Writing
Notes on reliability, cloud cost and AWS by Abhinav C.
- SLOs for features that answer differently every timeLLM-backed features break the usual assumptions: slow by design, a dependency you do not run, and "success" that a status code cannot see. What to measure instead.
- On AWS, the SQL Server bill is a license billSQL Server 2016 reached end of extended support in July 2026, so a lot of databases are moving now. Before resizing anything, know that cores, not hardware, drive the cost.
- The best OS hardening is having less OSCIS checklists harden a general-purpose server one setting at a time. Minimal operating systems remove most of what those checklists were protecting. With Amazon Linux 2 now past end of support, it is a good moment to choose.
- AWS got safer by default. Your old accounts did not.New S3 buckets block public access, encrypt objects and disable ACLs out of the box. Accounts created years ago kept their old settings. A short list of what to check.
- Kubernetes requests are a purchase orderEvery CPU and memory request reserves capacity you pay for, whether the pod uses it or not. Get them right and the cluster gets smaller and more stable.
- Your Helm chart is a public API. Version it like one.The day a second team installs your chart, values.yaml becomes an interface other people depend on. Most chart breakages are API breakages nobody called that.
- Your observability bill is a cardinality billMonitoring costs rarely grow with traffic. They grow with labels, log lines nobody reads and traces kept by default. Here is where the money goes and how to get it back.
- Reserved concurrency is the best circuit breaker in AWSMost people meet Lambda concurrency as a limit to raise. Used deliberately, it is a dial that protects your database, isolates noisy functions and works as a kill switch.