Revision · Cloud & Scaling

This part treated the cloud as renting someone else’s data center behind an API, and worked through how to use that well: what the cloud is, how to scale it, and how to keep the bill and a dead region from hurting you.

What this part covered

• Capacity is an API call — provisioning went from a multi-week ticket to a POST that returns a machine in seconds, which is what makes IaC, autoscaling, and self-healing even conceivable.
• Cloud models (IaaS/PaaS/SaaS) — rungs on one abstraction ladder where the shared-responsibility line slides upward, trading control for convenience; climb as high as a workload tolerates and drop down only when a hard requirement forces it.
• Core primitives — every cloud reduces to compute, storage (block/object/file), networking (VPC, subnets, security groups), and IAM as the spine — with IAM misconfiguration the new dominant failure mode.
• Capex to opex — you stop over-buying for a peak and pay by the hour, which is liberating but dangerous because a metered bill scales with carelessness as well as with usage.
• Autoscaling — a reconciliation loop on load metrics that scales horizontally (more stateless instances), reactively or on a schedule, with scale-to-zero paying nothing while idle at the cost of cold-start latency.
• Managed services — buying away the “undifferentiated heavy lifting” of backups, failover, and patching, weighed against lock-in, lost control, and steep data-egress fees; the honest default is to buy, especially databases.
• Cost and FinOps — cost as an engineering feedback loop (tag, report unit economics, optimize, govern), layering reserved, on-demand, and spot purchasing to roughly halve a naive on-demand bill.
• Multi-region and DR — RTO and RPO define recovery, the DR ladder trades cost for speed (backup-and-restore → pilot-light → warm-standby → active-active), and synchronous versus asynchronous replication is a hard consistency-versus-latency choice.
• DR must be rehearsed — an untested DR plan is a hope, not a plan; rebuildable infrastructure-as-code is what lets you practice recovery on a calm Tuesday.

The takeaway

The cloud removes owning, racking, powering, and capacity-planning hardware by hand, and returns elasticity, reproducibility, and re-derivable recovery — but only if you tame it. A console full of click-created resources is just ClickOps with someone else’s hardware, so everything here assumes cloud resources live in a repo, which is why this part and the IaC part are two halves of the same discipline.