The Message Knows the Way.

🎭 Asya is a Kubernetes-native actor mesh for async AI/ML pipelines. Each actor is a pure Python function. Each message carries its own route. No central orchestrator. No SDK. No waiting.

AI workloads are batch jobs — bursty, GPU-heavy, multi-step. Forcing them into synchronous REST patterns creates three problems:

1. Wasted resources — GPU pods sit idle between bursts; you pay for 24/7 what you need for minutes
2. Cascading failures — one slow step blocks the entire pipeline; retry logic is bolted on
3. Tight coupling — models, transports, and orchestration are tangled in application code

Write a pure Python function. Deploy it as a Kubernetes CRD. Asya handles the rest:

# handler.py — your code, zero dependencies
def process(state: dict) -> dict:
    state["output"] = model.predict(state["input"])
    return state

# actor.yaml — the platform's job
apiVersion: asya.sh/v1alpha1
kind: AsyncActor
metadata:
  name: my-model
spec:
  image: my-model:latest
  handler: handler.process
  scaling:
    minReplicas: 0
    maxReplicas: 50

Two files, two owners: data scientist owns the handler, platform team owns the spec.

Actors communicate through queues. Each message carries its own route.

• Decentralized routing — each envelope carries prev/curr/next queues; no coordinator to fail
• Scale to zero — KEDA watches each actor's queue independently; GPU pods cost nothing when idle
• Pure Python handlers — no SDK imports, no decorators; test with assert process(payload) == expected
• Dynamic pipelines — routes are data, not code; actors rewrite routing at runtime for agentic patterns, LLM judges, and branching flows
• Multi-transport — SQS, RabbitMQ, GCP Pub/Sub; switch without changing handler code
• A2A + MCP gateway — expose actor pipelines as A2A agents or MCP tools with streaming and pause/resume

helm repo add asya https://asya.sh/charts
helm repo update asya
helm install asya asya/asya-playground --namespace asya-demo --create-namespace

See the full Quickstart Guide for cluster setup, verification, and sending your first message.

See CONTRIBUTING.md. Prerequisites: Go 1.24+, Python 3.13+, Docker, Make.

Born from 3 years of production AI workloads at Delivery Hero. Open source and growing.

Apache 2.0 licensed. 🎭 Maintainer: Artem Yushkovskiy (@atemate)