Dynamic Graphs#

The ability for portions of a graph to be duplicated at runtime.

Relevant APIs#

DynamicOutDeclare that an op will return dynamic outputs
DynamicOutputThe object that an op will yield repeatedly, each containing a value and a unique mapping_key


The basic unit of computation in Dagster is the op. In certain cases it is desirable to run the same op multiple times on different pieces of similar data.

Dynamic outputs are the tool Dagster provides to allow resolving the pieces of data at runtime and having downstream copies of the ops created for each piece.

Using Dynamic Outputs#

A Static Job#

Here we start with a contrived example of a job containing a single expensive op:

def data_processing():
    large_data = load_big_data()
    interesting_result = expensive_processing(large_data)
    return analyze(interesting_result)

def naive():

While, the implementation of expensive_computation can internally do something to parallelize the compute, if anything goes wrong with any part we have to restart the whole computation.

A Dynamic Job#

With this motivation we will break up the computation using Dynamic Outputs. First we will define our new op that will use dynamic outputs. First we use DynamicOut to declare that this op will return dynamic outputs. Then in the function we yield a number of DynamicOutput objects that each contain a value and a unique mapping_key.

def load_pieces():
    large_data = load_big_data()
    for idx, piece in large_data.chunk():
        yield DynamicOutput(piece, mapping_key=idx)

Then after creating ops for our downstream operations, we can put them all together in a job.

def dynamic_graph():
    pieces = load_pieces()
    results = pieces.map(compute_piece)

Within our @job decorated composition function, the object representing the dynamic output can not be passed directly to another op. Either map or collect must be invoked on it.

map takes a Callable which receives a single argument. This callable is evaluated once, and any invoked op that is passed the input argument will establish dependencies. The ops downstream of a dynamic output will be cloned for each dynamic output, and identified using the associated mapping_key. The return value from the callable is captured and wrapped in an object that allows for subsequent map or collect calls.

collect creates a fan-in dependency over all the dynamic copies. The dependent op will receive a list containing all the values.

Advanced Mapping Examples#

Returning Dynamic Outputs#

In addition to yielding, DynamicOutput objects can also be returned as part of a list.

from dagster import DynamicOut, DynamicOutput, op
from typing import List

def return_dynamic() -> List[DynamicOutput[str]]:
    outputs = []
    for idx, page_key in get_pages():
        outputs.append(DynamicOutput(page_key, mapping_key=idx))
    return outputs
DynamicOutput can be used as a generic type annotation describing the expected type of the output.


The following two examples are equivalent ways to establish a sequence of ops that occur for each dynamic output.

def chained():
    results = dynamic_values().map(echo).map(echo).map(echo)

def chained_alt():
    def _for_each(val):
        a = echo(val)
        b = echo(a)
        return echo(b)

    results = dynamic_values().map(_for_each)

Additional Arguments#

A lambda or scoped function can be used to pass non-dynamic outputs along side dynamic ones in map downstream.

def other_arg():
    non_dynamic = one()
    dynamic_values().map(lambda val: add(val, non_dynamic))

Multiple Outputs#

Multiple outputs are returned via a namedtuple, where each entry can be used via map or collect.

        "values": DynamicOut(),
        "negatives": DynamicOut(),
def multiple_dynamic_values():
    for i in range(2):
        yield DynamicOutput(i, output_name="values", mapping_key=f"num_{i}")
        yield DynamicOutput(-i, output_name="negatives", mapping_key=f"neg_{i}")

def multiple():
    # can unpack on assignment (order based)
    values, negatives = multiple_dynamic_values()
    process(negatives.map(echo).collect())  # can use map or collect as usual

    # or access by name
    outs = multiple_dynamic_values()