Scheduling State

Overview

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 ready to run, 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, queue, stacks, 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.

  • ready: deque(key)

    A deque of keys that are ready to run now but haven’t yet been sent to a worker to run. These keys show no affinity to any particular worker and so can be run equally well by anyone. This is a common pool of tasks.

  • stacks: {worker: [keys]}:

    Keys that are ready to run and show some affinity to a particular worker. These keys typically have dependencies that are known to be on that worker or have user-defined restrictions that require them to run on certain nodes. A worker pulls from this stack first before taking tasks from the common pool of ready tasks.

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

    Keys that are currently running on 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 a set of all workers that are currently running them. This is redundant with processing and just here for faster indexed querying. It is rare for these sets to be larger than size one but does happen in rare cases when a task might be submitted redundantly to multiple workers.

  • 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.

  • 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.

  • 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.

  • 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.

  • stealable: [[key]]

    A list of stacks of stealable keys, ordered by stealability. For more information see Work Stealing

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'

who_has[key].add(worker)
has_what[worker].add(key)

nbytes[key] = nbytes

processing[worker].remove(key)
del rprocessing[key]

if key in who_wants:
    send_done_message_to_clients(who_wants[key])

for dep in dependencies[key]:
   waiting_data[dep].remove(key)

for dep in dependents[key]:
   waiting[dep].remove(key)

if stacks[worker]:
    next_task = stacks[worker].pop()
elif ready:
    next_task = ready.pop()
else:
    idle.add(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', 'queue', 'stacks', '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
  • Queue (ready): Ready to be computed by any worker
  • Stacks (ready): Ready to be computed by a particular preferred worker
  • 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)
        recommendations.update(new)

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

Stimuli

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.

API

class distributed.scheduler.Scheduler(center=None, loop=None, max_buffer_size=8353576960.0, delete_interval=500, synchronize_worker_interval=60000, ip=None, services=None, allowed_failures=3, validate=False, steal=True, **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 Tornado IOStreams 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 127.0.0.1:8786

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

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

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

State

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, stacks, 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
  • stacks: {worker: [keys]}:
    List of keys waiting to be sent to each worker
  • processing: {worker: {key: cost}}:
    Set of keys currently in execution on each worker and their expected duration
  • stack_durations: {worker: [ints]}:
    Expected durations of stacked tasks
  • stacks_duration: {worker: int}:
    Total duration of all tasks in each workers stack
  • rprocessing: {key: {worker}}:
    Set of workers 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
  • 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.
  • loose_restrictions: {key}:
    Set of keys for which we are allow to violate restrictions (see above) if not valid workers are present.
  • 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.
  • stealable: [[key]]
    A list of stacks of stealable keys, ordered by stealability
  • 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
  • streams: [IOStreams]:
    A list of Tornado IOStreams 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
  • scheduler_queues: [Queues]:
    A list of Tornado Queues from which we accept stimuli
  • report_queues: [Queues]:
    A list of Tornado Queues on which we report results
add_client(stream, client=None)[source]

Add client to network

We listen to all future messages from this IOStream.

add_keys(stream=None, address=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_plugin(plugin)[source]

Add external plugin to scheduler

See https://distributed.readthedocs.io/en/latest/plugins.html

add_worker(stream=None, address=None, keys=(), ncores=None, name=None, coerce_address=True, nbytes=None, now=None, host_info=None, **info)[source]

Add a new worker to the cluster

broadcast(stream=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

change_worker_cores(stream=None, worker=None, diff=0)[source]

Add or remove cores from a worker

This is used when a worker wants to spin off a long-running task

cleanup()[source]

Clean up queues and coroutines, prepare to stop

clear_data_from_workers()[source]

Send delete signals to clear unused data from workers

This watches the .deleted_keys attribute, which stores a set of keys to be deleted from each worker. This function is run periodically by the ._delete_periodic_callback to actually remove the data.

This runs every self.delete_interval milliseconds.

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

Remove keys from client desired list

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

Send cleanup signal to all coroutines then wait until finished

close_streams()[source]

Close all active IOStreams

coerce_address(addr)[source]

Coerce possible input addresses to canonical form

Handles lists, strings, bytes, tuples, or aliases

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

ensure_occupied()[source]

Run ready tasks on idle workers

Work stealing policy

If some workers are idle but not others, if there are no globally ready tasks, and if there are tasks in worker stacks, then we start to pull preferred tasks from overburdened workers and deploy them back into the global pool in the following manner.

We determine the number of tasks to reclaim as the number of all tasks in all stacks times the fraction of idle workers to all workers. We sort the stacks by size and walk through them, reclaiming half of each stack until we have enough task to fill the global pool. We are careful not to reclaim tasks that are restricted to run on certain workers.

ensure_occupied_queue(worker, count)[source]

Send at most count tasks from the ready queue to the specified worker

ensure_occupied_stacks(worker)[source]

Send tasks to worker while it has tasks and free cores

These tasks may come from the worker’s own stacks or from the global ready deque.

We update the idle workers set appropriately.

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

Provides a data stream to external requester

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

finished()[source]

Wait until all coroutines have ceased

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

Collect data in from workers

get_versions(stream)[source]

Basic information about ourselves and our cluster

handle_messages(in_queue, report, client=None)[source]

The master client coroutine. Handles all inbound messages from clients.

This runs once per Client IOStream or Queue.

See also

Scheduler.worker_stream
The equivalent function for workers
handle_queues(scheduler_queue, report_queue)[source]

Register new control and report queues to the Scheduler

Queues are not in common use. This may be deprecated in the future.

identity(stream)[source]

Basic information about ourselves and our cluster

issaturated(worker, latency=0.005)[source]

Determine if a worker has enough work to avoid being idle

A worker is saturated if the following criteria are met

  1. It is working on at least as many tasks as it has cores
  2. The expected time it will take to complete all of its currently assigned tasks is at least a full round-trip time. This is relevant when it has many small tasks
rebalance(stream=None, keys=None, workers=None)[source]

Rebalance keys so that each worker stores roughly equal bytes

Policy

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(client=None)[source]

Remove client from network

remove_plugin(plugin)[source]

Remove external plugin from scheduler

remove_worker(stream=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(stream=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.

Parameters:

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)[source]

Publish updates to all listening Queues and Streams

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

restart(environment=None)[source]

Restart all workers. Reset local state.

scatter(stream=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(port=8786, start_queues=True)[source]

Clear out old state and restart all running coroutines

start_ipython(stream=None)[source]

Start an IPython kernel

Returns Jupyter connection info dictionary.

steal_time_ratio(key, bandwidth=100000000.0)[source]

The compute to communication time ratio of a key

Returns:

ratio: The compute/communication time ratio of the task

loc: The self.stealable bin into which this key should go

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

Stop execution on a list of keys

stimulus_missing_data(keys=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

Scheduler.transitions
transitive version of this function

Examples

>>> self.transition('x', 'waiting')
{'x': 'ready'}
transition_story(*keys)[source]

Get all transitions that touch one of the input keys

transitions(recommendations)[source]

Process transitions until none are left

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

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

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

See also

Scheduler.mark_key_in_memory

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

Add new computations to the internal dask graph

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

work_steal()[source]

Steal tasks from saturated workers to idle workers

This moves tasks from the bottom of the stacks of over-occupied workers to the stacks of idling workers.

worker_stream(worker)[source]

Listen to responses from a single worker

This is the main loop for scheduler-worker interaction

See also

Scheduler.handle_messages
Equivalent coroutine for clients
workers_list(workers)[source]

List of qualifying workers

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

distributed.scheduler.decide_worker(dependencies, stacks, stack_duration, processing, who_has, has_what, restrictions, loose_restrictions, nbytes, ncores, key)[source]

Decide which worker should take task

>>> dependencies = {'c': {'b'}, 'b': {'a'}}
>>> stacks = {'alice:8000': ['z'], 'bob:8000': []}
>>> processing = {'alice:8000': set(), 'bob:8000': set()}
>>> who_has = {'a': {'alice:8000'}}
>>> has_what = {'alice:8000': {'a'}}
>>> nbytes = {'a': 100}
>>> ncores = {'alice:8000': 1, 'bob:8000': 1}
>>> restrictions = {}
>>> loose_restrictions = set()

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

>>> decide_worker(dependencies, stacks, processing, who_has, has_what,
...               restrictions, loose_restrictions, nbytes, ncores, 'b')
'alice:8000'

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, stacks, processing, who_has, has_what,
...               restrictions, loose_restrictions, nbytes, ncores, 'b')
'bob:8000'

Optionally provide restrictions of where jobs are allowed to occur

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

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}
>>> stacks = {'alice:8000': [], 'bob:8000': []}
>>> decide_worker(dependencies, stacks, processing, who_has, has_what,
...               {}, set(), nbytes, ncores, 'c')
'bob:8000'