Ops

The foundational unit of computation in Dagster.

Defining ops

@dagster.op(compute_fn=None, *, name=None, description=None, ins=None, out=None, config_schema=None, required_resource_keys=None, tags=None, version=None, retry_policy=None)[source]

Create an op with the specified parameters from the decorated function.

Ins and outs will be inferred from the type signature of the decorated function if not explicitly provided.

The decorated function will be used as the op’s compute function. The signature of the decorated function is more flexible than that of the compute_fn in the core API; it may:

  1. Return a value. This value will be wrapped in an Output and yielded by the compute function.

  2. Return an Output. This output will be yielded by the compute function.

  3. Yield Output or other event objects. Same as default compute behavior.

Note that options 1) and 2) are incompatible with yielding other events – if you would like to decorate a function that yields events, it must also wrap its eventual output in an Output and yield it.

@op supports async def functions as well, including async generators when yielding multiple events or outputs. Note that async ops will generally be run on their own unless using a custom Executor implementation that supports running them together.

Parameters:

  • name (Optional[str]) – Name of op. Must be unique within any GraphDefinition using the op.

  • description (Optional[str]) – Human-readable description of this op. If not provided, and the decorated function has docstring, that docstring will be used as the description.

  • ins (Optional[Dict[str, In]]) – Information about the inputs to the op. Information provided here will be combined with what can be inferred from the function signature.

  • out (Optional[Union[Out, Dict[str, Out]]]) – Information about the op outputs. Information provided here will be combined with what can be inferred from the return type signature if the function does not use yield.

  • config_schema (Optional[ConfigSchema) – The schema for the config. If set, Dagster will check that config provided for the op matches this schema and fail if it does not. If not set, Dagster will accept any config provided for the op.

  • required_resource_keys (Optional[Set[str]]) – Set of resource handles required by this op.

  • tags (Optional[Dict[str, Any]]) – Arbitrary metadata for the op. Frameworks may expect and require certain metadata to be attached to a op. Values that are not strings will be json encoded and must meet the criteria that json.loads(json.dumps(value)) == value.

  • version (Optional[str]) – (Experimental) The version of the op’s compute_fn. Two ops should have the same version if and only if they deterministically produce the same outputs when provided the same inputs.

  • retry_policy (Optional[RetryPolicy]) – The retry policy for this op.

Examples

@op
def hello_world():
    print('hello')

@op
def echo(msg: str) -> str:
    return msg

@op(
    ins={'msg': In(str)},
    out=Out(str)
)
def echo_2(msg): # same as above
    return msg

@op(
    out={'word': Out(), 'num': Out()}
)
def multi_out() -> Tuple[str, int]:
    return 'cool', 4
class dagster.OpDefinition(compute_fn, name, ins=None, outs=None, description=None, config_schema=None, required_resource_keys=None, tags=None, version=None, retry_policy=None)[source]

Defines an op, the functional unit of user-defined computation.

For more details on what a op is, refer to the Ops Overview .

End users should prefer the @op decorator. OpDefinition is generally intended to be used by framework authors or for programatically generated ops.

Parameters:

  • name (str) – Name of the op. Must be unique within any GraphDefinition or JobDefinition that contains the op.

  • input_defs (List[InputDefinition]) – Inputs of the op.

  • compute_fn (Callable) –

    The core of the op, the function that performs the actual computation. The signature of this function is determined by input_defs, and optionally, an injected first argument, context, a collection of information provided by the system.

    This function will be coerced into a generator or an async generator, which must yield one Output for each of the op’s output_defs, and additionally may yield other types of Dagster events, including AssetMaterialization and ExpectationResult.

  • output_defs (List[OutputDefinition]) – Outputs of the op.

  • config_schema (Optional[ConfigSchema) – The schema for the config. If set, Dagster will check that the config provided for the op matches this schema and will fail if it does not. If not set, Dagster will accept any config provided for the op.

  • description (Optional[str]) – Human-readable description of the op.

  • tags (Optional[Dict[str, Any]]) – Arbitrary metadata for the op. Frameworks may expect and require certain metadata to be attached to a op. Users should generally not set metadata directly. Values that are not strings will be json encoded and must meet the criteria that json.loads(json.dumps(value)) == value.

  • required_resource_keys (Optional[Set[str]]) – Set of resources handles required by this op.

  • version (Optional[str]) – (Experimental) The version of the op’s compute_fn. Two ops should have the same version if and only if they deterministically produce the same outputs when provided the same inputs.

  • retry_policy (Optional[RetryPolicy]) – The retry policy for this op.

Examples

def _add_one(_context, inputs):
    yield Output(inputs["num"] + 1)

OpDefinition(
    name="add_one",
    ins={"num": In(int)},
    outs={"result": Out(int)},
    compute_fn=_add_one,
)

Ins & outs

class dagster.In(dagster_type=<class 'dagster._core.definitions.utils.NoValueSentinel'>, description=None, default_value=<class 'dagster._core.definitions.utils.NoValueSentinel'>, root_manager_key=None, metadata=None, asset_key=None, asset_partitions=None, input_manager_key=None)[source]

Defines an argument to an op’s compute function.

Inputs may flow from previous op’s outputs, or be stubbed using config. They may optionally be typed using the Dagster type system.

Parameters:

  • dagster_type (Optional[Union[Type, DagsterType]]]) – The type of this input. Should only be set if the correct type can not be inferred directly from the type signature of the decorated function.

  • description (Optional[str]) – Human-readable description of the input.

  • default_value (Optional[Any]) – The default value to use if no input is provided.

  • root_manager_key (Optional[str]) – (Experimental) The resource key for the RootInputManager used for loading this input when it is not connected to an upstream output.

  • metadata (Optional[Dict[str, RawMetadataValue]]) – A dict of metadata for the input.

  • asset_key (Optional[Union[AssetKey, InputContext -> AssetKey]]) – (Experimental) An AssetKey (or function that produces an AssetKey from the InputContext) which should be associated with this In. Used for tracking lineage information through Dagster.

  • asset_partitions (Optional[Union[Set[str], InputContext -> Set[str]]]) – (Experimental) A set of partitions of the given asset_key (or a function that produces this list of partitions from the InputContext) which should be associated with this In.

class dagster.Out(dagster_type=<class 'dagster._core.definitions.utils.NoValueSentinel'>, description=None, is_required=True, io_manager_key=None, metadata=None, asset_key=None, asset_partitions=None, asset_partitions_def=None)[source]

Defines an output from an op’s compute function.

Ops can have multiple outputs, in which case outputs cannot be anonymous.

Many ops have only one output, in which case the user can provide a single output definition that will be given the default name, “result”.

Outs may be typed using the Dagster type system.

Parameters:

  • dagster_type (Optional[Union[Type, DagsterType]]]) – The type of this output. Should only be set if the correct type can not be inferred directly from the type signature of the decorated function.

  • description (Optional[str]) – Human-readable description of the output.

  • is_required (bool) – Whether the presence of this field is required. (default: True)

  • io_manager_key (Optional[str]) – The resource key of the output manager used for this output. (default: “io_manager”).

  • metadata (Optional[Dict[str, Any]]) – A dict of the metadata for the output. For example, users can provide a file path if the data object will be stored in a filesystem, or provide information of a database table when it is going to load the data into the table.

  • asset_key (Optional[AssetKey]) – (Experimental) An AssetKey which should be associated with this Out. Used for tracking lineage information through Dagster.

  • asset_partitions (Optional[Union[Set[str], OutputContext -> Set[str]]]) – (Experimental) A set of partitions of the given asset_key (or a function that produces this list of partitions from the OutputContext) which should be associated with this Out.

Execution

class dagster.RetryPolicy(max_retries=1, delay=None, backoff=None, jitter=None)[source]

A declarative policy for when to request retries when an exception occurs during op execution.

Parameters:

  • max_retries (int) – The maximum number of retries to attempt. Defaults to 1.

  • delay (Optional[Union[int,float]]) – The time in seconds to wait between the retry being requested and the next attempt being started. This unit of time can be modulated as a function of attempt number with backoff and randomly with jitter.

  • backoff (Optional[Backoff]) – A modifier for delay as a function of retry attempt number.

  • jitter (Optional[Jitter]) – A randomizing modifier for delay, applied after backoff calculation.

class dagster.Backoff(value)[source]

A modifier for delay as a function of attempt number.

LINEAR: attempt_num * delay EXPONENTIAL: ((2 ^ attempt_num) - 1) * delay

class dagster.Jitter(value)[source]

A randomizing modifier for delay, applied after backoff calculation.

FULL: between 0 and the calculated delay based on backoff: random() * backoff_delay PLUS_MINUS: +/- the delay: backoff_delay + ((2 * (random() * delay)) - delay)

Events

The objects that can be yielded by the body of ops’ compute functions to communicate with the Dagster framework.

(Note that Failure and RetryRequested are intended to be raised from ops rather than yielded.)

Event types

class dagster.Output(value, output_name='result', metadata_entries=None, metadata=None)[source]

Event corresponding to one of a op’s outputs.

Op compute functions must explicitly yield events of this type when they have more than one output, or when they also yield events of other types, or when defining a op using the OpDefinition API directly.

Outputs are values produced by ops that will be consumed by downstream ops in a job. They are type-checked at op boundaries when their corresponding Out or the downstream In is typed.

Parameters:

  • value (Any) – The value returned by the compute function.

  • output_name (Optional[str]) – Name of the corresponding out. (default: “result”)

  • metadata_entries (Optional[Union[MetadataEntry, PartitionMetadataEntry]]) – (Experimental) A set of metadata entries to attach to events related to this Output.

  • metadata (Optional[Dict[str, Union[str, float, int, Dict, MetadataValue]]]) – Arbitrary metadata about the failure. Keys are displayed string labels, and values are one of the following: string, float, int, JSON-serializable dict, JSON-serializable list, and one of the data classes returned by a MetadataValue static method.

class dagster.AssetMaterialization(asset_key, description=None, metadata_entries=None, partition=None, metadata=None)[source]

Event indicating that an op has materialized an asset.

Op compute functions may yield events of this type whenever they wish to indicate to the Dagster framework (and the end user) that they have produced a materialized value as a side effect of computation. Unlike outputs, asset materializations can not be passed to other ops, and their persistence is controlled by op logic, rather than by the Dagster framework.

Op authors should use these events to organize metadata about the side effects of their computations, enabling tooling like the Assets dashboard in Dagit.

Parameters:

  • asset_key (Union[str, List[str], AssetKey]) – A key to identify the materialized asset across job runs

  • description (Optional[str]) – A longer human-readable description of the materialized value.

  • metadata_entries (Optional[List[Union[MetadataEntry, PartitionMetadataEntry]]]) – Arbitrary metadata about the materialized value.

  • partition (Optional[str]) – The name of the partition that was materialized.

  • metadata (Optional[Dict[str, RawMetadataValue]]) – Arbitrary metadata about the asset. Keys are displayed string labels, and values are one of the following: string, float, int, JSON-serializable dict, JSON-serializable list, and one of the data classes returned by a MetadataValue static method.

class dagster.ExpectationResult(success, label=None, description=None, metadata_entries=None, metadata=None)[source]

Event corresponding to a data quality test.

Op compute functions may yield events of this type whenever they wish to indicate to the Dagster framework (and the end user) that a data quality test has produced a (positive or negative) result.

Parameters:

  • success (bool) – Whether the expectation passed or not.

  • label (Optional[str]) – Short display name for expectation. Defaults to “result”.

  • description (Optional[str]) – A longer human-readable description of the expectation.

  • metadata_entries (Optional[List[MetadataEntry]]) – Arbitrary metadata about the expectation.

  • metadata (Optional[Dict[str, RawMetadataValue]]) – Arbitrary metadata about the failure. Keys are displayed string labels, and values are one of the following: string, float, int, JSON-serializable dict, JSON-serializable list, and one of the data classes returned by a MetadataValue static method.

class dagster.TypeCheck(success, description=None, metadata_entries=None, metadata=None)[source]

Event corresponding to a successful typecheck.

Events of this type should be returned by user-defined type checks when they need to encapsulate additional metadata about a type check’s success or failure. (i.e., when using as_dagster_type(), @usable_as_dagster_type, or the underlying PythonObjectDagsterType() API.)

Solid compute functions should generally avoid yielding events of this type to avoid confusion.

Parameters:

  • success (bool) – True if the type check succeeded, False otherwise.

  • description (Optional[str]) – A human-readable description of the type check.

  • metadata_entries (Optional[List[MetadataEntry]]) – Arbitrary metadata about the type check.

  • metadata (Optional[Dict[str, RawMetadataValue]]) – Arbitrary metadata about the failure. Keys are displayed string labels, and values are one of the following: string, float, int, JSON-serializable dict, JSON-serializable list, and one of the data classes returned by a MetadataValue static method.

class dagster.Failure(description=None, metadata_entries=None, metadata=None, allow_retries=None)[source]

Event indicating op failure.

Raise events of this type from within op compute functions or custom type checks in order to indicate an unrecoverable failure in user code to the Dagster machinery and return structured metadata about the failure.

Parameters:

  • description (Optional[str]) – A human-readable description of the failure.

  • metadata_entries (Optional[List[MetadataEntry]]) – Arbitrary metadata about the failure.

  • metadata (Optional[Dict[str, RawMetadataValue]]) – Arbitrary metadata about the failure. Keys are displayed string labels, and values are one of the following: string, float, int, JSON-serializable dict, JSON-serializable list, and one of the data classes returned by a MetadataValue static method.

  • allow_retries (Optional[bool]) – Whether this Failure should respect the retry policy or bypass it and immediately fail. Defaults to True, respecting the retry policy and allowing retries.

class dagster.RetryRequested(max_retries=1, seconds_to_wait=None)[source]

An exception to raise from an op to indicate that it should be retried.

Parameters:

  • max_retries (Optional[int]) – The max number of retries this step should attempt before failing

  • seconds_to_wait (Optional[Union[float,int]]) – Seconds to wait before restarting the step after putting the step in to the up_for_retry state

Example

@op
def flakes():
    try:
        flakey_operation()
    except Exception as e:
        raise RetryRequested(max_retries=3) from e

Event metadata

Dagster uses metadata to communicate arbitrary user-specified metadata about structured events.

class dagster.MetadataValue[source]

Utility class to wrap metadata values passed into Dagster events so that they can be displayed in Dagit and other tooling.

@op
def emit_metadata(context, df):
    yield AssetMaterialization(
        asset_key="my_dataset",
        metadata={
            "my_text_label": "hello",
            "dashboard_url": MetadataValue.url("http://mycoolsite.com/my_dashboard"),
            "num_rows": 0,
        },
    )
class dagster.MetadataEntry(label, description=None, entry_data=None, value=None)[source]

The standard structure for describing metadata for Dagster events.

Lists of objects of this type can be passed as arguments to Dagster events and will be displayed in Dagit and other tooling.

Should be yielded from within an IO manager to append metadata for a given input/output event. For other event types, passing a dict with MetadataValue values to the metadata argument is preferred.

NOTE: MetadataEntry static constructors are deprecated. Instead you should use: MetadataEntry(<label>, value=MetadataValue.<type>(<value>)).

Parameters:

  • label (str) – Short display label for this metadata entry.

  • description (Optional[str]) – A human-readable description of this metadata entry.

  • value (MetadataValue) – Typed metadata entry data. The different types allow for customized display in tools like dagit.

Metadata types

All metadata types inherit from MetadataValue. The following types are defined:

class dagster.DagsterAssetMetadataValue(asset_key)[source]

Representation of a dagster asset.

Parameters:

asset_key (AssetKey) – The dagster asset key

class dagster.DagsterRunMetadataValue(run_id)[source]

Representation of a dagster run.

Parameters:

run_id (str) – The run id

class dagster.FloatMetadataValue(value)[source]

Container class for float metadata entry data.

Parameters:

value (Optional[float]) – The float value.

class dagster.IntMetadataValue(value)[source]

Container class for int metadata entry data.

Parameters:

value (Optional[int]) – The int value.

class dagster.JsonMetadataValue(data)[source]

Container class for JSON metadata entry data.

Parameters:

data (Dict[str, Any]) – The JSON data.

class dagster.MarkdownMetadataValue(md_str)[source]

Container class for markdown metadata entry data.

Parameters:

md_str (Optional[str]) – The markdown as a string.

class dagster.PathMetadataValue(path)[source]

Container class for path metadata entry data.

Parameters:

path (Optional[str]) – The path as a string or conforming to os.PathLike.

class dagster.PythonArtifactMetadataValue(module, name)[source]

Container class for python artifact metadata entry data.

Parameters:

  • module (str) – The module where the python artifact can be found

  • name (str) – The name of the python artifact

class dagster.TableMetadataValue(records, schema)[source]

Container class for table metadata entry data.

Parameters:

  • records (TableRecord) – The data as a list of records (i.e. rows).

  • schema (Optional[TableSchema]) – A schema for the table.

class dagster.TableSchemaMetadataValue(schema)[source]

Representation of a schema for arbitrary tabular data.

Parameters:

schema (TableSchema) – The dictionary containing the schema representation.

class dagster.TextMetadataValue(text)[source]

Container class for text metadata entry data.

Parameters:

text (Optional[str]) – The text data.

class dagster.UrlMetadataValue(url)[source]

Container class for URL metadata entry data.

Parameters:

url (Optional[str]) – The URL as a string.

Tables

These APIs provide the ability to express table schemas (TableSchema) and table rows/records (TableRecord) in Dagster. Currently the only use case for TableSchemas and TableRecords is to wrap them in their corresponding metadata classes TableMetadataValue and TableSchemaMetadataValue for attachment to events or Dagster types.

class dagster.TableRecord(**data)[source]

Represents one record in a table. All passed keyword arguments are treated as field key/value pairs in the record. Field keys are arbitrary strings– field values must be strings, integers, floats, or bools.

class dagster.TableSchema(columns, constraints=None)[source]

Representation of a schema for tabular data. Schema is composed of two parts:

  • A required list of columns (TableColumn). Each column specifies a name, type, set of constraints, and (optional) description. type defaults to string if unspecified. Column constraints (TableColumnConstraints) consist of boolean properties unique and nullable, as well as a list of strings other containing string descriptions of all additional constraints (e.g. “<= 5”).

  • An optional list of table-level constraints (TableConstraints). A table-level constraint cannot be expressed in terms of a single column, e.g. col a > col b. Presently, all table-level constraints must be expressed as strings under the other attribute of a TableConstraints object.

# example schema
TableSchema(
    constraints = TableConstraints(
        other = [
            "foo > bar",
        ],
    ),
    columns = [
        TableColumn(
            name = "foo",
            type = "string",
            description = "Foo description",
            constraints = TableColumnConstraints(
                required = True,
                other = [
                    "starts with the letter 'a'",
                ],
            ),
        ),
        TableColumn(
            name = "bar",
            type = "string",
        ),
        TableColumn(
            name = "baz",
            type = "custom_type",
            constraints = TableColumnConstraints(
                unique = True,
            )
        ),
    ],
)
Parameters:
class dagster.TableConstraints(other)[source]

Descriptor for “table-level” constraints. Presently only one property, other is supported. This contains strings describing arbitrary table-level constraints. A table-level constraint is a constraint defined in terms of multiple columns (e.g. col_A > col_B) or in terms of rows.

Parameters:

other (List[str]) – Descriptions of arbitrary table-level constraints.

class dagster.TableColumn(name, type='string', description=None, constraints=None)[source]

Descriptor for a table column. The only property that must be specified by the user is name. If no type is specified, string is assumed. If no constraints are specified, the column is assumed to be nullable (i.e. required = False) and have no other constraints beyond the data type.

Parameters:

  • name (List[str]) – Descriptions of arbitrary table-level constraints.

  • type (Optional[str]) – The type of the column. Can be an arbitrary string. Defaults to “string”.

  • description (Optional[str]) – Description of this column. Defaults to None.

  • constraints (Optional[TableColumnConstraints]) – Column-level constraints. If unspecified, column is nullable with no constraints.

class dagster.TableColumnConstraints(nullable=True, unique=False, other=None)[source]

Descriptor for a table column’s constraints. Nullability and uniqueness are specified with boolean properties. All other constraints are described using arbitrary strings under the other property.

Parameters:

  • nullable (Optional[bool]) – If true, this column can hold null values.

  • unique (Optional[bool]) – If true, all values in this column must be unique.

  • other (List[str]) – Descriptions of arbitrary column-level constraints not expressible by the predefined properties.

Asset key

Dagster uses AssetKey to build an index on Materialization events. Assets materialized with an AssetKey are highlighted in dagit on the Assets dashboard.

class dagster.AssetKey(path)[source]

Object representing the structure of an asset key. Takes in a sanitized string, list of strings, or tuple of strings.

Example usage:

from dagster import op

@op
def emit_metadata(context, df):
    yield AssetMaterialization(
        asset_key=AssetKey('flat_asset_key'),
        metadata={"text_metadata": "Text-based metadata for this event"},
    )

@op
def structured_asset_key(context, df):
    yield AssetMaterialization(
        asset_key=AssetKey(['parent', 'child', 'grandchild']),
        metadata={"text_metadata": "Text-based metadata for this event"},
    )

@op
def structured_asset_key_2(context, df):
    yield AssetMaterialization(
        asset_key=AssetKey(('parent', 'child', 'grandchild')),
        metadata={"text_metadata": "Text-based metadata for this event"},
    )
Parameters:

path (Sequence[str]) – String, list of strings, or tuple of strings. A list of strings represent the hierarchical structure of the asset_key.