Staged Rollout

Staged Rollout

Overview

A change reaches all users gradually, behind a switch you can flip back, while someone watches the signals. Big-bang releases convert a small bug into a full outage. A staged rollout converts the same bug into a contained blip caught at 1% traffic.

Core principle: Every outward-facing change ships behind a flag, advances in stages with a monitoring window between them, and has a rollback plan written before launch — not improvised during the incident.

This depends on the change being observable first: see [[instrument-observability]] for the signals the monitoring window watches.

When to Use

• Any user-facing or externally-observable change (new endpoint, UI, behavior change, schema migration)
• Risky internal changes where a bad deploy degrades many users at once

Skip only for trivially reversible, low-blast-radius changes (a copy fix, an internal doc). When unsure, stage it — the cost is one flag.

Process

1. Pre-launch checklist. Confirm before any traffic shift:

• Validation and review already passed (see [[validate]] / [[review-pr]])
• Instrumentation is live and the dashboards read real data
• The feature flag exists and defaults to off
• The rollback plan is written (step 4)

1. Feature-flag gating. Put the new behavior behind a flag, default off. The flag must be flippable at runtime without a redeploy — that is what makes rollback fast.
2. Staged / canary rollout. Advance through stages, not in one jump. A typical ramp:

| Stage | Audience | Hold for | |-------|----------|----------| | Canary | 1% (or internal/dogfood) | a monitoring window | | Early | 10% | a monitoring window | | Majority | 50% | a monitoring window | | Full | 100% | — |

Monitoring window: between each stage, watch the RED metrics and alerts long enough to span real traffic before promoting. Do not promote on a clean dashboard you've watched for thirty seconds. If a symptom breaches, stop and roll back — do not push forward hoping it settles.

1. Written rollback plan. Before launch, document: the exact trigger conditions (which metric/alert at which threshold), the precise rollback action (flip flag X to off; revert migration Y), who can execute it, and the expected recovery time. A rollback that requires a redeploy is too slow — prefer the flag.

Rationalizations

| Excuse | Reality | |--------|---------| | "It's a small change, ship to everyone" | Small changes cause large outages. The flag costs minutes; the outage costs hours. | | "I'll write the rollback plan if something breaks" | Mid-incident is the worst time to design recovery. Write it cold, before launch. | | "Dashboard looked clean, I promoted right away" | A clean dashboard watched for seconds proves nothing. Hold the window across real traffic. | | "Rollback is just redeploy the old version" | A redeploy is minutes of continued damage. A flag flip is seconds. Gate behind a flag. | | "Canary at 1% will take forever" | 1% surfaces the catastrophic bugs cheaply. That's the point. |

Red Flags

• Going straight to 100% with no canary stage
• A flag that needs a redeploy to toggle
• Promoting between stages without watching the metrics in between
• No documented trigger condition or named rollback action before launch
• "We'll roll it back by reverting the commit" as the entire plan

Verification

• [ ] Pre-launch checklist complete; flag exists and defaults to off
• [ ] The change is gated behind a runtime-flippable feature flag
• [ ] A staged/canary ramp is defined with a monitoring window between stages
• [ ] A written rollback plan exists naming trigger conditions, the exact rollback action, the owner, and expected recovery time
• [ ] The plan was written before launch, not after the first stage

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: etr
• Source: etr/groundwork
• License: MIT

Install and usage instructions live in the source repository linked above.