How to Reduce the Cost and Increase the Speed of a Serverless IoT Environment

Based in Istanbul, Turkey, Arçelik A.S. offers home appliances and related services in more than 130 countries. Research and development are a major part of the group’s strategy, and the company has begun developing appliances that connect to the Internet of Things (IoT).

By understanding how customers use home appliances, Arçelik can make better decisions about product development. And if after-sales teams in the field know what might be wrong with an appliance by applying early diagnostics, they could make the repair in one visit. In some cases, they could even resolve the issue remotely, saving time and money.

To process the 5 GB of data that 32,000 IoT appliances generate each day, Arçelik built a serverless environment on Amazon Web Services (AWS). The company had previously built similar architecture to monitor its manufacturing facilities and knew that Amazon Kinesis Data Streams could receive the data and stream it to an Amazon Simple Storage Service (Amazon S3) backup bucket in less than a second.

Software specialist Onur Yılmaz and his team built AWS Lambda functions in isolated environments, which help the company manage the events fired by many types of appliances in a single data pipeline. “AWS Lambda removes the complexity of dealing with servers and saves us money because we don’t have to pay for idle compute time," says Yılmaz. "The isolation we’ve built in helps us scale on the fly and improves fault tolerance.” 

The Arçelik team found it difficult to get full visibility into this complex environment using Amazon CloudWatch logs alone. At a serverless meetup in Ankara, Yılmaz met Emrah Şamdan, VP of products at Thundra, an APN Advanced Technology and AWS DevOps Competency Partner in the AWS Partner Network (APN).

Şamdan explains, “It can be hard to understand why a certain Lambda function is running longer than expected for different kind of events. That’s where Thundra helps. It lets you identify bottlenecks by visualizing serverless transactions with distributed and local tracing combined. And removing bottlenecks helps reduce the running time of Lambda functions.”

Arçelik’s AI and Analytics Team Lead Arda Taşcı says “It was super easy to integrate Thundra into our serverless data pipeline with Lambda Layers and start observing our functions.” Immediately, the team found examples of Lambda cold starts—invocation latencies when functions had not been used for a long period—and bottlenecks where functions were accessing the database.

“Thundra helped us reduce AWS Lambda processing time—and corresponding cost—by around 50 percent,” says Yılmaz. “And we get insight that helps us choose the best timeout values and memory thresholds for Lambda functions.” In addition, the improved observability Thundra provides continues to be valuable as Arçelik develops its pipeline to handle new data types from new IoT products.