Flyte v1.3.0

Flyte offers several key features:

• Author in pure Python: Write workflows in actual Python, no DSL required.
• Durable by default: Build fault-tolerant workflows that retry automatically and pick up where they left off.
• Dynamic workflow execution: Workflows can make on-the-fly decisions at runtime with real-time logic, conditions, and retries.
• Self-healing workflows: Workflows can autonomously recover from failures and continue.
• Run locally: Test and debug tasks in your local environment using the same Python SDK that runs in production.
• Infra-aware orchestration: Provision and scale resources for workflows.
• Integrations: Includes integrations with Apache Spark, BigQuery, PyTorch Elastic, Ray, Snowflake, and Weights & Biases.

Flyte is used for orchestrating AI, ML, and agentic workflows. It is suitable for building dynamic, self-healing workflows that require infrastructure awareness. Use cases include:

• AI/ML pipeline orchestration: Build and scale dynamic AI/ML workflows.
• Agentic workflows: Create workflows that adapt at runtime with stateful execution and automatic failure recovery.
• Data processing: Run Spark jobs on ephemeral clusters, query BigQuery or Snowflake, and more.
• Distributed training: Use PyTorch Elastic or Ray for multi-node distributed training.
• Experiment tracking: Integrate with Weights & Biases for experiment and inference-time tracking.
• Case studies: Companies like Spotify, Wayve, Cradle, Warner Bros. Discovery, Kineo, and MethaneSAT use Flyte to accelerate their AI/ML development and reduce costs.

Flyte is open-source and free to use. The page mentions "Try for free" for Union.ai, the enterprise platform built on Flyte, which offers additional features like massive scale, real-time inference, live remote debugging, and white-glove support. Pricing details for Union.ai are not specified on the page.