Debugging Azure Networking for Elastic Cloud Serverless

Summary

Summary of Findings Elastic's Site Reliability Engineering team (SRE) observed unstable throughput and packet loss in Elastic Cloud Serverless running on Azure Kubernetes Service (AKS). After investigation, we identified the primary contributing factors to be RX ring buffer overflows and kernel input queue saturation on SR-IOV interfaces. To address this, we increased RX buffer sizes and adjusted the netdev backlog, which significantly improved network stability. Setting the Scene Elastic Cloud Serverless is a fully managed solution that allows you to deploy and use Elastic for your use cases without managing the underlying infrastructure. Built on Kubernetes, it represents a shift in how you interact with Elasticsearch. Instead of managing clusters, nodes, data tiers, and scaling, you create serverless projects that are fully managed and automatically scaled by Elastic. This abstraction of infrastructure decisions allows you to focus solely on gaining value and insight from your data. Elastic Cloud Serverless is generally available (GA) on AWS, GCP and currently in Technical Preview on Azure. As part of preparing Elastic Cloud Serverless GA on Azure, we have been conducting extensive performance and scalability tests to ensure that our users get a consistent and reliable user experience. In this post, we’ll take you behind the scenes of a deep technical investigation into a surprising performance issue that affected Serverless Elasticsearch in our Azure Kubernetes clusters. At first, the network seemed like the least likely place to look, especially with a high-speed 100 Gb/s interface on the host backing it. But as we dug deeper, with help from the Microsoft Azure team, that’s exactly where the problem led us. Unexpected Results! While the high-level architectures and system design patterns of the major cloud provider’s systems are often similar, the implementations are different, and these differences can have dramatic impacts on a system’s performance characteri...