We Live in Memory: Making Λ and GraphQL Respond in 70MS
summary
Claims a production GraphQL API runs at 70ms average (most requests ~24ms), cuts ops cost from ~$30k/month to ~$3k, and stays stable under spikes. Implementation uses serverless GraphQL on Lambda, Fastly CDN edge caching, and Redis as primary datastore with an append-only log and RDB snapshots. A custom graph layer (Gradius) and in-progress Nemesis DB power labeled-edge lookups, batching and pipelining, and the stack demonstrably works in production.
AI-written summary · about
chapters (12)
- 0:00Intro
- 0:43We Live in Memory
- 2:46Goals
- 5:42Functions
- 7:30API
- 8:32Database
- 15:13Nemesis DB
- 16:53GraphQL Server
- 21:01GraphQL Resolvers
- 22:51Data Loader
- 25:07Lambda Deploy Tools
- 26:30Redis
related talks
34mGoing Serverless with GraphQL
26mWhen GraphQL Gets Expensive: Performance & Cost Patterns in Production Serverless Architectures
29mDeploying GraphQL in a Serverless Context
17mBuild cost effective Asynchronous GraphQL APIs using Serverless.
41mUsing Serverless & ARM64 for Real-Time Observability
36mEfficient and Performant GraphQL Serverless APIs
23mServer-less architecture in the wild
39mArchitecture of scalable and resilient NodeJS apps with GraphQL & event-driven serverless