Flyte

Flyte is an open source platform for orchestrating workflows. It started at Lyft, is licensed under Apache 2.0, and is hosted by the Linux Foundation AI & Data. The current release, Flyte 2, is pitched as a durable runtime for training, serving, and agent workloads. Teams write pipelines as plain async Python functions. When a node is preempted or a task runs out of memory, Flyte resubmits it with corrected resources. When code fails, it resumes from the failed task instead of rerunning the whole pipeline. It can fan out across thousands of tasks and supports caching and versioned workflows. It can also host long-running services such as vLLM endpoints, FastAPI apps, and Streamlit dashboards, which scale to zero when idle. Pricing comes in three options. The self-hosted open source edition is free and runs on your own Kubernetes cluster. Union.ai's managed Team plan has a $950 monthly minimum that is credited against metered usage, billed per action, vCPU-hour, GB-hour, and GPU-hour. Enterprise is priced on a custom annual commitment and adds SSO, SCIM, customer-managed keys, multi-cloud support, and air-gapped deployment. Integrations cover Spark, Ray, Dask, PyTorch Elastic, Snowflake, BigQuery, Databricks, MLflow, Weights and Biases, LangGraph, CrewAI, and the OpenAI Agents SDK. The main drawback is operational. Self-hosting requires Kubernetes expertise and containerized tasks, and reviewers describe the learning curve as steep for teams without platform engineers. Because Flyte is aimed at ML work, some teams end up running a second orchestrator for general data pipelines.