Apache Spark for Data Processing – The Natural Choice?

The following article is an excerpt from our new ebook: Comparing Data Lake Engineering Tools: Spark Platforms vs Upsolver

Apache Spark promised a bright future for data lakes. Has it lived up to expectations?

Since the middle of the last decade, Apache Spark has become the de-facto standard for large-scale distributed data processing. This open-source framework leveraged in-memory MapReduce and promised to simplify and accelerate data science projects, which had previously been mired by the difficulties of the Hadoop framework from which Spark evolved. In a 2016 Stack Overflow survey, Spark was rated as the highest paying and 2nd most trending technology, indicating the promise that the development community saw in it.

However, Apache Spark is now over a decade old and no longer the new kid on the block, but an established part of the big data development toolkit – either in its original open-source format or in the various proprietary tools it has progenated – including Databricks, Elastic MapReduce, and Google Cloud’s Dataproc. 

Has Spark lived up to its promise? Is it delivering value to organizations in the age of cloud computing? Does it serve data engineers in building data lake infrastructure, as well as data scientists using it for ad-hoc data exploration?

In our new ebook, we highlight the ways in which Spark and Spark-based managed services fit into the modern data engineering landscape, and see how they compare to Upsolver, a cloud-native data processing solution built for data lakes. You can get the full ebook in the link above, or continue reading the excerpt below.

Data processing: a common roadblock

While the advantages of data lakes make them indispensable in a modern big data architecture, many organizations struggle to realize value from data lake projects, and instead find themselves mired in lengthy and complex implementations. The reason for this is that while data lakes make data very easy to store, when it comes to operationalizing the data for analysis, things get tricky.

Data needs to be processed before it can be put to use

Unlike a data warehouse, where data is preprocessed and structured before it is written into database tables, data lake storage consists of objects – often millions of them – stored in folder-like structure in a cloud repository such as Amazon S3. Data will often originate from multiple sources and land in structured, semi-structured and unstructured formats, and even within the same source schema and fields change over time. Data being ‘poured’ into the lake in this fashion often leads to the infamous problem of data swamps – with the lake becoming a practically unusable mess of unstructured, uncataloged and unidentifiable data.

Eventually, we do not wish to store data merely for the sake of storing it; data is meant to serve the business through some kind of production use case – analytics, machine learning or powering data-driven applications. Supporting these varied use cases means transforming the raw data into structured and optimized datasets that can be made available  to different consumers. This is where data processing tools come in.

Why we need to process the data

There are multiple data processing challenges that need to be addressed in order to make data lake storage analytics-ready:

  • Optimizing the object store for querying: Query engines struggle to handle large volumes of small files. This problem becomes acute when dealing with streaming sources such as application logs or IoT devices, which can generate thousands of event logs per second, each stored in a separate JSON, XML or CSV file.These files need to be converted into a file system which is optimized for fast analytical queries, built on columnar formats such as Apache Parquet or ORC.
  • Perform stateful transformations: Since we want the data to be queryable from the lake by different consumers, we would like to be able to unify data from multiple sources. In order to do so, compute-intensive operations such as joins and aggregations need to be performed in-memory by a separate compute engine. 
  • Manage the metadata layer: Metadata is stored in a separate catalog service such as a Apache Hive Metastore or AWS Glue Catalog (a managed version of a Hive metastore), which needs to be properly partitioned and kept in-sync with the physical file locations on disk. 
  • Perform Upserts: Performing operations like upserts (inserts and updates)  and deletes on a data lake file system, which isn’t indexed and is built for append-only use cases, will often require separate data engineering effort using tools such as Apache Hudi or Iceberg.

The open source approach – Spark and Managed Spark Platforms

Spark, first introduced in 2009 and released under the open-source Apache license 2013, offered a modern alternative to Hadoop MapReduce. Spark offers a flexible real-time compute engine that supports complex transformations, and its relative popularity ensures there is a large open source community that continues to support it.  Due to its performance and support for popular programming languages like Java, Scala and Python, Spark has been a popular environment for creating data lake processing jobs. It has also grown thanks to additions such as Spark Streaming.

Why many companies struggle with Spark

Many organizations struggle to see get to production with Spark as it has a very high technical entry barrier and requires extensive dedicated engineering resources. 

Spark inherited the complexity of Hadoop and requires specialized engineers that understand distributed systems, coding in Scala/Java, workflow orchestration and analytics best practices. This combination of skills isn’t immediately available to most engineering teams – and when it is, it’s often better utilized on higher-value work such as application development or machine learning.  

There is no real self-service version of Spark – Spark jobs are written in  low-level code (Scala/Java) and there are a lot of “levers” to pull in order to tune the data lake for performance. Programming transformations, state management and file system management is time intensive, prone to errors and beyond the technical scope of 95% of traditional database practitioners.

Cloud managed Spark

While Spark’s roots are rooted firmly in open-source and on-premises development, the move towards cloud computing sprung several proprietary products that built on Spark’s core features while adding security, reliability and managed cloud infrastructure. Notable products in this space include Databricks, Amazon EMR, and Google Dataproc. These tools add managed scale using the elasticity of the cloud, but do not address the limitations noted above.

ETL tools won’t save the day 

Data warehouses have a rich ecosystem of ETL and ELT tools that help ingest application data and orchestrate workflows. Some popular choices include  DBT open source, FiveTran, and Matillion. These tools can be used in 2 ways: 1) read data from sources, perform minimal transformations and load into tables (Extract-Transform-Load) or 2) run  transformations jobs on the database compute engine after landing the data (Extract-Load-Transform). 

However, this type of ETL flow relies on having a database to do the heavy lifting. If we want to keep our data in usable formats on the lake, avoid vendor lock-in and support the open architecture that the data lake is intended for, we need to transform and integrate data without relying on database compute resources. 

The Upsolver approach

Upsolver was founded by database engineers who wanted to build a modern cloud data lake without the complexity of Spark and Hadoop. Upsolver takes the traditional database approach (configuration, not code) and applies it to data lake processing by applying the following principles:

  • Give end users access to 100% of their data: Upsolver’s data processing is governed by a visual, SQL-based interface which can be used by all data practitioners including DBAs, data architects, analysts, data scientists, product managers and big data engineers. 
  • Hide database complexity to allow for focus on the data: users want to focus on writing business logic (transformations); the subsequent engineering work is abstracted away and automated. Data consistency (no loss, no duplications), ETL jobs orchestration, ETL state management and file system management are all handled automatically according to best practices. 
  • Have it your way: no-code, low-code or high-code : Upsolver’s data transformation language is based on SQL and is extensible with Python. Any use case that can be expressed with SQL, can be executed on Upsolver so the user doesn’t need to compromise between flexibility and ease-of-use. 
  • Elastic scaling: Upsolver never stores data on local server storage so processing can elastically scale according to the workload. The infinite throughput of cloud object storage is key to ensuring this.
  • No infrastructure to manage: Upsolver runs on fully managed clusters (dedicated per customer) that auto-scale on-demand. There is no need to backup Upsolver instances since all data is stored on cloud storage. If an instance crashes, a new one will spin up and continue the processing work. 
  • Streaming first: Upsolver processes all data as a stream, ensuring extremely low latency even when working with data that hits all the ‘V’s of big data – volume, velocity, variety, and veracity.
  • Low cost: Upsolver utilizes cloud object storage and low cost compute instances for processing, creating the lowest possible footprint of cloud provider costs. 

Want to learn more? Download our full ebook on Comparing Data Lake Engineering Tools to see feature-by-feature comparisons and find out more about the benefits of cloud data lakes and how you can be successful with them. Get the ebook now, free.

Try SQLake for free (early access). SQLake is Upsolver’s newest offering. It lets you build and run reliable data pipelines on streaming and batch data via an all-SQL experience. Try it for free. No credit card required.

Spark Data Processing – Frequently Asked Questions

Is Spark a processing engine?

Apache Spark is an engine for large-scale data processing. Spark can wrangle datasets that are typically too expensive and large to transform using relational databases. There are other open and closed-source options, each with its own advantages and disadvantages.

Is Apache Spark for real-time processing?

While Spark is faster than its predecessor MapReduce, it is still essentially a batch processing tool. Spark Streaming can be used for stream processing, as well as other tools including Upsolver.

Which type of data processing can Apache Spark handle?

A highly optimized implementation of Spark can handle batch processing, machine learning, streaming, and SQL workloads. However, achieving performance at scale will require processing jobs to be written in Scala, as well as continuously monitored and optimized.

Is Spark better than Hadoop?

Spark is based on Hadoop but is more popular than MapReduce because it is seen as easier to use. However, Spark still requires a lot of technical expertise and is not easily accessible to most engineers. Cloud-based managed services are available that make it easier to use Spark, but they still require a lot of technical expertise.

Published in: Blog , Cloud Architecture
Upsolver Team
Upsolver Team

Upsolver enables any data engineer to build continuous SQL data pipelines for cloud data lake. Our team of expert solution architects is always available to chat about your next data project. Get in touch

Keep up with the latest cloud best practices and industry trends

Get weekly insights from the technical experts at Upsolver.



All Templates

Explore our expert-made templates & start with the right one for you.