Building Efficient Systems: From Protosockets to Global Tables

​Talk 1: When HTTP is too heavy: Designing a minimal RPC layer with protosocket at Momento

​Modern systems often default to HTTP-based RPC frameworks like gRPC—but at high message rates, those abstractions can become the bottleneck. At Momento, Pratik and team built Protosocket, a lightweight, message-oriented layer over raw TCP that avoids the overhead of HTTP, gRPC, and WebSockets. By rethinking synchronization and message flow, they achieved 3× throughput and reduced server count by 65%. This talk explores how stripping away layers of abstraction can unlock surprising efficiency gains in production systems.

​About Speaker
Pratik is a distributed systems engineer at Figma, where he architects and scales database infrastructure that powers millions of designers and developers worldwide. With deep expertise spanning the entire backend stack - from client SDKs to storage systems - he has consistently delivered transformational performance improvements and cost optimizations across his career. Previously, Pratik worked at Momento on caching and high performance, databases at DynamoDB, and event processing at AWS Marketplace.

​
Talk 2: Evolving DynamoDB’s Global Tables for Multi-Region Scale

​Geo-replication is one of the hardest problems in distributed systems — balancing latency, availability, and consistency across continents is never easy. In this talk, Somu will dive deep into the evolution of DynamoDB’s Global Tables, exploring the design principles that make it possible to replicate data across regions while maintaining performance and reliability at scale.

​We’ll contrast eventually consistent regional DynamoDB with strongly consistent Global Tables, unpacking how replication conflicts are detected and resolved, what guarantees customers can expect, and the trade-offs behind design decisions. You’ll walk away with a clear understanding of how large-scale distributed databases think about consistency models, fault domains, and correctness across time and geography — lessons that apply far beyond DynamoDB itself.

​Whether you’re designing your own replicated data store, building multi-region applications, or simply curious how the world’s largest managed NoSQL database achieves global durability, this session will give you a front-row view into what it takes to build systems that never sleep.

About Speaker
Somu Perianayagam is a software engineer at Amazon Web Services (AWS), where he has worked on multiple serverless products including Amazon DynamoDB and Amazon Bedrock. At DynamoDB, he led core efforts in the replication engines that power both regional and global tables — the backbone of DynamoDB’s high-availability and geo-redundant architecture. Currently at Bedrock, AWS’s serverless inference platforms, he focuses on solving hard systems problems in admission control, scheduling, and inference serving for generative AI workloads. When not debugging replication races or optimizing inference paths, he enjoys spending time with his kids and bikes.