AWS Lambda

What It Is and Why It's Terrible

A presentation & live demo highlighting the the dark underbelly of "Serverless" solutions, picking on AWS Lambda because it is the most prevelant. We build an old-school HitCounter in Lambda + DynamoDB and compare this approach to CGI + SQLite.

Outline

• High Latency when spinning up new instances.
  • New "provisioned concurrency" model is AWS billing double-speak, why not buy a VM?
• Debugging is a nightmare:
  • Logs don't come through CloudWatch immediately
  • No ptrace or bpftrace for production debugging
  • No atop for understanding resource utilization.
• Weird packaging & deployment pattern that isn't used anywhere else.
  • In practice you must use a framework like Serverless or Zappa
  • They handle some but not all infra for you -- where should the line be drawn?
• Is "patching servers" really that hard? Yumcron, anyone?
• Providing secrets means using + paying for AWS Secrets manager
  • No secrets manager means keeping secrets in plaintext somewhere
  • Avoid secrets means Fully Relying on IAM, which can be easy to screw up
• Network performance is proportional to memory allocation
  • This makes you pay for way more than you need in order to get a responsive app
  • Speculation: this comes from oversubscribing RAM via KSM, can't do same with network
• Lambda saves money when apps are mostly off, but the developer time will never pay off.

Acknowledgements

• Abe Simpson image © 20th Century Fox
• Modern Problems meme © Dave Chappelle / Comedy Central
• Critical Feedback from @myoung34