gRPC Conf 2025

Netflix scales gRPC services by reducing call volume through caching and batching, applied through client and server interceptors and automatically wired into services. Over half of all Java-client gRPC traffic is served through on-box (Google/Guava, OHCache) and distributed (Netflix/EVcache) caches. We present how integrating caching directives directly into the proto specification allow us to automatically inject cache support into the stubs with minimal developer overhead. We also present our request batching technique to aggregate multiple stub invocations into a single remote call. This allows service owners to reduce call volume even when consuming applications apply suboptimal fetch patterns.

Subsetting is a well established technique that helps scale a distributed system, once the management overhead of a fully connected mesh is no longer viable. At Datadog, with hundreds of thousands of processes connected with gRPC, we have been suffering from this overhead for quite some time, and around one year ago we decided to do something about it. As, currently, there isn’t a native way in gRPC to do subsetting, we decided to start extending gRPC in-house to fit our case. This presentation is going to be a story about the many different ways we attempted to deal with subsetting in gRPC and lessons we learned. We will cover why subsetting is beneficial, the different implementations we tried, ways to eliminate imbalance generated by subsetting and how subsetting is helping us to use smart load balancing algorithms to manage overhead and drive reliability. We will finish by providing an update around our efforts to upstream these changes following the gRPC RFC process.

This talk covers why many of Swift's features make it a great fit for the server before diving into the gRPC Swift library. We'll review the current state of the library and our journey towards gRPC Swift v2. We'll describe how the new version takes advantage of some of Swift's latest features, such as its concurrency model, and compile-time data race checking. You'll also see how we leveraged Swift to build simple and expressive new APIs, as well as how easy it is to get started with gRPC Swift on both the server and client.

AI is changing how we develop, the workloads we run, and the hardware we use. Development tools are helping us more than ever. Model training is an intense batch workload. Come learn how AI aids using gRPC, how gRPC is used with these stressful workloads, and how gRPC itself is adapting.

In this session, we will cover the following: - Basics of client-side load balancing support in gRPC: - Here, we will cover the interaction between the gRPC channel and the load balancing policy. We will also go over the load balancing API in gRPC that allows our users to implement their own policies. - Overview of supported load balancing policies: - gRPC ships with a bunch of load balancing policy implementations. We will go through a subset of them and talk about ideal deployment scenarios for each of them. - Custom load balancing policy support in gRPC: - gRPC supports configuration of custom load balancing policies on the client by the control plane.

Datadog has been using gRPC for many years. Hundreds of engineers are using it on a daily basis, and hundreds of millions of requests flow through our system across hundreds of thousands of processes every second to serve the product. Despite this scale, our setup has remained very basic: DNS for service discovery, and all tuning and customizations embedded directly in client and server code. We recently adapted the xDS control plane that configures all our standalone load balancers to support direct proxyless gRPC integration. In this session, we will describe the gains users can expect from using xDS for service discovery instead of DNS: faster endpoint propagation time, centralization of service configuration, and access to more advanced load balancing and traffic management features. We will describe the effort that went into it, and how it is playing out in practice. By the end of the talk, attendees will better understand what xDS can do for their gRPC setup.