The paved path through production, not the vendor’s front page.
Every cloud decision has a wilds version — improvised, under-scoped, discovered the hard way in an incident review — and a paved version, already vetted by people who operated it at scale. We write the paved version. No cert-prep, no vendor spin, no “it depends” left unresolved.
The capacity-mode decision you can't cleanly reverse once a fleet of read replicas depends on it.
The wilds
- A blog post from 2019, a re:Invent talk, and whatever your team remembers from the last outage.
- Generic best-practice lists that read the same for a 10-person startup and a bank.
- Vendor docs optimized to sell the service, not to tell you where it breaks.
The paved path
- The one irreversible decision, named up front, before you commit to it.
- The exact metric and threshold where the service's failure mode shows up in production.
- The non-obvious guardrail that would have prevented the incident you haven't had yet.
No. 2Tracks
One method, run separately per providerAWS
Aurora, DynamoDB, EKS, ECS and the rest of the database & compute Tier 1 set.
Azure
Cosmos DB, AKS and the SQL/Postgres managed services.
GCP
AlloyDB, Spanner, GKE.
OCI
Exadata, OKE, Autonomous Database.
Terraform
State, modules and the provider-agnostic layer.
No. 3Same decision, four clouds
Some decisions repeat across every provider under a different name. We line them up side by side so the trade-off is visible instead of buried in four separate docs sets.