Buying snapshot
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Decision point | What Flyte and Union.ai state |
|---|---|
License | Apache 2.0; a Graduated project of the LF AI & Data Foundation |
Try it locally |
|
Self-hosted operation | Union.ai describes open-source Flyte as installed and operated by you on your own Kubernetes, with as many clusters as you operate |
Managed operation (Union.ai) | Union hosts the control plane; you deploy the Union operator into a Kubernetes cluster you manage in your own AWS, GCP, Azure or neo-cloud account, unless you pay Union to run clusters. Workflow inputs, outputs, logs and code bundles are served from inside your cloud |
Moving code to Union.ai | Same SDK; existing Flyte code pointed at a Union tenant runs with no rewrite |
Support | Open source: community on Slack and GitHub. Union Team: Standard, 12×5, Sev 1 in 4 business hours. Union Enterprise: Premier, 24×7 for Sev 1–2; Mission Critical is a paid upgrade |
Flyte 1 users | Flyte 1 is now maintained on the repository's master branch; flyte.org links migration resources |
Our take
Flyte is worth a pilot if your Python team runs training, batch inference or agent workloads that fail on infrastructure, not only on code. If your jobs are simple, linear pipelines, its durability features add less.
Evaluate Flyte
Comparisons
Where it stands out
Recovery from infrastructure failures. Flyte 2 recovers automatically from an OOM kill, a preempted spot node or a GPU that vanishes mid-run, without recovery code from you. These infrastructure retries do not use up the task's retries budget.
Plain Python workflows. Flyte 2 drops the DSL. Tasks call tasks inside
TaskEnvironments, and branching or looping is ordinaryif,forandtry/except.Reproducible runs. Executions are logged and versioned automatically, and
cache="auto"reuses a task's result for identical inputs.Spot capacity with a fallback. Interruptible tasks can run on spot or preemptible instances; when a spot instance is reclaimed, Flyte 2 falls back to on-demand on the final attempt.
Serving and sandboxes in the same runtime. Flyte 2 adds real-time inference for single node models and a built-in sandbox for running LLM-generated Python code safely.
Integrations for existing stacks. Listed integrations include Apache Spark, Ray, PyTorch Elastic, BigQuery, Snowflake and Weights & Biases, plus LangGraph, CrewAI, MLflow, Dask, Databricks and a community plug-in registry.
Product screenshot
Trade-offs to weigh
Multi-node Flyte 2 depends on Union.ai for now. The repository says the open-source Kubernetes backend for running Flyte 2 as a distributed, multi-node service is coming soon, and points teams that need a production-grade Flyte 2 backend today to Union.ai.
You pay for what you request, not what you use. Union.ai bills vCPU and memory as allocated from each task's resource request, to the second, so generous resource requests raise the bill.
Short history on Team. Team keeps data for 30 days; Union names retention beyond 30 days as one reason to move to Enterprise.
Pricing and plan conditions
Flyte itself is free. The prices below are for Union.ai, the managed Flyte platform. On Team, Union charges $950 on the first day of each billing cycle and issues it back as usage credit, so you pay the greater of your usage and $950. Unused credit expires at the end of the cycle, and usage beyond it is billed at the rate card. The first month's minimum is waived, and no card is needed at signup.
Usage has four meters: actions, vCPU, memory and GPU. An action counts once the task, trace or condition succeeds, however long it runs and however many retries it takes. Bands are marginal and reset each cycle. Union's own example: 2.5M actions in a cycle are first 100k × $0.0030 = $300, next 900k × $0.0015 = $1,350 and next 1.5M × $0.0008 = $1,200, for $2,850 before compute. Union says Team runs in your own cloud, so you keep your negotiated rates and your cloud credits.
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Plan or option | Published price | What it adds |
|---|---|---|
Flyte (open source) | Free | Apache 2.0. The full Flyte SDK and engine: durable tasks, caching, retries, type-safe data, the Flyte UI |
Union.ai Team | $950 / mo minimum, credited to usage | Everything in Flyte, plus the Union control plane: Zero Trust, remote debugging, replay and recover, Volumes, federated login with SCIM. 2 active clusters, 10,000 concurrent actions, 200,000 actions per run |
Union.ai Enterprise | Custom; annual commit, fully credited to usage | Everything in Team, plus significant discounts on every meter, private networking, CMEK and BYOK, SSO and SAML, custom RBAC; unlimited clusters, multi-region and multi-cloud; custom concurrency and fanout; self-hosted option |
Union-managed clusters (optional) | $500 / cluster / mo, capped at $2,500 | Union provisions, upgrades and operates the clusters in your cloud. Self-managed clusters carry no fee; first Union-managed cluster free in month one for new accounts |
Single-tenant control plane | $50k / yr | Listed under Enterprise |
Guided Install (one-time) | Standard $5,000 · Advanced from $15,000 | One-time service |
Scroll sideways to see all columns. With a keyboard, focus this table and use the arrow keys.
| Meter | Published rate | Monthly bands |
|---|---|---|
Actions | $0.0030 → $0.0003 per action | $0.0030 (0–100k / mo), $0.0015 (100k–1M), $0.0008 (1M–10M), $0.0003 (10M+) |
vCPU | $0.0230 → $0.0130 per vCPU-hour | $0.0230 (0–100k vCPU-hrs / mo), $0.0173 (100k–1M), $0.0130 (1M+) |
Memory | $0.0032 → $0.0018 per GB-hour | $0.0032 (0–600k GB-hrs / mo), $0.0024 (600k–6M), $0.0018 (6M+) |
GPU Class 1 (e.g. T4, L4, Inferentia2) | $0.068 / $0.056 / $0.045 per GPU-hour | 0–1k / 1k–10k / 10k+ hrs per month |
GPU Class 2 (e.g. A10G, L40S, A100) | $0.24 / $0.20 / $0.16 per GPU-hour | 0–1k / 1k–10k / 10k+ hrs per month |
GPU Class 3 (H100, H200, B200, B300) | $0.89 / $0.74 / $0.59 per GPU-hour | 0–1k / 1k–10k / 10k+ hrs per month |
Reserved or owned GPUs | Flat fee per GPU per month | Unlimited hours; quote from Union sales |
Public signals
About these signals
Verified factual signals from public sources. They indicate observable activity or interest, not total adoption, product quality, or cost.
329 GitHub commits 90d7.8k GitHub stars0 vulnerabilities across 1 package
See all signals from 5 sources
PyPI · flyte@2.12.1
Alternatives to Flyte
The reviewed substitutes for Flyte among the ML pipeline frameworks, and what would make each one the better answer.
Direct alternatives
Reviewed substitutes: products bought for the same job, where a team picks one.
- Kubeflow
- Kubeflow and Flyte both orchestrate machine learning pipelines on Kubernetes. Kubeflow is a broad open-source platform whose projects for notebooks, training and serving can be adopted one by one, while Flyte puts the focus on typed, versioned and cached Python tasks and has a managed version through Union.Applies to: Choosing between two products of the same kind for one job.
- Metaflow
- Metaflow and Flyte are open-source frameworks for writing ML and data pipelines in Python and running them in production. Metaflow focuses on the data scientist's experience, with dependency management and a smooth move from laptop to cloud, while Flyte focuses on Kubernetes-native execution with strong typing, caching and versioning.Applies to: Choosing between two products of the same kind for one job.
Other approaches
A different approach to the same problem. Each substitutes only for the workload named beside it.
- ZenML
- Open-source MLOps framework for building portable, production-ready ML pipelines — pluggable stack components, artifact versioning, and pipeline orchestration.Applies to: for teams that need Python or Java workflows with type-safe tasks and multi-tenant Kubernetes execution.
- MLflow
- MLflow and Flyte both keep machine learning work reproducible. MLflow logs runs, parameters, metrics and models and keeps a central model registry, while Flyte versions each typed pipeline task on Kubernetes and caches its outputs, so a team that wants run history to come from its pipeline tool can rely on Flyte for both jobs.Applies to: Deciding whether run tracking and pipeline orchestration come from one tool or two.