A minimal FastAPI service showing the async-native GrowthBook integration pattern for high-concurrency Python services:

• One process-wide GrowthBookClient, created and closed by FastAPI's lifespan hook. Never create a client per request.
• Async Redis sticky bucket service (AbstractAsyncStickyBucketService, growthbook >= 2.4.0) — sticky bucket reads and writes never block the event loop. get_all_assignments is overridden with one batched MGET.
• Per-request UserContext — the client holds no user state, so one instance serves every request concurrently.

pip install -r requirements.txt

# Optional but recommended: real Redis for sticky bucketing
docker compose up -d redis
export REDIS_URL=redis://localhost:6379/0

# Point at your GrowthBook instance
export GB_API_HOST=https://cdn.growthbook.io
export GB_CLIENT_KEY=sdk-your-key

uvicorn main:app --reload

Without REDIS_URL the example falls back to an in-process store so it runs out of the box (not for production — assignments are lost on restart and not shared across workers).

curl 'localhost:8000/checkout?user_id=user-123'
curl localhost:8000/healthz

Create a feature named checkout-experiment (an experiment rule with sticky bucketing enabled) and a flag new-checkout-flow in GrowthBook to see real variations; unknown features fall back to their defaults.

With a sync sticky bucket service, every network round-trip to your assignment store runs on (or is offloaded from) the event loop. The async interface lets the SDK await your store natively: reads are prefetched per evaluation, writes are fire-and-forget and drained on close(). See the SDK benchmark (tests/scripts/benchmark_async_client.py in growthbook-python) for the difference under load.