Sensors#

Sensors allow you to instigate runs based on any external state change.

Relevant APIs#

Overview#

Sensors are definitions in Dagster that allow you to instigate runs based on some external state change. For example, you can:

• Launch a run whenever a file appears in an s3 bucket
• Launch a run whenever another job materializes a specific asset
• Launch a run whenever an external system is down

A sensor defines an evaluation function that returns either:

• One or more RunRequest objects. Each run request launches a run.
• An optional SkipReason, which specifies a message which describes why no runs were requested.

The Dagster Daemon runs each sensor evaluation function on a tight loop. If you are using sensors, make sure to follow the instructions on the Dagster Daemon page for how to run your sensors.

Defining a sensor#

To define a sensor, use the @sensor decorator. The decorated function can optionally have a context as the first argument. The context is a SensorEvaluationContext.

Let's say you have a job that logs a filename that is specified in the op configuration of the process_file op:

from dagster import op, job


@op(config_schema={"filename": str})
def process_file(context):
    filename = context.op_config["filename"]
    context.log.info(filename)


@job
def log_file_job():
    process_file()

You can write a sensor that watches for new files in a specific directory and yields a RunRequest for each new file in the directory. By default, this sensor runs every 30 seconds.

import os
from dagster import sensor, RunRequest


@sensor(job=log_file_job)
def my_directory_sensor():
    for filename in os.listdir(MY_DIRECTORY):
        filepath = os.path.join(MY_DIRECTORY, filename)
        if os.path.isfile(filepath):
            yield RunRequest(
                run_key=filename,
                run_config={
                    "ops": {"process_file": {"config": {"filename": filename}}}
                },
            )

This sensor iterates through all the files in MY_DIRECTORY and yields a RunRequest for each file. Note that despite the yield syntax, the function will run to completion before any runs are submitted.

To write a sensor that materializes assets, you can build a job that materializes assets:

asset_job = define_asset_job("asset_job", "*")


@sensor(job=asset_job)
def materializes_asset_sensor():
    yield RunRequest(...)

Once a sensor is added to a repository with the job it yields a RunRequest for, it can be started and will start creating runs. You can start or stop sensors in Dagit, or by setting the default status to DefaultSensorStatus.RUNNING in code:

@sensor(job=asset_job, default_status=DefaultSensorStatus.RUNNING)
def my_running_sensor():
    ...

If you manually start or stop a sensor in Dagit, that will override any default status that is set in code.

Once your sensor is started, if you're running the dagster-daemon process as part of your deployment, the sensor will begin executing immediately, without needing to restart the dagster-daemon process.

Asset sensors#

A useful pattern is to create a sensor that checks for new AssetMaterialization events for a particular asset key. This can be used to kick off a job that computes downstream assets or notifies appropriate stakeholders.

One benefit of this pattern is that it enables cross-job and even cross-repository dependencies. Each job run instigated by an asset sensor is agnostic to the job that caused it.

Dagster provides a special asset sensor definition format for sensors that fire a single RunRequest based on a single asset materialization. Here is an example of a sensor that generates a RunRequest for every materialization for the asset key my_table:

from dagster import AssetKey, asset_sensor


@asset_sensor(asset_key=AssetKey("my_table"), job=my_job)
def my_asset_sensor(context, asset_event):
    yield RunRequest(
        run_key=context.cursor,
        run_config={
            "ops": {
                "read_materialization": {
                    "config": {
                        "asset_key": asset_event.dagster_event.asset_key.path,
                    }
                }
            }
        },
    )

Multi-asset sensors#

Multi-asset sensors, which can trigger job executions based on some combination of states from multiple asset materialization event streams, can be handled using the base sensor definition and manual cursor management based on the asset event streams. These asset event streams can be queried using the instance attribute off of the SensorEvaluationContext object.

import json
from dagster import EventRecordsFilter, DagsterEventType


@sensor(job=my_job)
def multi_asset_sensor(context):
    cursor_dict = json.loads(context.cursor) if context.cursor else {}
    a_cursor = cursor_dict.get("a")
    b_cursor = cursor_dict.get("b")

    a_event_records = context.instance.get_event_records(
        EventRecordsFilter(
            event_type=DagsterEventType.ASSET_MATERIALIZATION,
            asset_key=AssetKey("table_a"),
            after_cursor=a_cursor,
        ),
        ascending=False,
        limit=1,
    )
    b_event_records = context.instance.get_event_records(
        EventRecordsFilter(
            event_type=DagsterEventType.ASSET_MATERIALIZATION,
            asset_key=AssetKey("table_b"),
            after_cursor=b_cursor,
        ),
        ascending=False,
        limit=1,
    )

    if not a_event_records or not b_event_records:
        return

    # make sure we only generate events if both table_a and table_b have been materialized since
    # the last evaluation.
    yield RunRequest(run_key=None)

    # update the sensor cursor by combining the individual event cursors from the two separate
    # asset event streams
    context.update_cursor(
        json.dumps(
            {
                "a": a_event_records[0].storage_id,
                "b": b_event_records[0].storage_id,
            }
        )
    )

Idempotence and Cursors#

When instigating runs based on external events, you usually want to run exactly one job run for each event. There are two ways to define your sensors to avoid creating duplicate runs for your events: using run_key and using a cursor.

Idempotence using run keys#

In the example sensor above, the RunRequest is constructed with a run_key.

yield RunRequest(
    run_key=filename,
    run_config={"ops": {"process_file": {"config": {"filename": filename}}}},
)

Dagster guarantees that for a given sensor, at most one run is created for each RunRequest with a unique run_key. If a sensor yields a new run request with a previously used run_key, Dagster skips processing the new run request.

In the example, a RunRequest is requested for each file during every sensor evaluation. Therefore, for a given sensor evaluation, there already exists a RunRequest with a run_key for any file that existed during the previous sensor evaluation. Dagster skips processing duplicate run requests, so Dagster launches runs for only the files added since the last sensor evaluation. The result is exactly one run per file.

Run keys allow you to write sensor evaluation functions that declaratively describe what job runs should exist, and helps you avoid the need for more complex logic that manages state. However, when dealing with high-volume external events, some state-tracking optimizations might be necessary.

Sensor optimizations using cursors#

When writing a sensor that deals with high-volume events, it might not be feasible to yield a RunRequest during every sensor evaluation. For example, you may have an s3 storage bucket that contains thousands of files.

When writing a sensor for such event sources, you can maintain a cursor that limits the number of yielded run requests for previously processed events. The sensor context, provided to every sensor evaluation function, has a cursor property and a update_cursor method for sensors to track state across evaluations.

• cursor: A cursor field on SensorEvaluationContext that returns the last persisted cursor value from a previous evaluation.
• update_cursor: A method on SensorEvaluationContext that takes a string to persist and make available to future evaluations.

Here is a somewhat contrived example of our directory file sensor using a cursor for updated files.

@sensor(job=log_file_job)
def my_directory_sensor_cursor(context):
    last_mtime = float(context.cursor) if context.cursor else 0

    max_mtime = last_mtime
    for filename in os.listdir(MY_DIRECTORY):
        filepath = os.path.join(MY_DIRECTORY, filename)
        if os.path.isfile(filepath):
            fstats = os.stat(filepath)
            file_mtime = fstats.st_mtime
            if file_mtime <= last_mtime:
                continue

            # the run key should include mtime if we want to kick off new runs based on file modifications
            run_key = f"{filename}:{str(file_mtime)}"
            run_config = {"ops": {"process_file": {"config": {"filename": filename}}}}
            yield RunRequest(run_key=run_key, run_config=run_config)
            max_mtime = max(max_mtime, file_mtime)

    context.update_cursor(str(max_mtime))

For sensors that consume multiple event streams, you may need to serialize and deserialize a more complex data structure in and out of the cursor string to keep track of the sensor's progress over the multiple streams.

Evaluation Interval#

By default, the Dagster Daemon runs a sensor 30 seconds after that sensor's previous evaluation finishes executing. You can configure the interval using the minimum_interval_seconds argument on the @sensor decorator.

It's important to note that this interval represents a minimum interval between runs of the sensor and not the exact frequency the sensor runs. If you have a sensor that takes 30 seconds to complete, but the minimum_interval_seconds is 5 seconds, the fastest Dagster Daemon will run the sensor is every 35 seconds. The minimum_interval_seconds only guarantees that the sensor is not evaluated more frequently than the given interval.

For example, here are two sensors that specify two different minimum intervals:

@sensor(job=my_job, minimum_interval_seconds=30)
def sensor_A():
    yield RunRequest(run_key=None, run_config={})


@sensor(job=my_job, minimum_interval_seconds=45)
def sensor_B():
    yield RunRequest(run_key=None, run_config={})

These sensor definitions are short, so they run in less than a second. Therefore, you can expect these sensors to run consistently around every 30 and 45 seconds, respectively.

If a sensor evaluation function takes more than 60 seconds to return its results, the sensor evaluation will time out and the Dagster Daemon will move on to the next sensor without submitting any runs. This 60 second timeout only applies to the time it takes to run the sensor function, not to the execution time of the runs submitted by the sensor. To avoid timeouts, slower sensors can break up their work into chunks, using cursors to let subsequent sensor calls pick up where the previous call left off.

Skipping sensor evaluations#

For debugging purposes, it is often useful to describe why a sensor might not yield any runs for a given evaluation. The sensor evaluation function can yield a SkipReason with a string description that will be displayed in Dagit.

For example, here is our directory sensor that now provides a SkipReason when no files are encountered:

@sensor(job=log_file_job)
def my_directory_sensor_with_skip_reasons():
    has_files = False
    for filename in os.listdir(MY_DIRECTORY):
        filepath = os.path.join(MY_DIRECTORY, filename)
        if os.path.isfile(filepath):
            yield RunRequest(
                run_key=filename,
                run_config={
                    "ops": {"process_file": {"config": {"filename": filename}}}
                },
            )
            has_files = True
    if not has_files:
        yield SkipReason(f"No files found in {MY_DIRECTORY}.")

Testing sensors#

To quickly preview what an existing sensor will generate when evaluated, you can run the CLI command dagster sensor preview my_sensor_name.

In order to unit test sensors, you can invoke the sensor directly. This will return all the run requests yielded by the sensor. The config obtained from these can be validated using the validate_run_config function.

from dagster import validate_run_config


@sensor(job=log_file_job)
def sensor_to_test():
    yield RunRequest(
        run_key="foo",
        run_config={"ops": {"process_file": {"config": {"filename": "foo"}}}},
    )


def test_sensor():
    for run_request in sensor_to_test():
        assert validate_run_config(log_file_job, run_request.run_config)

Notice that since we did not use the context argument in our sensor, we don't have to provide a context object. However, if we do in fact need the context object for our sensor, we can provide it via build_sensor_context. Consider again the my_directory_sensor_cursor example.

@sensor(job=log_file_job)
def my_directory_sensor_cursor(context):
    last_mtime = float(context.cursor) if context.cursor else 0

    max_mtime = last_mtime
    for filename in os.listdir(MY_DIRECTORY):
        filepath = os.path.join(MY_DIRECTORY, filename)
        if os.path.isfile(filepath):
            fstats = os.stat(filepath)
            file_mtime = fstats.st_mtime
            if file_mtime <= last_mtime:
                continue

            # the run key should include mtime if we want to kick off new runs based on file modifications
            run_key = f"{filename}:{str(file_mtime)}"
            run_config = {"ops": {"process_file": {"config": {"filename": filename}}}}
            yield RunRequest(run_key=run_key, run_config=run_config)
            max_mtime = max(max_mtime, file_mtime)

    context.update_cursor(str(max_mtime))

This sensor makes use of the context argument, and thus to invoke it, we need to provide one.

from dagster import build_sensor_context


def test_my_directory_sensor_cursor():
    context = build_sensor_context(cursor="0")
    for run_request in my_directory_sensor_cursor(context):
        assert validate_run_config(log_file_job, run_request.run_config)

Monitoring sensors in Dagit#

You can monitor and operate sensors in Dagit. There are multiple views that help with observing sensor evaluations, skip reasons, and errors.

To view the sensors page, you can navigate to the "Sensors" tab from the Repository page. Here you can start and stop sensors using the toggle.

If you click on any sensor, you can monitor all sensor evaluations on a timeline and view a table of runs launched by the sensor:

Run status sensors#

If you want to act on the status of a job run, Dagster provides a way to create a sensor that reacts to run statuses. You can use run_status_sensor with a specified DagsterRunStatus to decorate a function that will run when the given status occurs. This can be used to send alerts to a monitoring service on job failure or report a run success.

Here is an example of a sensor that reports job success:

from dagster import run_status_sensor, RunStatusSensorContext, DagsterRunStatus


@run_status_sensor(pipeline_run_status=DagsterRunStatus.SUCCESS)
def my_slack_on_run_success(context: RunStatusSensorContext):
    slack_client = WebClient(token=os.environ["SLACK_DAGSTER_ETL_BOT_TOKEN"])

    slack_client.chat_postMessage(
        channel="#alert-channel",
        message=f'Job "{context.pipeline_run.pipeline_name}" succeeded.',
    )

Then, you can add the sensor to a repository so it can be enabled and used the same as other sensors:

from dagster import repository


@repository
def my_repository():
    return my_jobs + [my_slack_on_run_success]

Job failure sensor#

Dagster provides a set of special run status sensor decorators for defining sensors that monitor job failure events. You can use run_failure_sensor to decorate a function that will run when a run fails.

For example, you can write a sensor that sends a slack message when it runs using this decorator:

import os
from dagster import run_failure_sensor, RunFailureSensorContext
from slack_sdk import WebClient


@run_failure_sensor
def my_slack_on_run_failure(context: RunFailureSensorContext):
    slack_client = WebClient(token=os.environ["SLACK_DAGSTER_ETL_BOT_TOKEN"])

    slack_client.chat_postMessage(
        channel="#alert-channel",
        message=f'Job "{context.pipeline_run.pipeline_name}" failed. Error: {context.failure_event.message}',
    )

Dagster also provides the following out-of-box job failure sensors:

from dagster_slack import make_slack_on_run_failure_sensor

slack_on_run_failure = make_slack_on_run_failure_sensor(
    "#my_channel", os.getenv("MY_SLACK_TOKEN")
)

from dagster import make_email_on_run_failure_sensor


email_on_run_failure = make_email_on_run_failure_sensor(
    email_from="no-reply@example.com",
    email_password=os.getenv("ALERT_EMAIL_PASSWORD"),
    email_to=["xxx@example.com", "xyz@example.com"],
)

Besides, if you would like to set up success or failure handling policies on ops, you can find more information on the Op Hooks page.

Testing run status sensors#

As with other sensors, you can directly invoke run status sensors. However, the context provided via run_status_sensor and run_failure_sensor contain objects that are typically only available during run time. Below you'll find code snippets that demonstrate how to build the context so that you can directly invoke your function in unit tests. Feel free to copy the following into your unit tests.

If you had written a status sensor like this (assuming you implemented the function email_alert elsewhere)

@run_status_sensor(pipeline_run_status=DagsterRunStatus.SUCCESS)
def my_email_sensor(context: RunStatusSensorContext):
    message = f'Job "{context.pipeline_run.pipeline_name}" succeeded.'
    email_alert(message)

We can first write a simple job that will succeed

@op
def succeeds():
    return 1


@job
def my_job_succeeds():
    succeeds()

Then we can execute this job and pull the attributes we need to build the context. We provide a function build_run_status_sensor_context that will return the correct context object

# execute the job
instance = DagsterInstance.ephemeral()
result = my_job_succeeds.execute_in_process(instance=instance)

# retrieve the DagsterRun
dagster_run = result.dagster_run

# retrieve a success event from the completed execution
dagster_event = result.get_job_success_event()

# create the context
run_status_sensor_context = build_run_status_sensor_context(
    sensor_name="my_email_sensor",
    dagster_instance=instance,
    dagster_run=dagster_run,
    dagster_event=dagster_event,
)

# run the sensor
my_email_sensor(run_status_sensor_context)

We have provided convenience functions ExecuteInProcessResult.get_job_success_event and ExecuteInProcessResult.get_job_failure_event for retrieving DagsterRunStatus.SUCCESS and DagsterRunStatus.FAILURE events, respectively. If you have a run status sensor triggered on another status, you can retrieve all events from result and filter based on your event type.

We can use the same pattern to build the context for run_failure_sensor. If we wanted to test this run failure sensor

@run_failure_sensor
def my_email_failure_sensor(context: RunFailureSensorContext):
    message = f'Job "{context.pipeline_run.pipeline_name}" failed. Error: {context.failure_event.message}'
    email_alert(message)

We first need to make a simple job that will fail

from dagster import op, job


@op
def fails():
    raise Exception("failure!")


@job
def my_job_fails():
    fails()

Then we can execute the job and create our context

from dagster import DagsterInstance, build_run_status_sensor_context

# execute the job
instance = DagsterInstance.ephemeral()
result = my_job_fails.execute_in_process(instance=instance, raise_on_error=False)

# retrieve the DagsterRun
dagster_run = result.dagster_run

# retrieve a failure event from the completed job execution
dagster_event = result.get_job_failure_event()

# create the context
run_failure_sensor_context = build_run_status_sensor_context(
    sensor_name="my_email_failure_sensor",
    dagster_instance=instance,
    dagster_run=dagster_run,
    dagster_event=dagster_event,
).for_run_failure()

# run the sensor
my_email_failure_sensor(run_failure_sensor_context)

Note the additional function call RunStatusSensorContext.for_run_failure after creating the context. The context provided by run_failure_sensor is a subclass of the context provided by run_status_sensor and can be built using this additional call.

Examples#

S3 sensors#

For jobs that should initiate new runs for new paths in an s3 bucket, the dagster-aws package provides the useful helper function get_s3_keys.

Here is an example of a sensor that listens to a particular s3 bucket my_s3_bucket:

from dagster_aws.s3.sensor import get_s3_keys


@sensor(job=my_job)
def my_s3_sensor(context):
    since_key = context.cursor or None
    new_s3_keys = get_s3_keys("my_s3_bucket", since_key=since_key)
    if not new_s3_keys:
        return SkipReason("No new s3 files found for bucket my_s3_bucket.")
    last_key = new_s3_keys[-1]
    run_requests = [RunRequest(run_key=s3_key, run_config={}) for s3_key in new_s3_keys]
    context.update_cursor(last_key)
    return run_requests

Using resources in sensors#

If you want to use resources within your sensor, you can use the build_resources API to perform the initialization.

from dagster import resource, build_resources, sensor


@resource
def the_credentials():
    ...


@resource(required_resource_keys={"credentials"})
def the_db_connection(init_context):
    get_the_db_connection(init_context.resources.credentials)


@sensor(job=the_job)
def uses_db_connection():
    with build_resources(
        {"db_connection": the_db_connection, "credentials": the_credentials}
    ) as resources:
        conn = resources.db_connection
        ...

If a resource you want to initialize has dependencies on other resources, those can be included in the dictionary passed to build_resources. For more in-depth usage, check out the Initializing Resources Outside of Execution section.

See it in action#

For more examples of sensors, check out the following in our Hacker News example:

• Sensor factory