Writing
Notes on reliability, cloud cost and AWS by Abhinav C.
- Your container image is a product you ship to yourselfEvery image carries an operating system, a runtime and hundreds of packages you did not choose. In 2026 regulators and customers are starting to ask what is inside. Be able to answer.
- S3 quietly became a coordination serviceStrong consistency, then conditional writes, then managed tables. S3 can now do jobs that used to need a database or a lock table. Most architectures have not caught up.
- Savings Plans are bonds, not couponsA commitment is a fixed-income position on your own future usage. Finance teams have managed that kind of risk for a century. Borrow their playbook: ladder it.
- Ansible and Chef moved upstream. Most teams never noticed.Configuration management used to fix running servers. Now its best job is building the image, once, before anything runs. What changed, and where in-place config still earns its keep.
- An alert is a contract with a sleeping personEvery page promises that something is wrong, that it matters now, and that a human can fix it. Break any of those and on-call stops trusting the pager, AI triage or not.
- EMR clusters should die youngA long-running EMR cluster is a pet with a very large food bill. Treat clusters as disposable compute for a single job, and S3 as the only place data lives.
- The GPU line item is the new EC2 line itemAI workloads moved the biggest number on the bill. The playbook that tamed EC2 still applies, with two new metrics: GPU utilization and cost per request.
- NodeMinus: put the safety check where the delete happensA small desktop app that deletes node_modules folders. The UI asks for confirmation, but the real guardrail lives in the Rust backend, where no frontend bug can skip it.