Custom Collections

For many problems, the built-in Dask collections (dask.array, dask.dataframe, dask.bag, and dask.delayed) are sufficient. For cases where they aren’t, it’s possible to create your own Dask collections. Here we describe the required methods to fulfill the Dask collection interface.


This is considered an advanced feature. For most cases the built-in collections are probably sufficient.

Before reading this you should read and understand:


The Dask Collection Interface

To create your own Dask collection, you need to fulfill the following interface. Note that there is no required base class.

It is recommended to also read Internals of the Core Dask Methods to see how this interface is used inside Dask.


The Dask graph.


dskMapping, None

The Dask graph. If None, this instance will not be interpreted as a Dask collection, and none of the remaining interface methods will be called.

If the collection also specifies __dask_layers__(), then dsk must be a HighLevelGraph or None.


The output keys for the Dask graph.

Note that there are additional constraints on keys for a Dask collection than those described in the task graph specification documentation. These additional constraints are described below.



A possibly nested list of keys that represent the outputs of the graph. After computation, the results will be returned in the same layout, with the keys replaced with their corresponding outputs.

All keys must either be non-empty strings or tuples where the first element is a non-empty string, followed by zero or more arbitrary hashables. The non-empty string is commonly known as the collection name. All collections embedded in the dask package have exactly one name, but this is not a requirement.

These are all valid outputs:

  • []

  • ["x", "y"]

  • [[("y", "a", 0), ("y", "a", 1)], [("y", "b", 0), ("y", "b", 1)]


This method only needs to be implemented if the collection uses HighLevelGraph to implement its dask graph.



Tuple of names of the HighLevelGraph layers which contain all keys returned by __dask_keys__().

static __dask_optimize__(dsk, keys, \*\*kwargs)

Given a graph and keys, return a new optimized graph.

This method can be either a staticmethod or a classmethod, but not an instancemethod.

Note that graphs and keys are merged before calling __dask_optimize__; as such, the graph and keys passed to this method may represent more than one collection sharing the same optimize method.

If not implemented, defaults to returning the graph unchanged.



The merged graphs from all collections sharing the same __dask_optimize__ method.


A list of the outputs from __dask_keys__ from all collections sharing the same __dask_optimize__ method.


Extra keyword arguments forwarded from the call to compute or persist. Can be used or ignored as needed.



The optimized Dask graph.

static __dask_scheduler__(dsk, keys, \*\*kwargs)

The default scheduler get to use for this object.

Usually attached to the class as a staticmethod, e.g.:

>>> import dask.threaded
>>> class MyCollection:
...     # Use the threaded scheduler by default
...     __dask_scheduler__ = staticmethod(dask.threaded.get)

Return the finalizer and (optional) extra arguments to convert the computed results into their in-memory representation.

Used to implement dask.compute.



A function with the signature finalize(results, *extra_args). Called with the computed results in the same structure as the corresponding keys from __dask_keys__, as well as any extra arguments as specified in extra_args. Should perform any necessary finalization before returning the corresponding in-memory collection from compute. For example, the finalize function for dask.array.Array concatenates all the individual array chunks into one large numpy array, which is then the result of compute.


Any extra arguments to pass to finalize after results. If no extra arguments should be an empty tuple.


Return the rebuilder and (optional) extra arguments to rebuild an equivalent Dask collection from a persisted or rebuilt graph.

Used to implement dask.persist().



A function with the signature rebuild(dsk, *extra_args, rename : Mapping[str, str] = None). dsk is a Mapping which contains at least the output keys returned by __dask_keys__(). The callable should return an equivalent Dask collection with the same keys as self, but with the results that are computed through a different graph. In the case of dask.persist(), the new graph will have just the output keys and the values already computed.

If the optional parameter rename is specified, it indicates that output keys may be changing too; e.g. if the previous output of __dask_keys__() was [("a", 0), ("a", 1)], after calling rebuild(dsk, *extra_args, rename={"a": "b"}) it must become [("b", 0), ("b", 1)]. The rename mapping may not contain the collection name(s); in such case the associated keys do not change. It may contain replacements for unexpected names, which must be ignored.


Any extra arguments to pass to rebuild after dsk. If no extra arguments are necessary, it must be an empty tuple.


It’s also recommended to define __dask_tokenize__, see Implementing Deterministic Hashing.

Internals of the Core Dask Methods

Dask has a few core functions (and corresponding methods) that implement common operations:

  • compute: Convert one or more Dask collections into their in-memory counterparts

  • persist: Convert one or more Dask collections into equivalent Dask collections with their results already computed and cached in memory

  • optimize: Convert one or more Dask collections into equivalent Dask collections sharing one large optimized graph

  • visualize: Given one or more Dask collections, draw out the graph that would be passed to the scheduler during a call to compute or persist

Here we briefly describe the internals of these functions to illustrate how they relate to the above interface.


The operation of compute can be broken into three stages:

  1. Graph Merging & Optimization

    First, the individual collections are converted to a single large graph and nested list of keys. How this happens depends on the value of the optimize_graph keyword, which each function takes:

    • If optimize_graph is True (default), then the collections are first grouped by their __dask_optimize__ methods. All collections with the same __dask_optimize__ method have their graphs merged and keys concatenated, and then a single call to each respective __dask_optimize__ is made with the merged graphs and keys. The resulting graphs are then merged.

    • If optimize_graph is False, then all the graphs are merged and all the keys concatenated.

    After this stage there is a single large graph and nested list of keys which represents all the collections.

  2. Computation

    After the graphs are merged and any optimizations performed, the resulting large graph and nested list of keys are passed on to the scheduler. The scheduler to use is chosen as follows:

    • If a get function is specified directly as a keyword, use that

    • Otherwise, if a global scheduler is set, use that

    • Otherwise fall back to the default scheduler for the given collections. Note that if all collections don’t share the same __dask_scheduler__ then an error will be raised.

    Once the appropriate scheduler get function is determined, it is called with the merged graph, keys, and extra keyword arguments. After this stage, results is a nested list of values. The structure of this list mirrors that of keys, with each key substituted with its corresponding result.

  3. Postcompute

    After the results are generated, the output values of compute need to be built. This is what the __dask_postcompute__ method is for. __dask_postcompute__ returns two things:

    • A finalize function, which takes in the results for the corresponding keys

    • A tuple of extra arguments to pass to finalize after the results

    To build the outputs, the list of collections and results is iterated over, and the finalizer for each collection is called on its respective results.

In pseudocode, this process looks like the following:

def compute(*collections, **kwargs):
    # 1. Graph Merging & Optimization
    # -------------------------------
    if kwargs.pop('optimize_graph', True):
        # If optimization is turned on, group the collections by
        # optimization method, and apply each method only once to the merged
        # sub-graphs.
        optimization_groups = groupby_optimization_methods(collections)
        graphs = []
        for optimize_method, cols in optimization_groups:
            # Merge the graphs and keys for the subset of collections that
            # share this optimization method
            sub_graph = merge_graphs([x.__dask_graph__() for x in cols])
            sub_keys = [x.__dask_keys__() for x in cols]
            # kwargs are forwarded to ``__dask_optimize__`` from compute
            optimized_graph = optimize_method(sub_graph, sub_keys, **kwargs)
        graph = merge_graphs(graphs)
        graph = merge_graphs([x.__dask_graph__() for x in collections])
    # Keys are always the same
    keys = [x.__dask_keys__() for x in collections]

    # 2. Computation
    # --------------
    # Determine appropriate get function based on collections, global
    # settings, and keyword arguments
    get = determine_get_function(collections, **kwargs)
    # Pass the merged graph, keys, and kwargs to ``get``
    results = get(graph, keys, **kwargs)

    # 3. Postcompute
    # --------------
    output = []
    # Iterate over the results and collections
    for res, collection in zip(results, collections):
        finalize, extra_args = collection.__dask_postcompute__()
        out = finalize(res, **extra_args)

    # `dask.compute` always returns tuples
    return tuple(output)


Persist is very similar to compute, except for how the return values are created. It too has three stages:

  1. Graph Merging & Optimization

    Same as in compute.

  2. Computation

    Same as in compute, except in the case of the distributed scheduler, where the values in results are futures instead of values.

  3. Postpersist

    Similar to __dask_postcompute__, __dask_postpersist__ is used to rebuild values in a call to persist. __dask_postpersist__ returns two things:

    • A rebuild function, which takes in a persisted graph. The keys of this graph are the same as __dask_keys__ for the corresponding collection, and the values are computed results (for the single-machine scheduler) or futures (for the distributed scheduler).

    • A tuple of extra arguments to pass to rebuild after the graph

    To build the outputs of persist, the list of collections and results is iterated over, and the rebuilder for each collection is called on the graph for its respective results.

In pseudocode, this looks like the following:

def persist(*collections, **kwargs):
    # 1. Graph Merging & Optimization
    # -------------------------------
    # **Same as in compute**
    graph = ...
    keys = ...

    # 2. Computation
    # --------------
    # **Same as in compute**
    results = ...

    # 3. Postpersist
    # --------------
    output = []
    # Iterate over the results and collections
    for res, collection in zip(results, collections):
        # res has the same structure as keys
        keys = collection.__dask_keys__()
        # Get the computed graph for this collection.
        # Here flatten converts a nested list into a single list
        subgraph = {k: r for (k, r) in zip(flatten(keys), flatten(res))}

        # Rebuild the output dask collection with the computed graph
        rebuild, extra_args = collection.__dask_postpersist__()
        out = rebuild(subgraph, *extra_args)


    # dask.persist always returns tuples
    return tuple(output)


The operation of optimize can be broken into two stages:

  1. Graph Merging & Optimization

    Same as in compute.

  2. Rebuilding

    Similar to persist, the rebuild function and arguments from __dask_postpersist__ are used to reconstruct equivalent collections from the optimized graph.

In pseudocode, this looks like the following:

def optimize(*collections, **kwargs):
    # 1. Graph Merging & Optimization
    # -------------------------------
    # **Same as in compute**
    graph = ...

    # 2. Rebuilding
    # -------------
    # Rebuild each dask collection using the same large optimized graph
    output = []
    for collection in collections:
        rebuild, extra_args = collection.__dask_postpersist__()
        out = rebuild(graph, *extra_args)

    # dask.optimize always returns tuples
    return tuple(output)


Visualize is the simplest of the 4 core functions. It only has two stages:

  1. Graph Merging & Optimization

    Same as in compute.

  2. Graph Drawing

    The resulting merged graph is drawn using graphviz and outputs to the specified file.

In pseudocode, this looks like the following:

def visualize(*collections, **kwargs):
    # 1. Graph Merging & Optimization
    # -------------------------------
    # **Same as in compute**
    graph = ...

    # 2. Graph Drawing
    # ----------------
    # Draw the graph with graphviz's `dot` tool and return the result.
    return dot_graph(graph, **kwargs)

Adding the Core Dask Methods to Your Class

Defining the above interface will allow your object to used by the core Dask functions (dask.compute, dask.persist, dask.visualize, etc.). To add corresponding method versions of these, you can subclass from dask.base.DaskMethodsMixin which adds implementations of compute, persist, and visualize based on the interface above.

Example Dask Collection

Here we create a Dask collection representing a tuple. Every element in the tuple is represented as a task in the graph. Note that this is for illustration purposes only - the same user experience could be done using normal tuples with elements of dask.delayed:

# Saved as
import dask
from dask.base import DaskMethodsMixin, replace_name_in_key
from dask.optimization import cull

# We subclass from DaskMethodsMixin to add common dask methods to our
# class. This is nice but not necessary for creating a dask collection.
class Tuple(DaskMethodsMixin):
    def __init__(self, dsk, keys):
        # The init method takes in a dask graph and a set of keys to use
        # as outputs.
        self._dsk = dsk
        self._keys = keys

    def __dask_graph__(self):
        return self._dsk

    def __dask_keys__(self):
        return self._keys

    def __dask_optimize__(dsk, keys, **kwargs):
        # We cull unnecessary tasks here. Note that this isn't necessary,
        # dask will do this automatically, this just shows one optimization
        # you could do.
        dsk2, _ = cull(dsk, keys)
        return dsk2

    # Use the threaded scheduler by default.
    __dask_scheduler__ = staticmethod(dask.threaded.get)

    def __dask_postcompute__(self):
        # We want to return the results as a tuple, so our finalize
        # function is `tuple`. There are no extra arguments, so we also
        # return an empty tuple.
        return tuple, ()

    def __dask_postpersist__(self):
        # We need to return a callable with the signature
        # rebuild(dsk, *extra_args, rename: Mapping[str, str] = None)
        return Tuple._rebuild, (self._keys,)

    def _rebuild(dsk, keys, *, rename=None):
        if rename is not None:
            keys = [replace_name_in_key(key, rename) for key in keys]
        return Tuple(dsk, keys)

    def __dask_tokenize__(self):
        # For tokenize to work we want to return a value that fully
        # represents this object. In this case it's the list of keys
        # to be computed.
        return self._keys

Demonstrating this class:

>>> from dask_tuple import Tuple
>>> from operator import add, mul

# Define a dask graph
>>> dsk = {"k0": 1,
...        ("x", "k1"): 2,
...        ("x", 1): (add, "k0", ("x", "k1")),
...        ("x", 2): (mul, ("x", "k1"), 2),
...        ("x", 3): (add, ("x", "k1"), ("x", 1))}

# The output keys for this graph.
# The first element of each tuple must be the same across the whole collection;
# the remainder are arbitrary, unique hashables
>>> keys = [("x", "k1"), ("x", 1), ("x", 2), ("x", 3)]

>>> x = Tuple(dsk, keys)

# Compute turns Tuple into a tuple
>>> x.compute()
(2, 3, 4, 5)

# Persist turns Tuple into a Tuple, with each task already computed
>>> x2 = x.persist()
>>> isinstance(x2, Tuple)
>>> x2.__dask_graph__()
{('x', 'k1'): 2, ('x', 1): 3, ('x', 2): 4, ('x', 3): 5}
>>> x2.compute()
(2, 3, 4, 5)

Checking if an object is a Dask collection

To check if an object is a Dask collection, use dask.base.is_dask_collection:

>>> from dask.base import is_dask_collection
>>> from dask import delayed

>>> x = delayed(sum)([1, 2, 3])
>>> is_dask_collection(x)
>>> is_dask_collection(1)

Implementing Deterministic Hashing

Dask implements its own deterministic hash function to generate keys based on the value of arguments. This function is available as dask.base.tokenize. Many common types already have implementations of tokenize, which can be found in dask/

When creating your own custom classes, you may need to register a tokenize implementation. There are two ways to do this:

  1. The __dask_tokenize__ method

    Where possible, it is recommended to define the __dask_tokenize__ method. This method takes no arguments and should return a value fully representative of the object.

  2. Register a function with dask.base.normalize_token

    If defining a method on the class isn’t possible or you need to customize the tokenize function for a class whose super-class is already registered (for example if you need to sub-class built-ins), you can register a tokenize function with the normalize_token dispatch. The function should have the same signature as described above.

In both cases the implementation should be the same, where only the location of the definition is different.


Both Dask collections and normal Python objects can have implementations of tokenize using either of the methods described above.


>>> from dask.base import tokenize, normalize_token

# Define a tokenize implementation using a method.
>>> class Foo:
...     def __init__(self, a, b):
...         self.a = a
...         self.b = b
...     def __dask_tokenize__(self):
...         # This tuple fully represents self
...         return (Foo, self.a, self.b)

>>> x = Foo(1, 2)
>>> tokenize(x)
>>> tokenize(x) == tokenize(x)  # token is deterministic

# Register an implementation with normalize_token
>>> class Bar:
...     def __init__(self, x, y):
...         self.x = x
...         self.y = y

>>> @normalize_token.register(Bar)
... def tokenize_bar(x):
...     return (Bar, x.x, x.x)

>>> y = Bar(1, 2)
>>> tokenize(y)
>>> tokenize(y) == tokenize(y)
>>> tokenize(y) == tokenize(x)  # tokens for different objects aren't equal

For more examples, see dask/ or any of the built-in Dask collections.