Flyte

Flyte 2 is a Python-authored runtime for durable AI workloads, covering training, serving, and agent execution. It is aimed at teams comfortable running software locally or on a cluster and needing workflows that can recover from infrastructure failures. The open-source runtime keeps entry cost at zero, while managed capacity and support become a significant commitment.

Overview

Flyte’s central proposition is durable execution for AI pipelines rather than a general-purpose visual automation tool. You define workflows as Python async functions and run them on a local devbox or your own cluster. That deployment choice suits engineering teams that want control of their environment; it is less suitable for buyers seeking a fully hosted service with no operational responsibility.

Flyte originated inside Lyft in 2016, was open-sourced in 2020, and joined the Linux Foundation AI & Data in 2022. Union.ai presents itself as the enterprise Flyte platform. Its security position is aimed at organizations with strict data boundaries: Union.ai states that the data plane runs in the customer’s cloud and that workflow inputs, outputs, logs, code, and reports do not pass through its control plane. Its security page lists SOC 2 Type I, SOC 2 Type II, and HIPAA.

Key features

Plain Python async functions make workflow authoring familiar to Python developers. A TaskEnvironment puts container images, resource requirements, secrets, and retry policy beside the task definition, so execution needs are explicit rather than hidden in separate configuration. Teams that do not want to write Python will find little reason to select this approach.

Failure recovery is Flyte’s strongest practical differentiator. A failed task can resume, and the runtime can recover from out-of-memory kills, preempted nodes, or disappearing GPUs. That matters when training or agent runs are costly to repeat. Automatic retries, event triggers, and execution observability reinforce the same operational model, although teams still need to manage the underlying deployment.

Branching, loops, and runtime decisions support dynamic workflows whose next actions depend on results produced during execution. This is useful for agents and other workloads that cannot be reduced to a fixed sequence. Native app integrations include Streamlit, FastAPI, vLLM, SGLang, and Ollama, with replicas able to scale to zero while idle. Integrations also include Apache Spark, Ray, PyTorch Elastic, BigQuery, Snowflake, Weights & Biases, OpenAI Agents SDK, LangGraph, CrewAI, Pydantic AI, MLflow, Dask, and Databricks.

Pricing

Flyte OSS — free. The Apache 2.0 open-source runtime runs locally or on your cluster. It is the sensible starting point for individual developers and teams willing to provide their own infrastructure, with no managed-service capacity or support commitment included.

Team — $950 a month. This tier is billed at $950 per month plus usage, with a 30-day free trial. The $950 monthly minimum is credited to usage. It includes 1,000 concurrent actions, 30-day data retention, one cluster, a 25,000-action-per-run fanout limit, and one-business-day Sev 1 support. The minimum makes Team difficult to justify for occasional or small workloads, but it gives an operating team defined concurrency, retention, and response terms.

Enterprise — custom pricing. Enterprise lists custom concurrent actions and retention, 3+ clusters, a 50,000+ actions-per-run fanout limit, and custom white-glove support. It is the appropriate commercial tier when one cluster, Team’s concurrency, or its retention period is insufficient; buyers must obtain a tailored quote.

Platforms

Flyte is available through its API and web interfaces on Linux, with a self-hosted deployment model. The documented local path uses pip install flyte followed by flyte start devbox; production use can run on a user’s cluster. That flexibility helps teams align deployment with existing cloud controls, but it places platform operations in the buyer’s remit.

Who it's for

Choose Flyte for Python-heavy teams building durable training, serving, or agent workflows; workloads exposed to preemption, GPU loss, or memory failures; and organizations that need local or cluster deployment with customer-cloud data handling. The OSS plan fits experimentation and self-managed production. Team fits organizations that can absorb a $950 monthly minimum and need defined managed capacity. Enterprise fits multi-cluster deployments and buyers requiring custom limits or white-glove support.

Pros and cons

• Pros: Python async authoring keeps workflow code close to application logic.
• Pros: Resume and infrastructure-failure recovery reduce expensive reruns.
• Pros: Dynamic branches, loops, and runtime decisions suit agent behavior.
• Pros: Broad AI, data, and serving integrations reduce connector work.
• Pros: Local and cluster deployment, plus customer-cloud data-plane positioning, support controlled environments.
• Cons: Self-hosting means the customer operates the deployment.
• Cons: Team’s $950 monthly minimum is a high floor for modest usage.
• Cons: Team is limited to one cluster, 1,000 concurrent actions, 30-day retention, and 25,000 actions per run.

Alternatives

For a different workflow and data-operations focus, compare DataOps Platforms and Workload Automation Software. Teams specifically evaluating machine-learning lifecycle tooling can review MLOps Platforms, while broader orchestration requirements belong in Workflow Orchestration Software.

Kestra is a free, Apache 2.0 platform available through API, Linux, macOS, self-hosting, and web interfaces; choose it when that broader platform footprint is preferable. Apache NiFi is free to download under Apache License 2.0 across API, Linux, macOS, self-hosting, web, and Windows, making it an alternative for teams prioritizing a no-cost distribution. DataKitchen TestGen offers a free-forever open-source plan with one user, one connection, and one project, suited to a narrower testing scope. Apache Airflow is free to download and use under the Apache License. Apache Hop is an open-source Apache-licensed platform for Linux, macOS, self-hosting, and Windows, with Java 21 required for release 2.19.0. Metaflow is free under Apache 2.0, with production scaling run in your own cloud account or Kubernetes cluster. Keboola has a free tier with 120 compute minutes in the first month, then 60 minutes monthly, plus paid compute. Luigi is a free open-source Python package without registration walls.

Verdict

Flyte is a strong choice for Python teams that need durable, dynamic AI workflows and want to deploy locally or on their own cluster. Its recovery model and integrations are the main reasons to choose it. Look elsewhere if you need a no-operations hosted experience, non-Python workflow authoring, or cannot justify Team’s $950 monthly minimum once self-managed OSS is no longer enough.