Within the body of an op, it is possible to communicate with the Dagster framework either by yielding an event, logging an event, or raising an exception. This page describes these different possibilities and the scenarios in which you might use them.
Within the body of an op, a stream of structured events can be yielded or logged. These events will be processed by Dagster and recorded in the event log, along with some additional context about the op that emitted it.
It is also possible to raise Dagster-specific exceptions, which will indicate to the framework to halt the op execution entirely and perform some action.
Often, it may be useful to attach some arbitrary information to an event or exception that is not captured by its basic parameters. Through the MetadataEntry and MetadataValue classes, we provide a consistent interface for specifying this metadata as a set of "entries". Individual entries are instances of the MetadataEntry class, which is a simple container for a string label and a typed value. The available value types are accessible through a static API defined on MetadataValue. These include simple datatypes (MetadataValue.float, MetadataValue.int, MetadataValue.text), as well as more complex information such as markdown and json (MetadataValue.md, MetadataValue.json). Depending on the type of its value, a MetadataEntry will be rendered in Dagit in a more useful format than a simple unstructured string.
Metadata is attached to events at construction time. There are currently two available metadata specification keyword arguments on all event constructors: (1) metadata takes a dictionary mapping string labels to MetadataValue objects. Under the hood, this is resolved to a list of MetadataEntry objects. (2) metadata_entries takes a list of MetadataEntry objects directly. The metadata_entries argument is deprecated-- new code should use the metadata dictionary form.
Yielding events from within the body of an op is a useful way of communicating with the Dagster framework. The most critical event to the functionality of Dagster is the Output event, which allows output data to be passed on from one op to the next. However, we also provide interfaces to inform Dagster about external assets and data quality checks during the run of an op.
Because returning a value from an op is such a fundamental part of creating a data pipeline, we have a few different interfaces for this functionality.
For many use cases, Dagster ops can be used directly with python's native type annotations without additional modification. Check out the docs on Op Outputs to learn more about this functionality. Dagster also provides the Output object, which opens up additional functionality to outputs when using Dagster, such as specifying output metadata and conditional branching, all while maintaining coherent type annotations.
Output objects can be either returned or yielded. The Output type is also generic, for use with return annotations:
from dagster import Output, op
from typing import Tuple
# Using Output as type annotation without inner type@opdefmy_output_op()-> Output:return Output("some_value")# A single output with a parameterized type annotation@opdefmy_output_generic_op()-> Output[int]:return Output(5)# Multiple outputs using parameterized type annotation@op(out={"int_out": Out(),"str_out": Out()})defmy_multiple_generic_output_op()-> Tuple[Output[int], Output[str]]:return(Output(5), Output("foo"))
When Output objects are yielded, type annotations cannot be used. Instead, type information can be specified using the out argument of the op decorator.
from dagster import Output, op
@op(out={"out1": Out(str),"out2": Out(int)})defmy_op_yields():yield Output(5, output_name="out2")yield Output("foo", output_name="out1")
If there is information specific to an op output that you would like to log, you can use an Output object to attach metadata to the op's output. To do this, use the metadata parameter on the object, which expects a mapping of string labels to metadata values.
The EventMetadata class contains a set of static wrappers to customize the display of certain types of structured metadata.
The following example demonstrates how you might use this functionality:
from dagster import MetadataValue, Output, op
@opdefmy_metadata_output(context)-> Output:
df = get_some_data()return Output(
df,
metadata={"text_metadata":"Text-based metadata for this event","dashboard_url": MetadataValue.url("http://mycoolsite.com/url_for_my_data"),"raw_count":len(df),"size (bytes)": calculate_bytes(df),},)
AssetMaterialization events tell Dagster that you have written some data asset to an external system. The classic example would be writing to a table in a database, but really any sort of persisted object that you would want to keep track of can be considered an asset.
Generally, you'd want to send this event directly after you persist the asset to your external system. All AssetMaterialization events must define an asset_key, which is a unique identifier to describe the asset you are persisting.
from dagster import AssetMaterialization, op
@opdefmy_asset_op(context):
df = get_some_data()
store_to_s3(df)
context.log_event(
AssetMaterialization(
asset_key="s3.my_asset",
description="A df I stored in s3",))
result = do_some_transform(df)return result
Asset materializations can also be yielded:
from dagster import AssetMaterialization, Output, op
@opdefmy_asset_op_yields():
df = get_some_data()
store_to_s3(df)yield AssetMaterialization(
asset_key="s3.my_asset",
description="A df I stored in s3",)
result = do_some_transform(df)yield Output(result)
When yielding asset materializations, outputs must also be yielded via an Output.
To learn more about assets and how they are surfaced once you send this event, check out the Asset Catalog documentation.
Attaching metadata to Asset Materializations is an important way of tracking aspects of a given asset over time. This functions essentially identically to other events which accept a metadata parameter, allowing you to attach a set of structured labels and values to display.
from dagster import op, AssetMaterialization, MetadataValue
@opdefmy_metadata_materialization_op(context):
df = read_df()
remote_storage_path = persist_to_storage(df)
context.log_event(
AssetMaterialization(
asset_key="my_dataset",
description="Persisted result to storage",
metadata={"text_metadata":"Text-based metadata for this event","path": MetadataValue.path(remote_storage_path),"dashboard_url": MetadataValue.url("http://mycoolsite.com/url_for_my_data"),"size (bytes)": calculate_bytes(df),},))return remote_storage_path
AssetObservation events record metadata about assets. Unlike asset materializations, asset observations do not signify that an asset has been mutated.
Within ops and assets, you can log or yield AssetObservation events at runtime. Similar to attaching metadata to asset materializations, asset observations accept a metadata parameter, allowing you to track specific properties of an asset over time.
from dagster import AssetObservation, op
@opdefobservation_op(context):
df = read_df()
context.log_event(
AssetObservation(asset_key="observation_asset", metadata={"num_rows":len(df)}))return5
In the example above, an observation tracks the number of rows in an asset persisted to storage. This information can then be viewed on the Asset Details page.
To learn more about asset observations, check out the Asset Observation documentation.
Ops can emit structured events to represent the results of a data quality test. The data quality event class is the ExpectationResult. To generate an expectation result, we can log or yield an ExpectationResult event in our op.
from dagster import ExpectationResult, op
@opdefmy_expectation_op(context, df):
do_some_transform(df)
context.log_event(
ExpectationResult(success=len(df)>0, description="ensure dataframe has rows"))return df
Like many other event types in Dagster, there are a variety of types of metadata that can be associated with an expectation result event, all through the MetadataEntry class. Each expectation event optionally takes a list of metadata entries that are then displayed in the event log.
This example shows metadata entries of different types attached to the same expectation result:
from dagster import ExpectationResult, MetadataValue, op
@opdefmy_metadata_expectation_op(context, df):
df = do_some_transform(df)
context.log_event(
ExpectationResult(
success=len(df)>0,
description="ensure dataframe has rows",
metadata={"text_metadata":"Text-based metadata for this event","dashboard_url": MetadataValue.url("http://mycoolsite.com/url_for_my_data"),"raw_count":len(df),"size (bytes)": calculate_bytes(df),},))return df
Dagster also provides some op-specific exception classes, which can be raised to halt the execution of a op. The behavior after an exception is raised depends on the exception that you use. The exceptions are documented below.
Failure exceptions may be raised to indicate that an op has failed in an unrecoverable way, and that execution should stop. This may be useful if, in the body of your op, you detect an issue that would not raise an error on its own, but would still cause problems if the execution were to continue.
from dagster import Failure, op
@opdefmy_failure_op():
path ="/path/to/files"
my_files = get_files(path)iflen(my_files)==0:raise Failure(
description="No files to process",
metadata={"filepath": MetadataValue.path(path),"dashboard_url": MetadataValue.url("http://mycoolsite.com/failures"),},)return some_calculation(my_files)
RetryRequested exceptions are useful when you experience failures that are possible to recover from. For example, if you have a flaky operation that you expect to throw an exception once in a while, you can catch the exception and throw a RetryRequested to make Dagster halt and re-start op execution.
You can configure the number of retries to be attempted with the max_retries parameter.
from dagster import RetryRequested, op
@opdefmy_retry_op():try:
result = flaky_operation()except Exception as e:raise RetryRequested(max_retries=3)from e
return result