Four things, each with the work that backs it.
Kubernetes platform buildout
Multi-environment clusters with HA control planes, a virtual IP that survives node loss, ingress, and a documented rebuild path. Not a demo cluster: dev, staging and production, each isolated and each proven with a real workload.
Control-plane failover tested by killing the VIP holder: traffic moved in about 3 seconds. etcd fsync p99 measured at 8.0 to 13.6 ms against a 25 ms budget.
Monitoring that catches real failures
Most monitoring proves a process is running. That is the weakest possible signal. I build output checks: did the pipeline actually produce anything, did the backup actually contain data, did the certificate actually renew.
97 signals across 20 hosts. It caught a certificate 6 hours from expiry, a renewal timer that restarted a service 67 times in a day, and a trace pipeline that had silently ingested nothing for two days.
Backups you have actually restored
A backup that has never been restored is a hypothesis. I set up the backup, then break something on purpose and restore it, then automate a verifier that fails loudly when a restore would not work.
The first verifier here printed a 0-byte warning and RESTORE VERIFIED in the same run. It computed the right answer and ignored it. That is the bug class I look for.
CI and deploy pipelines
Pipelines that fail for the right reason. I prove a check works by making it go red on purpose before I trust it, because a check that has never failed is decoration.
On this site's own pipeline: lint, types, build, plus checks for stale figures, dead links, and unrendered pages. Each one red-run against a deliberately broken tree before it shipped.