All posts by Constantin Scoarță

How CyberSolutions built a scalable data pipeline using Amazon EMR Serverless and the AWS Data Lab

Post Syndicated from Constantin Scoarță original https://aws.amazon.com/blogs/big-data/how-cybersolutions-built-a-scalable-data-pipeline-using-amazon-emr-serverless-and-the-aws-data-lab/

This post is co-written by Constantin Scoarță and Horațiu Măiereanu from CyberSolutions Tech.

CyberSolutions is one of the leading ecommerce enablers in Germany. We design, implement, maintain, and optimize award-winning ecommerce platforms end to end. Our solutions are based on best-in-class software like SAP Hybris and Adobe Experience Manager, and complemented by unique services that help automate the pricing and sourcing processes.

We have built data pipelines to process, aggregate, and clean our data for our forecasting service. With the growing interest in our services, we wanted to scale our batch-based data pipeline to process more historical data on a daily basis and yet remain performant, cost-efficient, and predictable. To meet our requirements, we have been exploring the use of Amazon EMR Serverless as a potential solution.

To accelerate our initiative, we worked with the AWS Data Lab team. They offer joint engineering engagements between customers and AWS technical resources to create tangible deliverables that accelerate data and analytics initiatives. We chose to work through a Build Lab, which is a 2–5-day intensive build with a technical customer team.

In this post, we share how we engaged with the AWS Data Lab program to build a scalable and performant data pipeline using EMR Serverless.

Use case

Our forecasting and recommendation algorithm is fed with historical data, which needs to be curated, cleaned, and aggregated. Our solution was based on AWS Glue workflows orchestrating a set of AWS Glue jobs, which worked fine for our requirements. However, as our use case developed, it required more computations and bigger datasets, resulting into unpredictable performance and cost.

This pipeline performs daily extracts from our data warehouse and a few other systems, curates the data, and does some aggregations (such as daily average). Those will be consumed by our internal tools and generate recommendations accordingly. Prior to the engagement, the pipeline was processing 28 days’ worth of historical data in approximately 70 minutes. We wanted to extend that to 100 days and 365 days of data without having to extend the extraction window or factor in the resources configured.

Solution overview

While working with the Data Lab team, we decided to structure our efforts into two approaches. As a short-term improvement, we were looking into optimizing the existing pipeline based on AWS Glue extract, transform, and load (ETL) jobs, orchestrated via AWS Glue workflows. However, for the mid-term to long-term, we looked at EMR Serverless to run our forecasting data pipeline.

EMR Serverless is an option in Amazon EMR that makes it easy and cost-effective for data engineers and analysts to run petabyte-scale data analytics in the cloud. With EMR Serverless, we could run applications built using open-source frameworks such as Apache Spark (as in our case) without having to configure, manage, optimize, or secure clusters. The following factors influenced our decision to use EMR Serverless:

  • Our pipeline had minimal dependency on the AWS Glue context and its features, instead running native Apache Spark
  • EMR Serverless offers configurable drivers and workers
  • With EMR Serverless, we were able to take advantage of its cost tracking feature for applications
  • The need for managing our own Spark History Server was eliminated because EMR Serverless automatically creates a monitoring Spark UI for each job

Therefore, we planned the lab activities to be categorized as follows:

  • Improve the existing code to be more performant and scalable
  • Create an EMR Serverless application and adapt the pipeline
  • Run the entire pipeline with different date intervals

The following solution architecture depicts the high-level components we worked with during the Build Lab.

In the following sections, we dive into the lab implementation in more detail.

Improve the existing code

After examining our code decisions, we identified a step in our pipeline that consumed the most time and resources, and we decided to focus on improving it. Our target job for this optimization was the “Create Moving Average” job, which involves computing various aggregations such as averages, medians, and sums on a moving window. Initially, this step took around 4.7 minutes to process an interval of 28 days. However, running the job for larger datasets proved to be challenging – it didn’t scale well and even resulted in errors in some cases.

While reviewing our code, we focused on several areas, including checking data frames at certain steps to ensure that they contained content before proceeding. Initially, we used the count() API to achieve this, but we discovered that head() was a better alternative because it returns the first n rows only and is faster than count() for large input data. With this change, we were able to save around 15 seconds when processing 28 days’ worth of data. Additionally, we optimized our output writing by using coalesce() instead of repartition().

These changes managed to shave off some time, down to 4 minutes per run. However, we could achieve a better performance by using cache() on data frames before performing the aggregations, which materializes the data frame upon the following transformation. Additionally, we used unpersist() to free up executors’ memory after we were done with the mentioned aggregations. This led to a runtime of approximately 3.5 minutes for this job.

Following the successful code improvements, we managed to extend the data input to 100 days, 1 year, and 3 years. For this specific job, the coalesce() function wasn’t avoiding the shuffle operation and caused uneven data distribution per executor, so we switched back to repartition() for this job. By the end, we managed to get successful runs in 4.7, 12, and 57 minutes, using the same number of workers in AWS Glue (10 standard workers).

Adapt code to EMR Serverless

To observe if running the same job in EMR Serverless would yield better results, we configured an application that uses a comparable number of executors as in AWS Glue jobs. In the job configurations, we used 2 cores and 6 GB of memory for the driver and 20 executors with 4 cores and 16 GB of memory. However, we didn’t use additional ephemeral storage (by default, workers come with free 20 GB).

By the time we had the Build Lab, AWS Glue supported Apache Spark 3.1.1; however, we opted to use Spark 3.2.0 (Amazon EMR version 6.6.0) instead. Additionally, during the Build Lab, only x86_64 EMR Serverless applications were available, although it now also supports arm64-based architecture.

We adapted the code utilizing AWS Glue context to work with native Apache Spark. For instance, we needed to overwrite existing partitions and sync updates with the AWS Glue Data Catalog, especially when old partitions were replaced and new ones were added. We achieved this by setting spark.conf.set("spark.sql.sources.partitionOverwriteMode", "DYNAMIC") and using an MSCK REPAIR query to sync the relevant table. Similarly, we replaced the read and write operations to rely on Apache Spark APIs.

During the tests, we intentionally disabled the fine-grained auto scaling feature of EMR Serverless while running jobs, in order to observe how the code would perform with the same number of workers but different date intervals. We achieved that by setting spark.dynamicAllocation.enabled to disabled (the default is true).

For the same code, number of workers, and data inputs, we managed to get better performance results with EMR Serverless, which were 2.5, 2.9, 6, and 16 minutes for 28 days, 100 days, 1 year, and 3 years, respectively.

Run the entire pipeline with different date intervals

Because the code for our jobs was implemented in a modular fashion, we were able to quickly test all of them with EMR Serverless and then link them together to orchestrate the pipeline via Amazon Managed Workflows for Apache Airflow (Amazon MWAA).

Regarding performance, our previous pipeline using AWS Glue took around 70 minutes to run with our regular workload. However, our new pipeline, powered by Amazon MWAA-backed EMR Serverless, achieved similar results in approximately 60 minutes. Although this is a notable improvement, the most significant benefit was our ability to scale up to process larger amounts of data using the same number of workers. For instance, processing 1 year’s worth of data only took around 107 minutes to complete.

Conclusion and key takeaways

In this post, we outlined the approach taken by the CyberSolutions team in conjunction with the AWS Data Lab to create a high-performing and scalable demand forecasting pipeline. By using optimized Apache Spark jobs on customizable EMR Serverless workers, we were able to surpass the performance of our previous workflow. Specifically, the new setup resulted in 50–72% better performance for most jobs when processing 100 days of data, resulting in an overall cost savings of around 38%.

EMR Serverless applications’ features helped us have better control over cost. For example, we configured the pre-initialized capacity, which resulted in job start times of 1–4 seconds. And we set up the application behavior to start with the first submitted job and automatically stop after a configurable idle time.

As a next step, we are actively testing AWS Graviton2-based EMR applications, which come with more performance gains and lower cost.


About the Authors

 Constantin Scoarță is a Software Engineer at CyberSolutions Tech. He is mainly focused on building data cleaning and forecasting pipelines. In his spare time, he enjoys hiking, cycling, and skiing.

Horațiu Măiereanu is the Head of Python Development at CyberSolutions Tech. His team builds smart microservices for ecommerce retailers to help them improve and automate their workloads. In his free time, he likes hiking and traveling with his family and friends.

Ahmed Ewis is a Solutions Architect at the AWS Data Lab. He helps AWS customers design and build scalable data platforms using AWS database and analytics services. Outside of work, Ahmed enjoys playing with his child and cooking.