AWS Lambda Web Adapter

Cost model and performance pitfalls

• Lambda bills for duration, so blocking on slow or hung downstream APIs can explode costs compared to always‑on servers.
• Timeouts are essential to cap cost, but aggressive timeouts plus retries can backfire and increase total work and spend.
• Cloudflare Workers are contrasted as billing on CPU time, not wall time, making network waits cheaper there.
• Thread notes Lambda’s default (3s) and practical limits on timeouts; some argue you want short timeouts to avoid “blank checks” to AWS.

When Lambda fits vs. when it doesn’t

• Good fit: low‑TPS internal apps, spiky or infrequent workloads, job/queue processing, webhooks, “fire sale” / telemetry‑like bursts, and pet projects where scaling to zero matters.
• Poor fit: consistently high load, long‑running or IO‑heavy calls, high DB connection counts, or where development iteration speed and observability are critical.
• Several commenters say ECS/Fargate or traditional containers solve many problems more cleanly once you have steady traffic.

Lambda Web Adapter and reuse of containers

• Adapter enables running the same HTTP container on Lambda, ECS/Fargate, or EC2, reducing duplicate codebases and easing lift‑and‑shift from one compute option to another.
• Some see this as improving security/compliance by reusing a single well‑known container; others worry it encourages keeping legacy systems alive instead of fixing underlying issues.
• One pattern: deploy the same image as both Lambda and Fargate service behind an ALB, then direct traffic to whichever is more cost‑effective for that workload.

Architecture patterns: async, APIs, and pooling

• Async/claim‑check patterns, webhooks, Step Functions, and queues (e.g., SQS) are suggested to avoid Lambdas waiting on long operations.
• Direct Lambda‑to‑Lambda blocking calls are viewed as an antipattern; event/queue‑based designs are preferred for cost and resilience.
• Connection pooling works per‑Lambda instance lifetime; for databases like Postgres, unbounded concurrency leads to too many connections without a proxy (e.g., RDS Proxy).

Developer experience and “programming the machine”

• Opinions diverge on iteration speed: some find Lambda cycles acceptable, others find build/deploy/debug painfully slow compared to local dev + containers/Kubernetes.
• Many note significant time spent on IaC and wiring AWS services (“programming the machine”) that can rival or exceed application coding.
• Scaling to zero is valued for per‑branch ephemeral environments and previews, but some argue the complexity often outweighs small cost savings.