DevOps · Cluster
Deployment Strategies
Blue-green, canary, and progressive rollout.
- 01 Blue-Green Cutover Mechanics at the Routing Layer Where the traffic switch actually happens, why DNS record changes make poor cutover levers, and what the standby fleet must be able to carry before and after the switch.
- 02 Connection Draining and In-Flight Request Handling at Cutover How an instance retires without dropping requests: deregistration order, keep-alive reuse, TCP and TLS teardown, and the timers that decide whether clients see errors.
- 03 Weighted Canary Routing at the Edge and What the Weight Measures Whether a traffic weight applies to connections, requests, or users, how hashing turns percentages into fixed partitions, and why small canaries cannot resolve small error differences.
- 04 Health Signals for Gating a Progressive Rollout Why readiness probes cannot decide whether a new version is good, which signals can, and how probe timing and baseline choice set the number of bad requests before an abort.
- 05 Expand and Contract: Schema Changes That Survive an Overlap The three-phase pattern that keeps two application versions reading one schema, which DDL operations take heavy locks, and how to bound the lock window with batched backfills.
- 06 Session Affinity, Sticky Routing, and Version Overlap How affinity keeps a client on one instance, why it biases a canary sample, and what has to be true of the session payload and its keys for a pinned session to survive promotion.
- 07 Feature Flags Versus Traffic Routing as Rollout Controls Why a flag and a routing weight expose different populations, what it costs to keep both code paths in one binary, and how evaluation consistency and stale defaults cause failures.
- 08 Decision Rules for Aborting or Promoting a Rollout Why repeatedly checking a canary against a fixed significance threshold inflates false aborts, how sequential methods and error-budget burn rates avoid that, and what sample sizes the arithmetic demands.
- 09 Progressive Delivery Controllers as Reconciliation Loops How a rollout becomes a cluster object with a status field, how the controller moves traffic through a provider abstraction, and where the loop conflicts with autoscalers and GitOps sync.
- 10 Rollback Boundaries: Which Layers a Reversal Can Restore Code, configuration, schema, data, messages, and external effects each reverse differently; how to place the irreversible step last and keep a rollback window that is real.