The Serverless Lesson I Would Write On A Sticky Note

I have shipped enough serverless work in the last year that patterns are clear. There is one rule that would fit on a sticky note and would have saved me most of the pain.

The business problem is cost and reliability. Serverless promises pay-per-use and auto scale. Those promises come with conditions. Ignore the conditions and the bill surprises you.

The rule

If the workload runs for more than five minutes at a time or costs more than one dollar per run, do not use Lambda. Use a container or a small VM.

Why this rule

Lambda and similar functions are optimized for short bursty workloads. Long running work hits timeouts, warm up latency, and surprisingly high per-invocation overhead. The cost per compute hour of a busy Lambda is often worse than a small always-on VM.

The math I actually ran

A client had a daily report generator that ran for eight minutes on Lambda. The monthly bill was $180. Moved it to a t3.small EC2 instance that runs the same script on a cron schedule. Monthly bill: $18.

When serverless actually wins

• Webhook handlers that run for a second at a time
• Low volume APIs where a VM would sit idle 90 percent of the time
• Glue between AWS services where IAM is already set up

When I skip it

• Any job that takes more than a minute
• Anything with heavy cold start penalties (large ML libraries)
• Anything that needs a long lived database connection

The other lesson

DynamoDB single table is great for patterns you know in advance. The day you need ad hoc analytics, a relational database with SQL will save you a month. Plan for both. See the AWS Lambda pricing page and run the math before you commit. I should have a year ago.