The big news here is the recently released preview of HDInsight IO Cache, which is a new transparent data caching feature that provides customers with up to 9X performance improvement for Spark jobs, without an increase in costs.
There are many open source caching products that exist in the ecosystem: Alluxio, Ignite, and RubiX to name a few big ones. The IO Cache is also based on RubiX and what differentiates RubiX from other comparable caching products is its approach of using SSD and eliminating the need for explicit memory management. While other comparable caching products leverage the reservation of operating memory for caching the data.
Read on for more details.
Basically, the idea is to keep the fast stuff fast and the slow stuff slow. I wrote a paper 14 years ago on the challenges of real-time data warehousing. Fortunately, both the data streaming, database, and BI layers have all evolved significantly since then, and now there exist databases and other data storage engines which can support the feature trinity that is needed to do both real-time and historical analytics right, without a Lambda architecture:
- Accept real-time streams of data at high rates.
- Simultaneously respond to large volumes of queries, including on the most recently added data.
- Store all the history needed for analysis.
We call these engines “fast data sinks” and there are four main groups of them today:
It’s an interesting argument.
To access data to display in our dashboard we will use some Spring Boot 2.06 Java 8 microservices to call Apache Hive 3.1.0 tables in HDP 3.0 on Hadoop 3.1.
We will have our website hosted and make REST Calls to Apache NiFi, our microservices, YARN, and other APIs.
As you can see we can easily incorporate data from HDP 3 — Apache Hive 3.1.0 in Spring Boot Java applications with not much trouble. You can see the Maven build script (all code is in GitHub).
Our motivation is to put all this data somewhere and show it on a dashboard that can use REST APIs for data access and updates. We may choose to use Apache NiFi for all REST APIs or we can do some in Apache NiFi. We are still exploring. We can also decide to change the backend to HBase 2.0, Phoenix, Druid or a combination of these. We will see.
Read on for a series of screenshots and config files showing you how to set this up.
We are also focusing on efficiency across our platform. While on-premises platform efficiency helps manage costs in the long run, the immediate benefits of in-cloud deployments are realized by reducing total cost of ownership (TCO). We introduced Hive-on-Spark two years ago to meet this goal in collaboration with Intel which is our strategic partner. We have a longstanding collaboration with Intel to optimize Cloudera’s stack on Intel architecture for our customers’ benefit.
In Enterprise 6.0, taking our strategic partnership with Intel ahead for further efficiency gains in Hive, we introduce a major performance and efficiency enhancement in HoS called Parquet Vectorization. This feature enables the HoS engine to process a vector of columns instead of one row at a time by batching data rows together into column vectors and making each operator work on such column vectors. This leads to better utilization of CPU caches and achieves high instructions per cycle by efficiently using the CPU instruction pipeline. In addition, we include numerous other performance improvements. For example, Hive often scans a given table multiple times during self joins, self-unions, or shared sub-queries. To address this, Dynamic RDD caching in HoS reuses a single scan across all these operations. Similarly, when the same subquery is used repeatedly, HoS executes this only once instead of separately for each subquery invocation. Overall, with all these enhancements, in Enterprise 6.0 Hive can be up to 2.2X faster than Hive on the latest Enterprise 5.x release. The majority of these gains can be attributed to Parquet Vectorization for Hive-on-Spark.
This is another case where the Cloudera-Hortonworks merger will get interesting: Cloudera seemed to hitch its wagon to Impala and Hortonworks to Hive; will they support both as much as they each did independently, or will the new corporate overlords settle on one of the two?
First, if you are running most of your other applications and microservices on Kubernetes, it becomes the organizational path of least resistance. This is just like how organizations who standardized on VMs have found it very difficult to allocate physical machines with local disks for Kafka.
I see situations with larger organizations where deploying Kafka outside of Kubernetes causes significant organizational headache that involves many approvals. When this is the case, I usually say that this isn’t a good hill to die on. It is possible to run Kafka on Kubernetes, so just do it. You’ll get your environment allocated faster and will be able to use your time to do productive work rather than fight an organizational battle.
And if things go wrong, you’ll get much better service from your internal infrastructure teams, because you’ll be running in an environment that is familiar to them.
Read on for more benefits as well as a few drawbacks.
Whatever camp you sit in, the merger undoubtedly caught the attention of the 2,500 organizations that have adopted Cloudera’s Distribution of Hadoop (CDH) or the Hortonworks Data Platform (HDP) over the years — not to mention the thousands of other companies that have adopted open source Apache Hadoop platforms or Hadoop ecosystem components in the cloud. These Global 2000 companies have invested billions of dollars into building giant clusters to store and process many exabytes worth of data, and they’re not going to just turn them off overnight because the two biggest players suddenly decided to merge.
At the same time, these customers need to be reassured that Cloudera has a plan to maintain the investments they’ve already made in HDP and CDH platforms, both in a short-term or tactical sense, as well as in terms of Cloudera’s long-range strategy to evolve its platform to meet emerging future compute and storage needs.
Read on for more detail.
Real-time stream processing is becoming more prevalent on modern day data platforms, and with a myriad of processing technologies out there, where do you begin? Stream processing involves the consumption of messages from either queue/files, doing some processing in the middle (querying, filtering, aggregation) and then forwarding the result to a sink – all with a minimal latency. This is in direct contrast to batch processing which usually occurs on an hourly or daily basis. Often is this the case, both of these will need to be combined to create a new data set.
In terms of options for real-time stream processing on Azure you have the following:
Azure Stream Analytics
Spark Streaming / Storm on HDInsight
Spark Streaming on Databricks
Click through for more.
The Hadoop Distributed File System (HDFS) allows you to both federate storage across many computers as well as distribute files in a redundant manor across a cluster. HDFS is a key component to many storage clusters that possess more than a petabyte of capacity.
Each computer acting as a storage node in a cluster can contain one or more storage devices. This can allow several mechanical storage drives to both store data more reliably than SSDs, keep the cost per gigabyte down as well as go some way to exhausting the SATA bus capacity of a given system.
Hadoop ships with a feature-rich and robust JVM-based HDFS client. For many that interact with HDFS directly it is the go-to tool for any given task. That said, there is a growing population of alternative HDFS clients. Some optimise for responsiveness while others make it easier to utilise HDFS in Python applications. In this post I’ll walk through a few of these offerings.
Read on for reviews of those offerings.
In this little proof of concept work, we grab some of these flows process them in Apache NiFi and then store them in Apache Hive 3 tables for analytics. We should probably push the data to HBase for aggregates and Druid for time series. We will see as this expands.
There are also other data access options including the NiFi REST API and the NiFi Python APIs.
- Send notification when the NiFi starts, stops or died unexpectedly
- Two OOTB notifications
- Email notification service
- HTTP notification service
- It’s easy to write a custom notification service
AmbariReportingTask (global, per process group)
MonitorDiskUsage(Flowfile, content, provenance repositories)
Much of this is an overview of the tools and measures available.
While extract, transform, load (ETL) has its use cases, an alternative to ETL is data virtualization, which integrates data from disparate sources, locations, and formats, without replicating or moving the data, to create a single “virtual” data layer. The virtual data layer allows users to query data from many sources through a single, unified interface. Access to sensitive data sets can be controlled from a single location. The delays inherent to ETL need not apply; data can always be up to date. Storage costs and data governance complexity are minimized. See the pro’s and con’s of data virtualization via Data Virtualization vs Data Warehouse and Data Virtualization vs. Data Movement.
SQL Server 2019 big data clusters with enhancements to PolyBase act as a virtual data layer to integrate structured and unstructured data from across the entire data estate (SQL Server, Azure SQL Database, Azure SQL Data Warehouse, Azure Cosmos DB, MySQL, PostgreSQL, MongoDB, Oracle, Teradata, HDFS, Blob Storage, Azure Data Lake Store) using familiar programming frameworks and data analysis tools:
James covers some of the reasoning behind this and the shift from using Polybase to integrate data with Hadoop + Azure Blob Storage to using SQL Server as a data virtualization engine.