Draft. This page is reachable and it is not indexed until Vincent approves it.
Dagster vs Prefect comes down to the unit you want to manage. Dagster is built around data assets: the tables and files your pipeline should produce. Prefect is built around flows: Python functions it runs, retries, and tracks. Since July 2026 both belong to Prefect, and both remain supported products. Pick on the model, not on a benchmark.
By Vincent Charles. Last updated . Pricing checked Nov 19, 2026.
Vincent Charles is a data engineer and founder of dataengineers.club and Unchain Data (unchaindata.xyz). He has worked on data at Dailymotion's big data team, Carrefour and Binance, starting in SQL and GIS location intelligence.
Disclosure: dataengineers.club does not sell an orchestrator.
On this page
Key takeaways
- The core difference is the model. Dagster asks "which data assets should exist?" Prefect asks "which Python work should run, and how should it recover?"
- Same owner, two products. Prefect announced it was acquiring Dagster Labs on July 13, 2026. Both companies say Dagster keeps its name, open-source license, and pricing (Prefect, Dagster).
- Both are Apache 2.0 open source. Both also sell a managed cloud: Dagster+ and Prefect Cloud.
- Pricing units differ. Dagster+ bills a monthly plan fee plus credits (asset materializations and ops). Prefect Cloud bills a plan fee, with Team priced per user (Dagster pricing, Prefect pricing).
- Same pipeline, near-identical code. In our local test both tools produced the same 1,158-row daily summary, both retried an injected failure as documented, and the pipeline code was 94 lines in Dagster and 93 in Prefect. We do not publish a "faster" claim.
Definitions in one place
- Orchestrator: a tool that schedules, runs, retries, and records the steps of a data pipeline.
- Asset (Dagster): "an object in persistent storage, such as a table, file, or persisted machine learning model." An asset definition describes in code how to produce and update it (Dagster docs).
- Materialization (Dagster): running an asset's function and saving the result to storage.
- Flow (Prefect): a decorated Python function. Prefect tracks its states and can retry it, time it out, and deploy it (Prefect docs).
- Task (Prefect): a smaller unit of work inside a flow, with its own retries and caching (Prefect docs).
- Deployment (Prefect): a server-side definition of a flow that can be scheduled and triggered remotely (Prefect docs).
Who this is for
- Teams choosing their first Python orchestrator for a warehouse, dbt, or ML stack and narrowing it to these two.
- Airflow users who want to know whether either tool fixes a specific problem they have today. If you want the wider field, see our Airflow alternatives guide.
- Small data teams (one to ten engineers) deciding between self-hosting the open-source version and paying for a managed control plane.
- Workload types covered: batch ELT, dbt runs, file and API ingestion, ML training jobs. Streaming is out of scope for both tools and for this page.
- What we assume you already run: Python 3.11 or newer, Git, and a warehouse or object store. Dagster's install guide requires Python 3.11 or higher (Dagster docs).
Not for you if you need a real-time stream processor, or if you are comparing Temporal or Mage. We did not test those.
Methodology
- Status: test run on 2026-11-19. Results below come from our run notes and logs.
- Date of docs and pricing review: 2026-11-19.
- Versions installed: dagster 1.13.30 and prefect 3.9.4, the latest stable releases on PyPI that day (PyPI dagster, PyPI prefect), with DuckDB 1.3.2 in both pipelines.
- Machine: Debian 13 (Linux 6.12, x86_64), 8 vCPU, 15 GiB RAM, Python 3.13.5.
- Where it ran: local only, using each tool's own local UI (
dg devfor Dagster,prefect server startfor Prefect). No managed deploy. - What we ran: one three-step pipeline (ingest a public CSV, transform it, write a table), with an injected failure on the first attempt to exercise retries.
- What we measure: human time to first successful local run, lines of pipeline code, whether retries and the run view behave as documented, and the deployment path used.
- What we do not measure: throughput, latency, cost at scale, Kubernetes deploys, or any "developer experience score".
- Repo: repo coming soon (both implementations, pinned requirements, and raw logs).
- Not run: Dagster+ or Prefect Cloud, Docker, Kubernetes or work-pool deploys, a scheduled run firing, and any partitions or backfill job.
- Docs-only sections: the feature matrix, pricing, and day-2 notes are documented from vendor docs and pricing pages as of 2026-11-19, and say so.
Feature matrix
This table compares the documented feature surface. It is sourced from official docs, not from our test.
| Dimension | Dagster | Prefect | Source / as of |
|---|---|---|---|
| Center of the model | Assets: declared tables, files, models. Ops exist underneath, but Dagster recommends assets for new pipelines | Flows and tasks: decorated Python functions. Prefect 3 also tracks assets as workflow outputs | Dagster assets, Prefect flows, Prefect assets, 2026-11-19 |
| Local development | Scaffold with create-dagster, run the UI with dg dev, launch with dg launch | Run a flow as a plain Python script; start a local server and UI with prefect server start | Dagster quickstart, Prefect server, 2026-11-19 |
| Scheduling | Cron schedules, sensors, Declarative Automation | Schedules attached to deployments, event-driven automations | Dagster schedules, Prefect schedules, 2026-11-19 |
| Retries | RetryPolicy on ops and asset jobs (count, delay, backoff, jitter); run-level retries configured per instance | retries and retry_delay_seconds on flows and tasks, plus jitter and custom retry conditions | Dagster op retries, Dagster run retries, Prefect retries, 2026-11-19 |
| Partitions and backfills | Built in for assets | Not a documented Prefect feature: Prefect's full docs index has no partitions or backfills page | Dagster partitions, Prefect docs index, 2026-11-19 |
| Observability / UI | Asset graph, lineage, run view, asset catalog; Insights on Dagster+ | Flow run states, logs, deployments, automations in the UI | Docs above, 2026-11-19 |
| Deployment options | Self-host OSS; Dagster+ Serverless or Hybrid | Self-host Prefect server; Prefect Cloud with hybrid, push, or managed work pools | Dagster+, Prefect work pools, 2026-11-19 |
| Managed offer | Dagster+ | Prefect Cloud | Pricing pages, 2026-11-19 |
| License | Apache 2.0 | Apache 2.0 | Dagster LICENSE, Prefect LICENSE, PyPI metadata 2026-11-19 |
| Weaker fit | Highly dynamic, branch-at-runtime Python work where there is no clear data asset | Asset-centric platforms that lean on partitions, lineage, and a data catalog | Our reading of the docs above; not covered by our test |
The two products have moved toward each other. Prefect now documents assets with materialized, referenced, and external states (Prefect assets). Dagster's pricing page lists both "asset-based orchestration" and "workflow-based orchestration" (Dagster pricing). The default way of thinking still differs, and that is what you will feel day to day.
Pricing as of 2026-11-19
Pricing checked Nov 19, 2026 on each vendor's public pricing page. Both open-source versions are free to self-host. You pay for your own compute either way.
| Item | Dagster (Dagster+) | Prefect (Prefect Cloud) | Source |
|---|---|---|---|
| Free / open-source | Dagster OSS, Apache 2.0, self-hosted. Dagster+ has a 30-day free trial | Prefect OSS, Apache 2.0, self-hosted. Prefect Cloud Hobby tier is free: 2 users, up to 5 deployments, 500 minutes of Prefect Serverless, 7-day run retention | Dagster pricing, Prefect pricing |
| Entry paid plan | Solo: $10 per month plus $0.040 per credit | Starter: $100 per month (3 users, up to 20 deployments, 75 hours of Prefect Serverless) | Same |
| Team plan | Starter: $100 per month plus $0.035 per credit | Team: $100 per user per month (4 to 8 users, up to 100 deployments, 225 hours of Prefect Serverless) | Same |
| Enterprise | Pro: custom, contact sales | Enterprise: custom | Same |
| Billing unit | Plan fee plus credits. One credit equals one asset materialization or one op executed. Serverless compute is $0.010 per minute; Hybrid has no compute charge | Plan fee (per user on Team) with included serverless hours and deployment limits | Same |
| Compliance listed | SOC 2 Type II and GDPR on all plans; HIPAA marked as not included on Solo and Starter | SOC 2 Type II, HIPAA, and GDPR listed on the page; plan-level HIPAA scope not stated | Same |
| What we could not verify | Pro plan price: not on the public page as of 2026-11-19 | Enterprise price and overage rates beyond included serverless hours: not on the public page as of 2026-11-19 | Same |
Notes:
- The billing units are not comparable one to one. A Dagster+ bill grows with the number of asset materializations and ops. A Prefect Cloud bill grows mainly with seats and deployments. Model your own workload before deciding on cost.
- The acquisition announcement says pricing for both Prefect Cloud and Dagster+ is unchanged. Recheck both pages before you sign anything (Dagster).
- We did not buy either plan for this test and used no trial. Both runs were local and free.
Pick Dagster or Prefect if...
| Situation | Choose | Why |
|---|---|---|
| Your pipeline is a graph of warehouse tables, dbt models, and files | Dagster | Assets are the core object, with lineage and a catalog in the UI (docs) |
| You need partitioned data and reliable backfills | Dagster | Partitions and backfills are a built-in asset feature (docs) |
| You have existing Python scripts and want them tracked fast | Prefect | A flow is a decorated function that still runs as normal Python (docs) |
| Work is dynamic: API fan-out, ML jobs, agent workflows | Prefect | Flow and task retries, timeouts, and event automations are the core model (docs) |
| You want per-run infrastructure on Docker, Kubernetes, or serverless | Prefect | Work pools choose infrastructure per deployment (docs) |
| Cost scales with seats in your org, not with jobs | Dagster | Dagster+ Solo and Starter are a flat fee plus per-credit usage, not per seat (pricing) |
| You run Airflow today and it works | Neither, stay on Airflow | No tested gain justifies a rewrite. See Airflow alternatives |
What we actually ran
We ran this locally on 2026-11-19. Times come from our logs. The tester had already read the docs and written the plan, so treat the times as a lower bound, not a typical first-time experience.
The shared pipeline
Both implementations do the same three steps:
- Ingest: download the NYTimes COVID-19 US state-level time series (us-states.csv), 61,942 data rows covering 2020-01-21 to 2023-03-23.
- Transform: clean column names, cast types, and write Parquet.
- Write: a daily summary (sum of cases and deaths, count of state rows) to a local DuckDB table and a Parquet file.
An injected error makes the ingest step fail on its first attempt. Both pipelines recovered through the tool's own retry setting, with no manual rerun. Both produced 61,942 cleaned rows and 1,158 daily summary rows, with the same last-day totals.
How each version was built
- Dagster: scaffolded with
create-dagster, three assets, andRetryPolicy(max_retries=3, delay=1); run withdg launchand inspected indg dev. - Prefect: one
@flowcalling three@taskfunctions, withretries=3, retry_delay_seconds=1on ingest; run as a script, then against a local server.
Results
| Measure | Dagster | Prefect |
|---|---|---|
| Installed version and Python version | dagster 1.13.30, Python 3.13.5 | prefect 3.9.4, Python 3.13.5 |
| Package install (warm cache) | 2 seconds, 102 packages | 2 seconds, 104 packages |
| Time to first successful local run | 4 min 31 s from install to first success, including writing code and one failed run; the successful launch took 8.4 s | 11.5 s for the first script run (10.2 s of startup, 1.3 s of flow time), with no failed run because the SQL was already fixed |
| Lines of pipeline code (excluding imports, blank lines, comments, docstrings) | 94 | 93 |
| Did the injected failure retry as documented? | Yes, succeeded on attempt 2 | Yes, run_count=2 |
| What the run view showed for the failed attempt | Run SUCCESS; raw_data shows two attempts with retry and re-execution events | Run COMPLETED; task states Pending, Running, AwaitingRetry, Retrying, Completed. The error text is in the log, not the state message |
| Deployment path used | Local only (dg launch, dg dev) | Local only (script, then local server) |
Surprises and doc mismatches:
- Dagster:
dg list defsfailed withModuleNotFoundErroruntil we installed the scaffolded project in editable mode. A comma-separated asset selection string raisedDagsterInvalidAssetSelectionError;selection="*"worked. Our first launch failed on a DuckDB SQL error (DuckDB 1.3.2 rejected?placeholders insideread_csv_autoandCOPY), not a Dagster issue. Because the retry policy covered that step, Dagster made four attempts on the SQL error before failing: retries cannot tell a flaky source from a bug. - Prefect: the first script run spent 10.2 seconds in import and startup (11.5 seconds wall clock minus 1.3 seconds of flow time). A later run against a running local server took 2.6 seconds, 1.4 of them flow time.
How the code shape differed
The docs description matched what we wrote. Dagster declares three assets, and the dependency graph decides run order:
@asset(retry_policy=RETRY_POLICY)
def clean_data(context: AssetExecutionContext, raw_data: str) -> str:
...Prefect is an ordinary Python call chain inside a flow:
@flow(name="csv_pipeline")
def csv_pipeline() -> str:
raw = ingest_raw_data()
clean = transform_clean_data(raw)
return write_daily_summary(clean)The SQL was identical. At this size the difference is small: 94 versus 93 lines, though Dagster also needed a project layout, an editable install, and a job declaration, while Prefect ran as one module. We declared a daily schedule on the Dagster job but did not wait for it to fire, and we did not run a partitions or backfill job in either tool.
Day-2 operations
Day-2 operations are what happens after the first run works: retries, deploys, monitoring, and local iteration. This section is documented from vendor docs as of 2026-11-19, not from our test.
Retries and failure handling
- Dagster has two levels. A
RetryPolicyretries a single op or asset step inside the same run, with backoff and jitter options. Run retries rerun the whole run, and they also cover a crashed or terminated run process. On Dagster OSS you enable run retries indagster.yaml; on Dagster+ they are on by default (op retries, run retries). - Prefect sets
retriesandretry_delay_secondson a flow or a task. Task-level retries keep the retry narrow. Prefect also supports jitter, custom retry conditions, and global retry settings (retries guide).
Deployment
- Dagster OSS is self-hosted. Dagster+ comes as Serverless, where your code runs in Dagster's environment, or Hybrid, where you run the execution environment and Dagster+ manages the control plane (Dagster+ docs).
- Prefect uses deployments plus work pools. Hybrid pools need a worker in your infrastructure. Push pools submit runs to a serverless provider. Managed pools run on Prefect's infrastructure (work pools).
Observability
- Dagster centers the UI on the asset graph and lineage. Dagster+ adds Insights, alerts, an asset catalog, and branch deployments (Dagster+ docs).
- Prefect centers the UI on flow runs and their states. A run that never reaches a final state, for example because its host died, is called a "zombie" run in Prefect's docs (flows).
Local development
- Dagster:
dg devstarts the local UI against your project. - Prefect: a flow runs as a normal Python script.
prefect server startadds the local UI, backed by SQLite by default, with PostgreSQL as an option (server guide).
Failure modes to plan for
- Self-hosted Dagster: you own the database, daemon, and upgrades.
- Self-hosted Prefect: you own the server database and its migrations.
- Managed plans: retention limits apply. Prefect Cloud keeps 7 days of run history on Hobby and Starter, and 14 days on Team (pricing).
Limitations
Where each tool is a poor fit
- Dagster asks you to think in assets. If your work is a long chain of side effects (send emails, call APIs, trigger other systems) with no clear output table, the asset model adds structure you may not need.
- Prefect gives you flexible Python. If your main need is partitioned backfills, column lineage, and a data catalog, you will build more of that yourself.
- Both add a server, a database, and a deployment process. A cron job plus logging may be enough for one or two scripts.
What we could not test
- Scale, throughput, or cost under real load.
- Kubernetes, Docker, or cloud deploys. Both runs were local only.
- Enterprise features: SSO, RBAC, audit logs, HIPAA scope.
- How the July 2026 acquisition will change either roadmap. Both companies said no specific product changes had been announced at the time (Prefect FAQ).
- Temporal, Mage, Airflow, or any other orchestrator.
FAQ
Did Prefect buy Dagster?
Yes. On July 13, 2026, Prefect announced it was acquiring Dagster Labs, including the Dagster product, codebase, customers, and many of its team. Both companies say Dagster keeps its name, open-source license, and pricing, and that no customer migration is required. The combined company was expected to operate under the Prefect name from August 2026 ([Dagster](https://dagster.io/prefect), [Prefect FAQ](https://www.prefect.io/prefect-dagster-faq)). Checked 2026-11-19.
What is Dagster and Prefect?
Both are open-source Python orchestrators that schedule, run, and monitor data pipelines. Dagster models pipelines as data assets, such as tables and files that should exist. Prefect models them as flows and tasks, which are decorated Python functions it tracks and retries. Each has a managed cloud: Dagster+ and Prefect Cloud. Both are Apache 2.0 licensed as of 2026-11-19.
Why is Dagster better than Airflow?
It is not better for everyone. Dagster suits teams that want an asset-centric model with lineage, partitions, and backfills built in, plus local development through `dg dev` ([Dagster docs](https://docs.dagster.io/getting-started/quickstart)). If your Airflow setup is stable and your team knows it, the migration cost may outweigh the gain. See our [Airflow alternatives](/alternatives/airflow) guide for a wider comparison.
Is Prefect better than Airflow?
It depends on the work. Prefect fits dynamic Python workflows where you want to wrap existing functions with retries, timeouts, and tracking, and choose infrastructure per deployment through work pools ([Prefect docs](https://docs.prefect.io/v3/concepts/work-pools)). Airflow remains a reasonable choice for teams with mature DAGs, operators, and managed hosting already in place. We did not run Airflow in this test.
How much do Dagster and Prefect cost?
Both open-source versions are free to self-host. As of 2026-11-19, Dagster+ Solo is $10 per month plus $0.040 per credit and Starter is $100 per month plus $0.035 per credit. Prefect Cloud has a free Hobby tier, Starter at $100 per month, and Team at $100 per user per month. Enterprise prices for both are not public ([Dagster](https://dagster.io/pricing), [Prefect](https://www.prefect.io/pricing)).
Can I use Dagster and Prefect together?
Yes. Prefect's announcement frames Dagster as defining outcomes and Prefect as executing the work ([Prefect](https://www.prefect.io/prefect-acquires-dagster)). Dagster Labs publishes `dagster-prefect`, a library that launches Prefect deployments or background tasks from a Dagster asset, while Dagster handles scheduling, partitions, lineage, and retries. As of 2026-11-19 Dagster's docs label it a preview feature, not ready for production ([Dagster docs](https://docs.dagster.io/integrations/libraries/prefect)).
What about Temporal or Mage?
We did not test them here. Temporal describes itself as "an open-source platform for building crash-proof applications that resume exactly where they left off after failures" ([Temporal docs](https://docs.temporal.io/llms.txt)). Mage builds pipelines in a visual editor from Python, SQL, and R blocks ([Mage docs](https://docs.mage.ai/migrations/mage-oss-to-mage-pro)). Our [Airflow alternatives](/alternatives/airflow) guide covers Mage.
Sources
All links checked on 2026-11-19.
- Prefect, "Prefect acquires Dagster Labs" (July 13, 2026): https://www.prefect.io/prefect-acquires-dagster
- Dagster, "Dagster is now part of Prefect" (July 13, 2026): https://dagster.io/prefect
- Dagster blog, "Prefect is Acquiring Dagster" (July 13, 2026): https://dagster.io/blog/prefect-is-acquiring-dagster
- Prefect customer FAQ on the Dagster acquisition: https://www.prefect.io/prefect-dagster-faq
- Dagster pricing: https://dagster.io/pricing
- Prefect pricing: https://www.prefect.io/pricing
- Dagster docs, assets: https://docs.dagster.io/guides/build/assets
- Dagster docs, quickstart: https://docs.dagster.io/getting-started/quickstart
- Dagster docs, op retries: https://docs.dagster.io/guides/build/ops/op-retries
- Dagster docs, run retries: https://docs.dagster.io/deployment/execution/run-retries
- Dagster docs, schedules: https://docs.dagster.io/guides/automate/schedules
- Dagster docs, partitions and backfills: https://docs.dagster.io/guides/build/partitions-and-backfills
- Dagster docs, about Dagster+: https://docs.dagster.io/deployment/dagster-plus
- Prefect docs, flows: https://docs.prefect.io/v3/concepts/flows
- Prefect docs, tasks: https://docs.prefect.io/v3/concepts/tasks
- Prefect docs, assets: https://docs.prefect.io/v3/concepts/assets
- Prefect docs, retries: https://docs.prefect.io/v3/how-to-guides/workflows/retries
- Prefect docs, deployments: https://docs.prefect.io/v3/concepts/deployments
- Prefect docs, work pools: https://docs.prefect.io/v3/concepts/work-pools
- Prefect docs, schedules: https://docs.prefect.io/v3/concepts/schedules
- Prefect docs, local server: https://docs.prefect.io/v3/how-to-guides/self-hosted/server-cli
- Dagster license: https://github.com/dagster-io/dagster/blob/master/LICENSE
- Prefect license: https://github.com/PrefectHQ/prefect/blob/main/LICENSE
- PyPI, dagster: https://pypi.org/project/dagster/
- PyPI, prefect: https://pypi.org/project/prefect/
- Prefect docs index: https://docs.prefect.io/llms.txt
- Dagster docs, Dagster and Prefect integration: https://docs.dagster.io/integrations/libraries/prefect
- Temporal docs index: https://docs.temporal.io/llms.txt
- Mage docs, OSS vs Pro: https://docs.mage.ai/migrations/mage-oss-to-mage-pro
Get the newsletter
One email when a comparison on this page is actually finished.