blue-green-deployment-pattern

What problem does it solve?

This Skill prevents risky downtime during LLM serving upgrades by using a blue/green cutover approach that runs an isolated green instance, verifies it, switches traffic, and rolls back quickly if anything looks wrong.

Core Features & Use Cases

• Zero/Low-downtime cutover workflow: bring up a green vLLM instance on a new port with isolated dependencies, perform health and smoke checks, then switch to full production settings.
• Safety guardrails for real deployments: isolated venv, flashinfer cache handling, interactive approval gate, GPU memory fencing (smoke vs full), and a rollback runbook.
• Use Case: Upgrading vLLM (e.g., 0.19 → 0.20) while keeping the current serving endpoint stable, even when dependency changes and flashinfer/JIT cache issues would otherwise break the upgrade.

Quick Start

Run the standard cutover flow for this pattern from the repository using install.sh with the skill name.