metaflow-dagster

Deploy and run Metaflow flows as Dagster jobs.

metaflow-dagster generates a self-contained Dagster definitions file from any Metaflow flow, letting you schedule, monitor, and launch your pipelines through Dagster while keeping all your existing Metaflow code unchanged.

Install

pip install metaflow-dagster

Or from source:

git clone https://github.com/npow/metaflow-dagster.git
cd metaflow-dagster
pip install -e ".[test]"

Quick start

python my_flow.py dagster create dagster_defs.py
dagster dev -f dagster_defs.py

Usage

Generate and run a Dagster job

python my_flow.py dagster create dagster_defs.py
dagster dev -f dagster_defs.py

Or execute directly in Python:

from dagster_defs import MyFlow
result = MyFlow.execute_in_process()

All graph shapes are supported

# Linear
class SimpleFlow(FlowSpec):
    @step
    def start(self):
        self.value = 42
        self.next(self.end)
    @step
    def end(self): pass

# Split/join (static branch)
class BranchFlow(FlowSpec):
    @step
    def start(self):
        self.next(self.branch_a, self.branch_b)
    ...

# Conditional (dynamic branch — only one path runs at runtime)
class ConditionalFlow(FlowSpec):
    value = Parameter("value", default=42, type=int)
    @step
    def start(self):
        self.route = "high" if self.value >= 50 else "low"
        self.next({"high": self.high_branch, "low": self.low_branch}, condition="route")
    @step
    def high_branch(self): ...
    @step
    def low_branch(self): ...
    @step
    def join(self): ...

# Foreach fan-out
class ForeachFlow(FlowSpec):
    @step
    def start(self):
        self.items = [1, 2, 3]
        self.next(self.process, foreach="items")
    ...

Parametrised flows

Parameters defined with metaflow.Parameter are forwarded automatically as a typed Dagster Config class on the start op:

python param_flow.py dagster create param_flow_dagster.py

Then pass them via Dagster's run config when launching from the UI, or via a config file when using the CLI:

# config.yaml
ops:
  op_start:
    config:
      greeting: Hi
      count: 5

python -m dagster job execute -f param_flow_dagster.py -j ParametrizedFlow -c config.yaml

Step decorators (--with)

Inject Metaflow step decorators at deploy time without modifying the flow source:

# Run every step in a sandbox (e.g. metaflow-sandbox extension)
python my_flow.py dagster create my_flow_dagster.py --with=sandbox

# Multiple decorators are supported
python my_flow.py dagster create my_flow_dagster.py \
  --with=sandbox \
  --with='resources:cpu=4,memory=8000'

Retries and timeouts

@retry and @timeout on any step are picked up automatically. The generated op gets a Dagster RetryPolicy and an op_execution_timeout tag — no extra configuration needed:

class MyFlow(FlowSpec):
    @retry(times=3, minutes_between_retries=2)
    @timeout(seconds=300)
    @step
    def train(self):
        ...

Generates:

@op(retry_policy=RetryPolicy(max_retries=3, delay=120),
    tags={"dagster/op_execution_timeout": "300"})
def op_train(context): ...

Each Dagster retry passes the correct --retry-count to Metaflow so attempt numbering is consistent.

Environment variables

@environment(vars={...}) on a step passes those variables to the metaflow step subprocess:

@environment(vars={"TOKENIZERS_PARALLELISM": "false"})
@step
def embed(self): ...

Project namespace

If the flow uses @project(name=...), the Dagster job name is automatically prefixed:

@project(name="recommendations")
class TrainFlow(FlowSpec): ...

python train_flow.py dagster create out.py
# job name: recommendations_TrainFlow

Workflow timeout

Cap the total wall-clock time for the entire job run:

python my_flow.py dagster create my_flow_dagster.py --workflow-timeout 3600

Attach tags

Metaflow tags are forwarded to every metaflow step subprocess at compile time:

python my_flow.py dagster create my_flow_dagster.py --tag env:prod --tag version:2

Custom job name

python my_flow.py dagster create my_flow_dagster.py --name nightly_pipeline

Resource hints

@resources on a step forwards CPU, memory, and GPU hints to the underlying compute backend (e.g. @kubernetes, @batch, @sandbox) via --with=resources:cpu=N,memory=M,gpu=G:

class MyFlow(FlowSpec):
    @resources(cpu=4, memory=8000, gpu=1)
    @step
    def train(self):
        ...

The resource hints are also visible as Dagster op tags in the UI.

Event-driven sensors (@trigger / @trigger_on_finish)

Decorate your flow with @trigger or @trigger_on_finish to emit a SensorDefinition in the generated file automatically.

# Fire this flow when a named event is detected
@trigger(event="data.ready")
class MyFlow(FlowSpec):
    ...

# Fire this flow when UpstreamFlow completes successfully in Dagster
@trigger_on_finish(flow="UpstreamFlow")
class DownstreamFlow(FlowSpec):
    ...

The @trigger_on_finish sensor polls the Dagster run history and yields a RunRequest for each new successful run of the upstream job. The @trigger(event=...) sensor emits a stub with a TODO comment — wire it to your event source (webhook, message queue, etc.).

Both sensors are auto-included in the Definitions object:

defs = Definitions(jobs=[MyFlow], schedules=[...], sensors=[MyFlow_on_finish_0])

Resume a failed run

Re-run a failed Dagster job, skipping steps that already completed:

python my_flow.py dagster resume --run-id <dagster-run-id>

Pass --definitions-file if you want to write the resume definitions to a file first:

python my_flow.py dagster resume --run-id <dagster-run-id> \
    --definitions-file my_flow_resume.py

The resumed run reuses the original Metaflow run ID via --clone-run-id, so all completed step outputs are available under the same pathspec.

Configuration

Metadata service and datastore

By default, metaflow-dagster uses whatever metadata and datastore backends are active in your Metaflow environment. The generated file bakes in those settings at creation time so every step subprocess uses the same backend.

To use a remote metadata service or object store, configure them before running dagster create:

python my_flow.py \
  --metadata=service \
  --datastore=s3 \
  dagster create my_flow_dagster.py

Or via environment variables:

export METAFLOW_DEFAULT_METADATA=service
export METAFLOW_DEFAULT_DATASTORE=s3
python my_flow.py dagster create my_flow_dagster.py

Scheduling

If your flow has a @schedule decorator, the generated file includes a ScheduleDefinition automatically. No extra configuration needed.

How it works

metaflow-dagster compiles your Metaflow flow's DAG into a self-contained Dagster definitions file. Each Metaflow step becomes a @op. The generated file:

• runs each step as a subprocess via the standard metaflow step CLI
• passes --input-paths correctly for joins and foreach splits
• emits Metaflow artifact keys and a retrieval snippet to the Dagster UI after each step
• forwards @resources hints to the compute backend via --with=resources:...
• emits SensorDefinitions for @trigger and @trigger_on_finish decorators

Job graph

The compiled DAG is fully visible in Dagster — typed inputs, fan-out branches, and fan-in joins:

Launchpad

Parametrised flows get a typed config schema in the Dagster launchpad, populated from your Metaflow Parameter defaults:

Run timeline

Each Metaflow step appears as a Dagster op with real wall-clock timing. Parallel branches run concurrently:

Artifact retrieval

After each step, the op emits the artifact keys and a ready-to-copy retrieval snippet — without loading the values themselves:

from metaflow import Task
task = Task('BranchingFlow/dagster-d75a08c398a3/start/1')
task['value'].data

Step logs

Every op logs the exact metaflow step CLI command it ran. Flow print() output streams through Dagster's log panel:

Development

git clone https://github.com/npow/metaflow-dagster.git
cd metaflow-dagster
pip install -e ".[test]"

# Fast compilation tests only (seconds)
pytest -v -m "not e2e"

# Full end-to-end suite (compiles + runs via `dagster job execute`)
pytest -v -m e2e

The e2e tests compile each flow to a real Dagster definitions file, execute it via dagster job execute against a temporary SQLite-backed Dagster instance, and verify Metaflow artifacts on disk. No mocks, no webserver required.

License

Apache 2.0