Scheduling State


The life of a computation with Dask can be described in the following stages:

  1. The user authors a graph using some library, perhaps Dask.delayed or dask.dataframe or the submit/map functions on the client. They submit these tasks to the scheduler.
  2. The schedulers assimilates these tasks into its graph of all tasks to track and as their dependencies become available it asks workers to run each of these tasks.
  3. The worker receives information about how to run the task, communicates with its peer workers to collect dependencies, and then runs the relevant function on the appropriate data. It reports back to the scheduler that it has finished.
  4. The scheduler reports back to the user that the task has completed. If the user desires, it then fetches the data from the worker through the scheduler.

Most relevant logic is in tracking tasks as they evolve from newly submitted, to waiting for dependencies, to actively running on some worker, to finished in memory, to garbage collected. Tracking this process, and tracking all effects that this task has on other tasks that might depend on it, is the majority of the complexity of the dynamic task scheduler. This section describes the system used to perform this tracking.

For more abstract information about the policies used by the scheduler, see Scheduling Policies.

State Variables

We start with a description of the state that the scheduler keeps on each task. Each of the following is a dictionary keyed by task name (described below):

  • tasks: {key: task}:

    Dictionary mapping key to a serialized task.

    A key is the name of a task, generally formed from the name of the function, followed by a hash of the function and arguments, like 'inc-ab31c010444977004d656610d2d421ec'.

    The value of this dictionary is the task, which is an unevaluated function and arguments. This is stored in one of two forms:

    • {'function': inc, 'args': (1,), 'kwargs': {}}; a dictionary with the function, arguments, and keyword arguments (kwargs). However in the scheduler these are stored serialized, as they were sent from the client, so it looks more like {'function': b'\x80\x04\x95\xcb\...', 'args': b'...', }
    • {'task': (inc, 1)}: a tuple satisfying the dask graph protocol. This again is stored serialized.

    These are the values that will eventually be sent to a worker when the task is ready to run.

  • dependencies and dependents: {key: {keys}}:

    These are dictionaries which show which tasks depend on which others. They contain redundant information. If dependencies[a] == {b, c} then the task with the name of a depends on the results of the two tasks with the names of b and c. There will be complimentary entries in dependents such that a in dependents[b] and a in dependents[c] such as dependents[b] == {a, d}. Keeping the information around twice allows for constant-time access for either direction of query, so we can both look up a task’s out-edges or in-edges efficiently.

  • waiting and waiting_data: {key: {keys}}:

    These are dictionaries very similar to dependencies and dependents, but they only track keys that are still in play. For example waiting looks like dependencies, tracking all of the tasks that a certain task requires before it can run. However as tasks are completed and arrive in memory they are removed from their dependents sets in waiting, so that when a set becomes empty we know that a key is ready to run and ready to be allocated to a worker.

    The waiting_data dictionary on the other hand holds all of the dependents of a key that have yet to run and still require that this task stay in memory in services of tasks that may depend on it (its dependents). When a value set in this dictionary becomes empty its task may be garbage collected (unless some client actively desires that this task stay in memory).

  • task_state: {key: string}:

    The task_state dictionary holds the current state of every key. Current valid states include released, waiting, no-worker, processing, memory, and erred. These states are explained further below.

  • priority: {key: tuple}:

    The priority dictionary provides each key with a relative ranking. This ranking is generally a tuple of two parts. The first (and dominant) part corresponds to when it was submitted. Generally earlier tasks take precedence. The second part is determined by the client, and is a way to prioritize tasks within a large graph that may be important, such as if they are on the critical path, or good to run in order to release many dependencies. This is explained further in Scheduling Policy

    A key’s priority is only used to break ties, when many keys are being considered for execution. The priority does not determine running order, but does exert some subtle influence that does significantly shape the long term performance of the cluster.

  • processing: {worker: {key: cost}}:

    Keys that are currently allocated to a worker. This is keyed by worker address and contains the expected cost in seconds of running that task.

  • rprocessing: {key: worker}:

    The reverse of the processing dictionary. This is all keys that are currently running with the workers that is currently running them. This is redundant with processing and just here for faster indexed querying.

  • who_has: {key: {worker}}:

    For keys that are in memory this shows on which workers they currently reside.

  • has_what: {worker: {key}}:

    This is the transpose of who_has, showing all keys that currently reside on each worker.

  • released: {keys}

    The set of keys that are known, but released from memory. These have typically run to completion and are no longer necessary.

  • unrunnable: {key}

    The set unrunnable contains keys that are not currently able to run, probably because they have a user defined restriction (described below) that is not met by any available worker. These keys are waiting for an appropriate worker to join the network before computing.

  • host_restrictions: {key: {hostnames}}:

    A set of hostnames per key of where that key can be run. Usually this is empty unless a key has been specifically restricted to only run on certain hosts. These restrictions don’t include a worker port. Any worker on that hostname is deemed valid.

  • worker_restrictions: {key: {worker addresses}}:

    A set of complete host:port worker addresses per key of where that key can be run. Usually this is empty unless a key has been specifically restricted to only run on certain workers.

  • loose_restrictions: {key}:

    Set of keys for which we are allowed to violate restrictions (see above) if not valid workers are present and the task would otherwise go into the unrunnable set.

  • resource_restrictions: {key: {resource: quantity}}:

    Resources required by a task, such as {'GPU': 1} or {'memory': 1e9}. These names must match resources specified when creating workers.

  • worker_resources: {worker: {str: Number}}:

    The available resources on each worker like {'gpu': 2, 'mem': 1e9}. These are abstract quantities that constrain certain tasks from running at the same time.

  • used_resources: {worker: {str: Number}}:

    The sum of each resource used by all tasks allocated to a particular worker.

  • exceptions and tracebacks: {key: Exception/Traceback}:

    Dictionaries mapping keys to remote exceptions and tracebacks. When tasks fail we store their exceptions and tracebacks (serialized from the worker) here so that users may gather the exceptions to see the error.

  • exceptions_blame: {key: key}:

    If a task fails then we mark all of its dependent tasks as failed as well. This dictionary lets any failed task see which task was the origin of its failure.

  • suspicious_tasks: {key: int}

    Number of times a task has been involved in a worker failure. Some tasks may cause workers to fail (such as sys.exit(0)). When a worker fails all of the tasks on that worker are reassigned to others. This combination of behaviors can cause a bad task to catastrophically destroy all workers on the cluster, one after another. Whenever a worker fails we mark each task currently running on that worker as suspicious. If a task is involved in three failures (or some other fixed constant) then we mark the task as failed.

  • who_wants: {key: {client}}:

    When a client submits a graph to the scheduler it also specifies which output keys it desires. Those keys are tracked here where each desired key knows which clients want it. These keys will not be released from memory and, when they complete, messages will be sent to all of these clients that the task is ready.

  • wants_what: {client: {key}}:

    The transpose of who_wants.

  • nbytes: {key: int}:

    The number of bytes, as determined by sizeof, of the result of each finished task. This number is used for diagnostics and to help prioritize work.

Example Event and Response

Whenever an event happens, like when a client sends up more tasks, or when a worker finishes a task, the scheduler changes the state above. For example when a worker reports that a task has finished we perform actions like the following:

Task `key` finished by `worker`:

task_state[key] = 'memory'


nbytes[key] = nbytes

del rprocessing[key]

if key in who_wants:

for dep in dependencies[key]:

for dep in dependents[key]:

for task in ready_tasks():
    worker = best_wrker(task):
    send_task_to_worker(task, worker)

State Transitions

The code presented in the section above is just for demonstration. In practice writing this code for every possible event is highly error prone, resulting in hard-to-track-down bugs. Instead the scheduler moves tasks between a fixed set of states, notably 'released', 'waiting', 'no-worker', 'processing', 'memory', 'error'. Transitions between common pairs of states are well defined and, if no path exists between a pair, the graph of transitions can be traversed to find a valid sequence of transitions. Along with these transitions come consistent logging and optional runtime checks that are useful in testing.

Tasks fall into the following states with the following allowed transitions

Dask scheduler task states
  • Released: known but not actively computing or in memory
  • Waiting: On track to be computed, waiting on dependencies to arrive in memory
  • No-worker (ready, rare): Ready to be computed, but no appropriate worker exists
  • Processing: Actively being computed by one or more workers
  • Memory: In memory on one or more workers
  • Erred: Task has computed and erred
  • Forgotten (not actually a state): Task is no longer needed by any client and so it removed from state

Every transition between states is a separate method in the scheduler. These task transition functions are prefixed with transition and then have the name of the start and finish task state like the following.

def transition_released_waiting(self, key):

def transition_processing_memory(self, key):

def transition_processing_erred(self, key):

These functions each have three effects.

  1. They perform the necessary transformations on the scheduler state (the 20 dicts/lists/sets) to move one key between states.
  2. They return a dictionary of recommended {key: state} transitions to enact directly afterwards on other keys. For example after we transition a key into memory we may find that many waiting keys are now ready to transition from waiting to a ready state.
  3. Optionally they include a set of validation checks that can be turned on for testing.

Rather than call these functions directly we call the central function transition:

def transition(self, key, final_state):
    """ Transition key to the suggested state """

This transition function finds the appropriate path from the current to the final state. It also serves as a central point for logging and diagnostics.

Often we want to enact several transitions at once or want to continually respond to new transitions recommended by initial transitions until we reach a steady state. For that we use the transitions function (note the plural s).

def transitions(self, recommendations):
    recommendations = recommendations.copy()
    while recommendations:
        key, finish = recommendations.popitem()
        new = self.transition(key, finish)

This function runs transition, takes the recommendations and runs them as well, repeating until no further task-transitions are recommended.


Transitions occur from stimuli, which are state-changing messages to the scheduler from workers or clients. The scheduler responds to the following stimuli:

  • Workers
    • Task finished: A task has completed on a worker and is now in memory
    • Task erred: A task ran and erred on a worker
    • Task missing data: A task tried to run but was unable to find necessary data on other workers
    • Worker added: A new worker was added to the network
    • Worker removed: An existing worker left the network
  • Clients
    • Update graph: The client sends more tasks to the scheduler
    • Release keys: The client no longer desires the result of certain keys

Stimuli functions are prepended with the text stimulus, and take a variety of keyword arguments from the message as in the following examples:

def stimulus_task_finished(self, key=None, worker=None, nbytes=None,
                           type=None, compute_start=None, compute_stop=None,
                           transfer_start=None, transfer_stop=None):

def stimulus_task_erred(self, key=None, worker=None,
                        exception=None, traceback=None)

These functions change some non-essential administrative state and then call transition functions.

Note that there are several other non-state-changing messages that we receive from the workers and clients, such as messages requesting information about the current state of the scheduler. These are not considered stimuli.


class distributed.scheduler.Scheduler(center=None, loop=None, delete_interval=500, synchronize_worker_interval=60000, services=None, allowed_failures=3, extensions=[<class 'distributed.channels.ChannelScheduler'>, <class 'distributed.publish.PublishExtension'>, <class 'distributed.stealing.WorkStealing'>], validate=False, **kwargs)[source]

Dynamic distributed task scheduler

The scheduler tracks the current state of workers, data, and computations. The scheduler listens for events and responds by controlling workers appropriately. It continuously tries to use the workers to execute an ever growing dask graph.

All events are handled quickly, in linear time with respect to their input (which is often of constant size) and generally within a millisecond. To accomplish this the scheduler tracks a lot of state. Every operation maintains the consistency of this state.

The scheduler communicates with the outside world through Comm objects. It maintains a consistent and valid view of the world even when listening to several clients at once.

A Scheduler is typically started either with the dask-scheduler executable:

$ dask-scheduler
Scheduler started at

Or within a LocalCluster a Client starts up without connection information:

>>> c = Client()  
>>> c.cluster.scheduler  

Users typically do not interact with the scheduler directly but rather with the client object Client.


The scheduler contains the following state variables. Each variable is listed along with what it stores and a brief description.

  • tasks: {key: task}:

    Dictionary mapping key to a serialized task like the following: {'function': b'...', 'args': b'...'} or {'task': b'...'}

  • dependencies: {key: {keys}}:

    Dictionary showing which keys depend on which others

  • dependents: {key: {keys}}:

    Dictionary showing which keys are dependent on which others

  • task_state: {key: string}:

    Dictionary listing the current state of every task among the following: released, waiting, queue, no-worker, processing, memory, erred

  • priority: {key: tuple}:

    A score per key that determines its priority

  • waiting: {key: {key}}:

    Dictionary like dependencies but excludes keys already computed

  • waiting_data: {key: {key}}:

    Dictionary like dependents but excludes keys already computed

  • ready: deque(key)

    Keys that are ready to run, but not yet assigned to a worker

  • processing: {worker: {key: cost}}:

    Set of keys currently in execution on each worker and their expected duration

  • rprocessing: {key: worker}:

    The worker currently executing a particular task

  • who_has: {key: {worker}}:

    Where each key lives. The current state of distributed memory.

  • has_what: {worker: {key}}:

    What worker has what keys. The transpose of who_has.

  • released: {keys}

    Set of keys that are known, but released from memory

  • unrunnable: {key}

    Keys that we are unable to run

  • host_restrictions: {key: {hostnames}}:

    A set of hostnames per key of where that key can be run. Usually this is empty unless a key has been specifically restricted to only run on certain hosts.

  • worker_restrictions: {key: {workers}}:

    Like host_restrictions except that these include specific host:port worker names

  • loose_restrictions: {key}:

    Set of keys for which we are allow to violate restrictions (see above) if not valid workers are present.

  • resource_restrictions: {key: {str: Number}}:

    Resources required by a task, such as {'GPU': 1} or {'memory': 1e9}. These names must match resources specified when creating workers.

  • worker_resources: {worker: {str: Number}}:

    The available resources on each worker like {'gpu': 2, 'mem': 1e9}. These are abstract quantities that constrain certain tasks from running at the same time.

  • used_resources: {worker: {str: Number}}:

    The sum of each resource used by all tasks allocated to a particular worker.

  • exceptions: {key: Exception}:

    A dict mapping keys to remote exceptions

  • tracebacks: {key: list}:

    A dict mapping keys to remote tracebacks stored as a list of strings

  • exceptions_blame: {key: key}:

    A dict mapping a key to another key on which it depends that has failed

  • suspicious_tasks: {key: int}

    Number of times a task has been involved in a worker failure

  • deleted_keys: {key: {workers}}

    Locations of workers that have keys that should be deleted

  • wants_what: {client: {key}}:

    What keys are wanted by each client.. The transpose of who_wants.

  • who_wants: {key: {client}}:

    Which clients want each key. The active targets of computation.

  • nbytes: {key: int}:

    Number of bytes for a key as reported by workers holding that key.

  • ncores: {worker: int}:

    Number of cores owned by each worker

  • idle: {worker}:

    Set of workers that are not fully utilized

  • worker_info: {worker: {str: data}}:

    Information about each worker

  • host_info: {hostname: dict}:

    Information about each worker host

  • worker_bytes: {worker: int}:

    Number of bytes in memory on each worker

  • occupancy: {worker: time}

    Expected runtime for all tasks currently processing on a worker

  • services: {str: port}:

    Other services running on this scheduler, like HTTP

  • loop: IOLoop:

    The running Tornado IOLoop

  • comms: [Comm]:

    A list of Comms from which we both accept stimuli and report results

  • task_duration: {key-prefix: time}

    Time we expect certain functions to take, e.g. {'sum': 0.25}

  • coroutines: [Futures]:

    A list of active futures that control operation

add_client(comm, client=None)[source]

Add client to network

We listen to all future messages from this Comm.

add_keys(comm=None, worker=None, keys=())[source]

Learn that a worker has certain keys

This should not be used in practice and is mostly here for legacy reasons.


Add external plugin to scheduler


add_worker(comm=None, address=None, keys=(), ncores=None, name=None, resolve_address=True, nbytes=None, now=None, resources=None, host_info=None, **info)[source]

Add a new worker to the cluster

broadcast(comm=None, msg=None, workers=None, hosts=None, nanny=False)[source]

Broadcast message to workers, return all results

cancel_key(key, client, retries=5)[source]

Cancel a particular key and all dependents


Clean up queues and coroutines, prepare to stop

client_releases_keys(keys=None, client=None)[source]

Remove keys from client desired list

close(comm=None, fast=False)[source]

Send cleanup signal to all coroutines then wait until finished


Close all active Comms.

close_worker(stream=None, worker=None)[source]

Remove a worker from the cluster

This both removes the worker from our local state and also sends a signal to the worker to shut down. This works regardless of whether or not the worker has a nanny process restarting it

coerce_address(addr, resolve=True)[source]

Coerce possible input addresses to canonical form. resolve can be disabled for testing with fake hostnames.

Handles strings, tuples, or aliases.


Coerce the hostname of a worker.

correct_time_delay(worker, msg)[source]

Apply offset time delay in message times.

Clocks on different workers differ. We keep track of a relative “now” through periodic heartbeats. We use this known delay to align message times to Scheduler local time. In particular this helps with diagnostics.

Operates in place

feed(comm, function=None, setup=None, teardown=None, interval=1, **kwargs)[source]

Provides a data Comm to external requester

Caution: this runs arbitrary Python code on the scheduler. This should eventually be phased out. It is mostly used by diagnostics.


Wait until all coroutines have ceased

gather(comm=None, keys=None)[source]

Collect data in from workers


Basic information about ourselves and our cluster

get_worker_service_addr(worker, service_name)[source]

Get the (host, port) address of the named service on the worker. Returns None if the service doesn’t exist.

handle_client(comm, client=None)[source]

Listen and respond to messages from clients

This runs once per Client Comm or Queue.

See also

The equivalent function for workers

Listen to responses from a single worker

This is the main loop for scheduler-worker interaction

See also

Equivalent coroutine for clients

Basic information about ourselves and our cluster

rebalance(comm=None, keys=None, workers=None)[source]

Rebalance keys so that each worker stores roughly equal bytes


This orders the workers by what fraction of bytes of the existing keys they have. It walks down this list from most-to-least. At each worker it sends the largest results it can find and sends them to the least occupied worker until either the sender or the recipient are at the average expected load.


Remove client from network


Remove external plugin from scheduler

remove_worker(comm=None, address=None, safe=False)[source]

Remove worker from cluster

We do this when a worker reports that it plans to leave or when it appears to be unresponsive. This may send its tasks back to a released state.

replicate(comm=None, keys=None, n=None, workers=None, branching_factor=2, delete=True)[source]

Replicate data throughout cluster

This performs a tree copy of the data throughout the network individually on each piece of data.


keys: Iterable

list of keys to replicate

n: int

Number of replications we expect to see within the cluster

branching_factor: int, optional

The number of workers that can copy data in each generation

report(msg, client=None)[source]

Publish updates to all listening Queues and Comms

If the message contains a key then we only send the message to those comms that care about the key.


Restart all workers. Reset local state.

run_function(stream, function, args=(), kwargs={})[source]

Run a function within this process

See also


scatter(comm=None, data=None, workers=None, client=None, broadcast=False, timeout=2)[source]

Send data out to workers

send_task_to_worker(worker, key)[source]

Send a single computational task to a worker

start(addr_or_port=8786, start_queues=True)[source]

Clear out old state and restart all running coroutines


Start an IPython kernel

Returns Jupyter connection info dictionary.

stimulus_cancel(comm, keys=None, client=None)[source]

Stop execution on a list of keys

stimulus_missing_data(cause=None, key=None, worker=None, ensure=True, **kwargs)[source]

Mark that certain keys have gone missing. Recover.

stimulus_task_erred(key=None, worker=None, exception=None, traceback=None, **kwargs)[source]

Mark that a task has erred on a particular worker

stimulus_task_finished(key=None, worker=None, **kwargs)[source]

Mark that a task has finished execution on a particular worker

transition(key, finish, *args, **kwargs)[source]

Transition a key from its current state to the finish state

Returns:Dictionary of recommendations for future transitions

See also

transitive version of this function


>>> self.transition('x', 'waiting')
{'x': 'processing'}

Get all transitions that touch one of the input keys


Process transitions until none are left

This includes feedback from previous transitions and continues until we reach a steady state

update_data(comm=None, who_has=None, nbytes=None, client=None)[source]

Learn that new data has entered the network from an external source

See also


update_graph(client=None, tasks=None, keys=None, dependencies=None, restrictions=None, priority=None, loose_restrictions=None, resources=None)[source]

Add new computations to the internal dask graph

This happens whenever the Client calls submit, map, get, or compute.


Return set of currently valid worker addresses for key

If all workers are valid then this returns True. This checks tracks the following state:

  • worker_restrictions
  • host_restrictions
  • resource_restrictions
worker_objective(key, worker)[source]

Objective function to determine which worker should get the key

Minimize expected start time. If a tie then break with data storate.


List of qualifying workers

Takes a list of worker addresses or hostnames. Returns a list of all worker addresses that match


Find workers that we can close with low cost

This returns a list of workers that are good candidates to retire. These workers are idle (not running anything) and are storing relatively little data relative to their peers. If all workers are idle then we still maintain enough workers to have enough RAM to store our data, with a comfortable buffer.

This is for use with systems like distributed.deploy.adaptive.


memory_factor: Number

Amount of extra space we want to have for our stored data. Defaults two 2, or that we want to have twice as much memory as we currently have data.


to_close: list of workers that are OK to close

distributed.scheduler.decide_worker(dependencies, occupancy, who_has, valid_workers, loose_restrictions, objective, key)[source]

Decide which worker should take task

>>> dependencies = {'c': {'b'}, 'b': {'a'}}
>>> occupancy = {'alice:8000': 0, 'bob:8000': 0}
>>> who_has = {'a': {'alice:8000'}}
>>> nbytes = {'a': 100}
>>> ncores = {'alice:8000': 1, 'bob:8000': 1}
>>> valid_workers = True
>>> loose_restrictions = set()

We choose the worker that has the data on which ‘b’ depends (alice has ‘a’)

>>> decide_worker(dependencies, occupancy, who_has, has_what,
...               valid_workers, loose_restrictions, nbytes, ncores, 'b')

If both Alice and Bob have dependencies then we choose the less-busy worker

>>> who_has = {'a': {'alice:8000', 'bob:8000'}}
>>> has_what = {'alice:8000': {'a'}, 'bob:8000': {'a'}}
>>> decide_worker(dependencies, who_has, has_what,
...               valid_workers, loose_restrictions, nbytes, ncores, 'b')

Optionally provide valid workers of where jobs are allowed to occur

>>> valid_workers = {'alice:8000', 'charlie:8000'}
>>> decide_worker(dependencies, who_has, has_what,
...               valid_workers, loose_restrictions, nbytes, ncores, 'b')

If the task requires data communication, then we choose to minimize the number of bytes sent between workers. This takes precedence over worker occupancy.

>>> dependencies = {'c': {'a', 'b'}}
>>> who_has = {'a': {'alice:8000'}, 'b': {'bob:8000'}}
>>> has_what = {'alice:8000': {'a'}, 'bob:8000': {'b'}}
>>> nbytes = {'a': 1, 'b': 1000}
>>> decide_worker(dependencies, who_has, has_what,
...               {}, set(), nbytes, ncores, 'c')