Asya FAQs

Question 1

Is Asya suitable for real-time AI inference with very low latency?

Accepted Answer

No, Asya is not ideal for real-time inference requiring latencies below 100ms due to the inherent queueing overhead. It is optimized for asynchronous, event-driven, and batch processing workloads. For critical low-latency inference, alternative solutions like KServe or Seldon are more appropriate.

Question 2

How does Asya achieve cost efficiency for AI/ML workloads?

Accepted Answer

Asya leverages KEDA (Kubernetes Event-driven Autoscaling) to scale individual actors from zero to N instances based on their queue depth. This ensures compute resources are only consumed when active processing is required, significantly reducing idle costs, especially for GPU inference workloads.

Question 3

What is Asya and what problems does it solve?

Accepted Answer

Asya is a Kubernetes-native, queue-based actor framework designed for orchestrating AI/ML workloads. It solves the complexity of building multi-step pipelines and event-driven applications by automating scaling, routing, and retries for pure Python functions.

Question 4

What kind of code do data scientists write when using Asya?

Accepted Answer

Data scientists write pure Python functions or classes to handle payloads. Asya abstracts away all infrastructure concerns like queues, routing, and scaling, allowing them to focus entirely on their core AI/ML logic without needing decorators, DAGs, or boilerplate infrastructure code.

Question 5

How does Asya handle message routing and scaling on Kubernetes?

Accepted Answer

Asya employs a sidecar pattern. An Operator deploys your Python handler alongside a Go-based sidecar that manages queue consumption, robust message routing, and retries. KEDA then continuously monitors queue depth to automatically scale your actor pods up or down, even to zero.

Asya

About

Key Features

Use Cases