AUTHOR
Adservio team
DATE
July 28, 2021
CATEGORY
Digital Delivery
How to Build a Data Pipeline Using Kafka, Spark, and Hive

Digital Delivery

12 min

How to Build a Data Pipeline Using Kafka, Spark, and Hive

As innovators race toward real-time, the wise CIO is the one who embraces the fastest and most efficient means of analyzing data.

Companies are shifting analytics from a siloed activity to a core business function.

Business leaders now understand the importance of real-time data for informed decision-making.

Information analysis dramatically changes the way leaders use data to project outcomes.

Instead of making decisions based on a series of past events they can now use analytics to gain immediate insights.

These insights can help organizations deal with disruptive change, radical uncertainty, and opportunities, says Rita Sallam, Distinguished VP Analyst, Gartner.

Overall, real-time analytics enable faster, more precise decisions than those with stale information or none at all.

In this guide, we’ll discuss how to build a data pipeline using Spark, Kafka, and Hive to help drive your data initiatives.

Kafka, Spark, Hive Overview

Kafka, Spark, and Hive are three individual technologies that when combined help companies build robust data pipelines.

Here is a brief overview of and the benefits of each.

Kafka

Apache Kafka handles the first stage of a data pipeline. It is a platform used to ingest process real-time data streams.

It is a scalable, high-performance, low latency messaging system that works on a publisher-subscribe model.

Kafka’s fast, durable and fault tolerance capabilities make it ideal for use cases where RabitMQ, AMPMQ, and JMS aren’t suitable for high-volume processing.

This Kafka platform is popular among developers due to its simplicity. It is easy to set up, configure and get up and running in very little time.

Spark

Spark is an information processing platform that ingests information from data sources and makes simple work of processing tasks on large datasets.

Spark can handle several petabytes of information at a time.

The system is high-performing because it distributes tasks across multiple computers. It also performs multiple operations consecutively using memory rather than disk space.

Hive

The next stage of the pipeline is storage.

Hive is a data warehouse that is integrated with Hadoop and was designed to process and store massive amounts of information.

The platform is fault-tolerant, highly scalable, and flexible. As a cloud-based service, developers can quickly scale virtual servers up or down depending on the workload.

Hive has many benefits, one of which is its ease of use.

Hive uses a SQL-like language which should be familiar to many developers. Because it stores information on HDFS, it is a much more scalable solution than a traditional database.

Lastly, Hive has significant capacity capabilities that can support up to 100,000 queries/hour.

Build a Data Pipeline Using Kafka

Pipelines must provide a publisher-subscribe messaging system to process incoming data.

Messaging works on the concept of streaming events. These events represent real-world data coming from a variety of data sources.

The process works via the following components:

  • Events - An event represents some sort of change that took place in the system. It could be a record update, delete, or some other specified action.
  • Topics - Events are organized into topics. A topic is the equivalent of a filesystem that stores files.
  • Producers - Publishers are source systems that publish (write) events to the messaging queue.
  • Consumers - Consumers represent systems that require the updated information from the source. These consumers subscribe to the messaging queue and will receive updates as they occur.

Build a Data Pipeline Using Spark

Spark is a consumer of real-time data from Kafka and performs processing on the information received.

The components of Spark include data sources, receivers, and streaming operations.

Sources In Spark

  • Basic Sources - Sources such as file systems and socket connections are basic sources that can be accessed directly in the StreamingContext API.
  • Advanced Sources - Advanced sources are those that are only available via utility classes. Kafka is considered an advanced data source in the context of Spark and requires linking to additional dependencies to work.

Receivers

Receivers refer to Spark objects that receive data from source systems.

Spark has two types of receivers:

  • Reliable Receiver - This type of receiver sends an acknowledgment when Spark receives the data. A reliable receiver is critical for systems that require an acknowledgment for further processing.
  • Unreliable Receiver - An unreliable receiver does not send an acknowledgment. It is suitable for less complex systems that don’t require an acknowledgment that the data was received.

Streaming Operations

Spark performs two operations on the data received:

  • Transformations - Transformations represent the operations Spark performs on the data input. Examples include filtering data based on set criteria, mapping input values to a pair of values, counting the number of elements.
  • Output Operations - Output operations are concerned with pushing the data to external systems.

Build a Data Pipeline Using Hive

Traditional databases can’t handle the processing and storage requirements for large datasets.

Given that, the system must use a high-volume storage system that can handle petabytes of information.

Hive accomplishes this task using the concept of buckets.

What are Buckets?

Bucketing improves query performance for large data sets by breaking data down into ranges known as buckets.

With this approach, sampling is greatly improved because the data is already split into smaller chunks.

Common Pipeline Architectures

Pipelines can be created using two common architectures. Each offers specific features.

Deciding which to choose depends on the use case.

Lambda Architecture

The lambda architecture consists of three layers:

Batch

The system pushes a continuous feed of information into the batch layer and the speed layer simultaneously.

Once ingested, the batch layer computes, processes, and then stores it. It is then fed to the speed layer for further processing.

Kafka is often used at this layer to ingest information into the pipeline.

Speed (Stream Layer)

The speed layer receives the output of the batch layer. It uses this data to enrich the real-time information received during initial ingestion.

The dataset is then pushed to the serving layer. Spark is often used at this level.

Serving

The serving layer receives the output of the speed layer. Leaders can use this dataset for queries, analysis, and reporting.

Hive is the database often used at this layer for storage.

Kappa Architecture

The Kappa architecture is not a replacement for lambda. It is an alternate option that uses fewer code resources.

Kappa is recommended for instances where the lambda architecture is too complex.

This approach eliminates the batch layer. All processing happens at the streaming layer and the serving layer.

Both Kafka and Spark operate at the streaming layer while Hive operates at the serving layer.

5 Important Features of a Robust Data Pipeline

A pipeline must contain several functions to provide real-time analytics data. Your team will have a safer foundation for making better business decisions by ensuring your data pipeline contains the features below;

1. Analytics

Data pipelines support predictive analysis by relaying real-time information to key business systems.

Leaders can use this information to gain insight into key performance indicators that help them determine if the organization is meeting its goals.

Apache Spark facilitates this process by performing transformations on data received from the messaging system.

The system can perform operations on data such as mapping input values to a pair of values or counting the number of elements in a data stream.

2. High Volume Storage

Traditional databases can’t handle the processing and storage requirements for large datasets.

Given that, the system must use a high-volume storage system that can handle petabytes of information. Hive accomplishes this task using the concept of buckets.

Bucketing improves query performance for large data sets by breaking data down into ranges known as buckets.

With this approach, sampling is greatly improved because the data is already split into smaller chunks.

3. Self-Service Management

The tools, programming languages, and maintenance required of a pipeline require technical skills that may prove too steep for business users.

As a result, companies are finding new ways to democratize access to data.

This approach provides non-technical users less complex tools to access information.

These tools don’t require programming and allow users to build a new pipeline without involving IT.

4. Streamlined Data Pipeline Development

Streamlining data pipeline development is now easier than ever. Now IT has access to APIs, connectors and pre-built integrations to make the process more efficient.

These tools make it easy to scale or integrate new data sources.

5. Exactly-Once Processing (E1P)

Exactly-Once Processing (E1P) ensures that a message is written exactly once to a topic.

It ensures that a message is written only once regardless of how many times the producer sends the message.

The process prevents data loss in transit while avoiding receiving the same data multiple times.  

A data pipeline is essential to feeding real-time insight to business leaders.

Using these insights, they can make strategic decisions to respond to change and capitalize on market opportunities.

Contact us to learn how we can help you build data pipelines that empower your business with the information you need to define strategies for success.

Industry insights you won’t delete. Delivered to your inbox weekly.

Other posts

Any questions for us?