeline’s intermediate output, we were essentially compressing, serializing, and replicating three copies for a single read workload with terabytes of data. Instead, we went a step further: Remove the two temporary tables and combine all three Hive stages into a single Spark job that reads 60 TB of compressed data and performs a 90 TB shuffle and sort.
Maybe it’s just a mindset thing, but the part that impressed me was the number of pull requests for system improvements (and the number which were accepted).
The built-in machine learning library in Spark is broken into two parts: MLlib and KeystoneML.
MLlib: This is the principal library for machine learning tasks. It includes both algorithms and specialized data structures. Machine learning algorithms for clustering, regression, classification, and collaborative filtering are available. Data structures such as sparse and dense matrices and vectors, as well as supervised learning structures that act like vectors but denote the features of the data set from its labels, are also available. This makes feeding data into a machine learning algorithm incredibly straightforward and does not require writing a bunch of code to denote how the algorithm should organize the data inside itself.
KeystoneML: Like the oil pipeline it takes its name from, KeystoneML is built to help construct machine learning pipelines. The pipelines help prepare the data for the model, build and iteratively test the model, and tune the parameters of the model to squeeze out the best performance and capability.
Whereas Hadoop’s ecosystem is large and sprawling, the Spark ecosystem tends to be more tightly constrained. The nice part about Spark is that it plays nicely with the Hadoop ecosystem—you can have a cluster or architecture with Spark and Hadoop-centric technologies (Storm, Kafka, Hive, Flume, etc. etc.) working together quite nicely.
Notebooks are very helpful in building a pipeline even with compiled artifacts. Being able to visualize data and interactively experiment with transformations makes it much easier to write code in small, testable chunks. More importantly, the development of most data pipelines begins with exploration, which is the perfect use case for notebooks. As an example, Yesware regularly uses Databricks Notebooks to prototype new features for their ETL pipeline.
On the flip side, teams also run into problems as they use notebooks to take on more complex data processing tasks:
- Logic within notebooks becomes harder to organize. Exploratory notebooks start off as simple sequences of Spark commands that run in order. However, it is common to make decisions based on the result of prior steps in a production pipeline – which is often at odds with how notebooks are written during the initial exploration.
- Notebooks are not modular enough. Teams need the ability to retry only a subset of a data pipeline so that a failure does not require re-running the entire pipeline.
These are the common reasons that teams often re-implement notebook code for production. The re-implementation process is time-consuming, tedious, and negates the interactive properties of notebooks.
Those two reasons are why I’ve argued that you should sit down in front of a REPL and figure out what you’re doing with a particular data set. Once you’ve got it figured out, perform the operations in a notebook for posterity and to replicate your actions later. I’m curious to see how this gets adopted in practice.
For data scientists, we provide out-of-the-box integration with Jupyter (iPython), the most popular open source notebook in the world. Unlike other managed Spark offerings that might require you to install your own notebooks, we worked with the Jupyter OSS community to enhance the kernel to allow Spark execution through a REST endpoint.
We co-led “Project Livy” with Cloudera and other organizations to create an open source Apache licensed REST web service that makes Spark a more robust back-end for running interactive notebooks. As a result, Jupyter notebooks are now accessible within HDInsight out-of-the-box. In this scenario, we can use all of the services in Azure mentioned above with Spark with a full notebook experience to author compelling narratives and create data science collaborative spaces. Jupyter is a multi-lingual REPL on steroids. Jupyter notebook provides a collection of tools for scientific computing using powerful interactive shells that combine code execution with the creation of a live computational document. These notebook files can contain arbitrary text, mathematical formulas, input code, results, graphics, videos and any other kind of media that a modern web browser is capable of displaying. So, whether you’re absolutely new to R or Python or SQL or do some serious parallel/technical computing, the Jupyter Notebook in Azure is a great choice.
If you could only learn one new thing in 2016, Spark probably should be that thing. Also, I probably should agitate a bit more about wanting Spark support within Polybase…
The other type of data are data streams. Data streams can be visualized by water flowing from a tap to a sink. This process is not ending. The nice property of streams is that you can consume the stream while it is flowing. There is almost no latency involved for consuming a stream.
Apache Spark is fundamentally based on batches of data. By that, for all processing jobs at least some latency is introduced. Apache Flink on the other hand is fundamentally based on streams. Let’s take a look at some evidence for the difference in latency.
Read the whole thing.
Looking at the first 5 records of the RDD
This output is difficult to read. This is because we are asking PySpark to show us data that is in the RDD format. PySpark has a DataFrame functionality. If the Python version is 2.7 or higher, you can utilize the pandas package. However, pandas doesn’t work on Python versions 2.6, so we use the Spark SQL functionality to create DataFrames for exploration.
The full example is a fairly simple k-means clustering process, which is a great introduction to PySpark.
Beyond a time-bounded interaction, SparkSession provides a single point of entry to interact with underlying Spark functionality and allows programming Spark with DataFrame and Dataset APIs. Most importantly, it curbs the number of concepts and constructs a developer has to juggle while interacting with Spark.
In this blog and its accompanying Databricks notebook, we will explore SparkSession functionality in Spark 2.0.
This looks to be an easier method for integrating various parts of Spark in one user session. Read the whole thing.
Modeled using principles from the “AWS Dynamo” paper, Riak KV buckets are good for scenarios which require frequent, small data-sized operations in near real-time, especially workloads with reads, writes, and updates — something which might cause data corruption in some distributed databases or bring them to “crawl” under bigger workloads. In Riak, each data item is replicated on several nodes, which allows the database to process a huge number of operations with very low latency while having unique anti-corruption and conflict-resolution mechanisms. However, integration with Apache Spark requires a very different mode of operation — extracting large amounts of data in bulk, so that Spark can do its “magic” in memory over the whole data set. One approach to solve this challenge is to create a myriad of Spark workers, each asking for several data items. This approach works well with Riak, but it creates unacceptable overhead on the Spark side.
This is interesting in that it ties together two data platforms whose strengths are almost the opposite: one is great for fast, small writes of single records and the other is great for operating on large batches of data.