ARM64 Graviton migration across 99 Lambdas

The problem

The serverless backend ran 99 Lambdas across case, identity, files, notifications, and shared infra, all on x86 at default memory. Cost was creeping up with usage and several CPU bound functions like PDF parsing, token verification, and metric calculation were hitting timeouts under load. A blanket memory bump would have helped but doubled the bill.

The solution

Audited every function for native dependencies, then rolled out two changes in the SAM Globals block per service. Switched architecture to arm64 for Graviton 3, and right sized memory per function based on actual CloudWatch duration histograms rather than guesses. Shipped service by service behind a feature canary, with a one click rollback if duration p95 regressed. Caught two functions with x86 only native libs and pinned them, left the rest on ARM. Net result was lower cost on most paths and faster execution on CPU bound ones.

Architecture

Stack

Outcomes