Hey everyone! I'm so excited to finally share my experience from the last three months. I just finished migrating a legacy monolithic reporting application (Java/Spring, PostgreSQL) to a fully serverless stack on AWS. I wanted to give back to the community since I learned so much from posts here.
Here’s my breakdown of the good, bad, and the frankly ugly parts.
**The Good Stuff 😊**
* **Costs Plummeted:** Our non-peak hour and dev environment costs dropped by about 70%. Not having to pay for idle EC2 instances was a huge win.
* **Development Velocity:** Once we got the patterns down, spinning up new API endpoints (API Gateway + Lambda) and ETL jobs (Glue) felt incredibly fast. I loved how infrastructure became part of the code.
* **Observability:** With everything instrumented from the start, CloudWatch and X-Ray gave us amazing visibility into our data pipelines and API calls. Debugging *some* things became easier.
**The Bad & The Ugly 😅**
* **Cold Starts Are Real:** Our user-facing dashboard backed by Lambda would sometimes have a 3-4 second load on the first hit after inactivity. It really hurt the user experience until we added provisioned concurrency (which ate into our cost savings).
* **Distributed Complexity:** Tracing a request through API Gateway → 10+ Lambda functions → Step Functions → SQS → more Lambdas became a headache. The mental model shifted from a single app to a web of tiny parts.
* **Vendor Lock-In Fears:** We're now deeply tied to AWS-specific services (Step Functions, Glue, DynamoDB). I'm starting to wonder if using more "generic" containers (like Fargate) for some core parts would have been smarter.
My biggest question for you all who've done this:
* **For those who built self-serve analytics or reporting tools on serverless, how did you handle the cold start problem for your end-users?**
* **Any recommendations on ETL tools or patterns that play nicely with this architecture? I’ve been looking at dbt but am unsure how it fits.**
I'm still really glad we did it—the learning curve was steep but worth it! Would love to hear your stories and advice.
Provisioned concurrency helps but the cost model flips when you need high, consistent concurrency. I've seen Lambda bills spike 300% compared to a few reserved EC2 instances for always-on workloads.
What's your average request volume and p99 latency after the fix? Real numbers matter more than anecdotes.
Numbers don't lie.