Workers provide two functions:

  1. Compute tasks as directed by the scheduler

  2. Store and serve computed results to other workers or clients

Each worker contains a ThreadPool that it uses to evaluate tasks as requested by the scheduler. It stores the results of these tasks locally and serves them to other workers or clients on demand. If the worker is asked to evaluate a task for which it does not have all of the necessary data then it will reach out to its peer workers to gather the necessary dependencies.

A typical conversation between a scheduler and two workers Alice and Bob may look like the following:

Scheduler -> Alice:  Compute ``x <- add(1, 2)``!
Alice -> Scheduler:  I've computed x and am holding on to it!

Scheduler -> Bob:    Compute ``y <- add(x, 10)``!
                     You will need x.  Alice has x.
Bob -> Alice:        Please send me x.
Alice -> Bob:        Sure.  x is 3!
Bob -> Scheduler:    I've computed y and am holding on to it!

Storing Data

Data is stored locally in a dictionary in the .data attribute that maps keys to the results of function calls.

{'x': 3,
 'y': 13,
 '(df, 0)': pd.DataFrame(...),

This .data attribute is a MutableMapping that is typically a combination of in-memory and on-disk storage with an LRU policy to move data between them.

Read more: Worker Memory Management

Thread Pool

Each worker sends computations to a thread in a concurrent.futures.ThreadPoolExecutor for computation. These computations occur in the same process as the Worker communication server so that they can access and share data efficiently between each other. For the purposes of data locality all threads within a worker are considered the same worker.

If your computations are mostly numeric in nature (for example NumPy and Pandas computations) and release the GIL entirely then it is advisable to run dask-worker processes with many threads and one process. This reduces communication costs and generally simplifies deployment.

If your computations are mostly Python code and don’t release the GIL then it is advisable to run dask-worker processes with many processes and one thread per process:

$ dask-worker scheduler:8786 --nworkers 8 --nthreads 1

This will launch 8 worker processes each of which has its own ThreadPoolExecutor of size 1.

If your computations are external to Python and long-running and don’t release the GIL then beware that while the computation is running the worker process will not be able to communicate to other workers or to the scheduler. This situation should be avoided. If you don’t link in your own custom C/Fortran code then this topic probably doesn’t apply.

Command Line tool

Use the dask-worker command line tool to start an individual worker. For more details on the command line options, please have a look at the command line tools documentation.

Internal Scheduling

Internally tasks that come to the scheduler proceed through the following pipeline as distributed.worker_state_machine.TaskState objects. Tasks which follow this path have a runspec defined which instructs the worker how to execute them.

Dask worker task states

Data dependencies are also represented as TaskState objects and follow a simpler path through the execution pipeline. These tasks do not have a runspec defined and instead contain a listing of workers to collect their result from.

Dask worker dependency states

As tasks arrive they are prioritized and put into a heap. They are then taken from this heap in turn to have any remote dependencies collected. For each dependency we select a worker at random that has that data and collect the dependency from that worker. To improve bandwidth we opportunistically gather other dependencies of other tasks that are known to be on that worker, up to a maximum of 200MB of data (too little data and bandwidth suffers, too much data and responsiveness suffers). We use a fixed number of connections (around 10-50) so as to avoid overly-fragmenting our network bandwidth. In the event that the network comms between two workers are saturated, a dependency task may cycle between fetch and flight until it is successfully collected.

After all dependencies for a task are in memory we transition the task to the ready state and put the task again into a heap of tasks that are ready to run.

We collect from this heap and put the task into a thread from a local thread pool to execute.

Optionally, this task may identify itself as a long-running task (see Tasks launching tasks), at which point it secedes from the thread pool.

A task either errs or its result is put into memory. In either case a response is sent back to the scheduler.

Tasks slated for execution and tasks marked for collection from other workers must follow their respective transition paths as defined above. The only exceptions to this are when:

  • A task is stolen, in which case a task which might have been collected will instead be executed on the thieving worker

  • Scheduler intercession, in which the scheduler reassigns a task that was previously assigned to a separate worker to a new worker. This most commonly occurs when a worker dies during computation.


Dask workers are by default launched, monitored, and managed by a small Nanny process.

class distributed.nanny.Nanny(scheduler_ip=None, scheduler_port=None, scheduler_file=None, worker_port: int | str | Collection[int] | None = 0, nthreads=None, loop=None, local_directory=None, services=None, name=None, memory_limit='auto', reconnect=True, validate=False, quiet=False, resources=None, silence_logs=None, death_timeout=None, preload=None, preload_argv=None, preload_nanny=None, preload_nanny_argv=None, security=None, contact_address=None, listen_address=None, worker_class=None, env=None, interface=None, host=None, port: int | str | Collection[int] | None = None, protocol=None, config=None, **worker_kwargs)[source]

A process to manage worker processes

The nanny spins up Worker processes, watches then, and kills or restarts them as necessary. It is necessary if you want to use the Client.restart method, or to restart the worker automatically if it gets to the terminate fraction of its memory limit.

The parameters for the Nanny are mostly the same as those for the Worker with exceptions listed below.

env: dict, optional

Environment variables set at time of Nanny initialization will be ensured to be set in the Worker process as well. This argument allows to overwrite or otherwise set environment variables for the Worker. It is also possible to set environment variables using the option distributed.nanny.environ. Precedence as follows

  1. Nanny arguments

  2. Existing environment variables

  3. Dask configuration

See also

async close(timeout=5)[source]

Close the worker process, stop all comms.


A signal that we shouldn’t try to restart workers if they go away

This is used as part of the cluster shutdown process.

async instantiate() distributed.core.Status[source]

Start a local worker process

Blocks until the process is up and the scheduler is properly informed

async kill(timeout=2)[source]

Kill the local worker process

Blocks until both the process is down and the scheduler is properly informed

async start_unsafe()[source]

Start nanny, start local process, start watching

API Documentation

class distributed.worker_state_machine.TaskState(key: str, run_spec: SerializedTask | None = None, dependencies: set[TaskState] = <factory>, dependents: set[TaskState] = <factory>, waiting_for_data: set[TaskState] = <factory>, waiters: set[TaskState] = <factory>, state: TaskStateState = 'released', previous: TaskStateState | None = None, next: TaskStateState | None = None, duration: float | None = None, priority: tuple[int, ...] | None = None, who_has: set[str] = <factory>, coming_from: str | None = None, resource_restrictions: dict[str, float] = <factory>, exception: Serialize | None = None, traceback: Serialize | None = None, exception_text: str = '', traceback_text: str = '', type: type | None = None, suspicious_count: int = 0, startstops: list[StartStop] = <factory>, start_time: float | None = None, stop_time: float | None = None, metadata: dict = <factory>, nbytes: int | None = None, annotations: dict | None = None, done: bool = False)[source]

Holds volatile state relating to an individual Dask task.

Not to be confused with distributed.scheduler.TaskState, which holds similar information on the scheduler side.

annotations: dict | None = None

Arbitrary task annotations

coming_from: str | None = None

The worker that current task data is coming from if task is in flight

dependencies: set[TaskState]

The data needed by this key to run

dependents: set[TaskState]

The keys that use this dependency

done: bool = False

True if the task is in memory or erred; False otherwise

duration: float | None = None

Expected duration of the task

exception: Serialize | None = None

The exception caused by running a task if it erred (serialized)

exception_text: str = ''

string representation of exception

key: str

Task key. Mandatory.

metadata: dict

Metadata related to the task. Stored metadata should be msgpack serializable (e.g. int, string, list, dict).

nbytes: int | None = None

The size of the value of the task, if in memory

next: TaskStateState | None = None

The next state of the task. It is not None iff state == resumed.

previous: TaskStateState | None = None

The previous state of the task. It is not None iff state in (cancelled, resumed).

priority: tuple[int, ...] | None = None

The priority this task given by the scheduler. Determines run order.

resource_restrictions: dict[str, float]

Abstract resources required to run a task

run_spec: SerializedTask | None = None

A named tuple containing the function, args, kwargs and task associated with this TaskState instance. This defaults to None and can remain empty if it is a dependency that this worker will receive from another worker.

start_time: float | None = None

Time at which task begins running

startstops: list[StartStop]

Log of transfer, load, and compute times for a task

state: TaskStateState = 'released'

The current state of the task

stop_time: float | None = None

Time at which task finishes running

suspicious_count: int = 0

The number of times a dependency has not been where we expected it

traceback: Serialize | None = None

The traceback caused by running a task if it erred (serialized)

traceback_text: str = ''

string representation of traceback

type: type | None = None

The type of a particular piece of data

waiters: set[TaskState]

Subset of dependents that are not in memory

waiting_for_data: set[TaskState]

Subset of dependencies that are not in memory

who_has: set[str]

Addresses of workers that we believe have this data

class distributed.worker_state_machine.WorkerState(*, nthreads: int = 1, address: str | None = None, data: MutableMapping[str, object] | None = None, threads: dict[str, int] | None = None, plugins: dict[str, WorkerPlugin] | None = None, resources: Mapping[str, float] | None = None, total_out_connections: int = 9999, validate: bool = True, transition_counter_max: int | Literal[False] = False)[source]

State machine encapsulating the lifetime of all tasks on a worker.

Not to be confused with distributed.scheduler.WorkerState.


The data attributes of this class are implementation details and may be changed without a deprecation cycle.


The attributes of this class are all heavily correlated with each other. Do not modify them directly, ever, as it is extremely easy to obtain a broken state this way, which in turn will likely result in cluster-wide deadlocks.

The state should be exclusively mutated through handle_stimulus().

actors: dict[str, object]

Actor tasks. See Actors.

address: str

Worker <IP address>:<port>. This is used in decision-making by the state machine, e.g. to determine if a peer worker is running on the same host or not. This attribute may not be known when the WorkerState is initialised. It must be set before the first call to handle_stimulus().

property all_running_tasks: set[distributed.worker_state_machine.TaskState]

All tasks that are currently occupying a thread. They may or may not count towards the maximum number of threads.

These are:

  • ts.status in (executing, long-running)

  • ts.status in (cancelled, resumed) and ts.previous in (executing, long-running)

available_resources: dict[str, float]

{resource name: amount}. Current resources that aren’t being currently consumed by task execution. Always less or equal to total_resources. See Worker Resources.

busy_workers: set[str]

Peer workers that recently returned a busy status. Workers in this set won’t be asked for additional dependencies for some time.

comm_nbytes: int

The total number of bytes in flight

comm_threshold_bytes: int

Ignore total_out_connections as long as comm_nbytes is less than this value.

constrained: HeapSet[TaskState]

Priority heap of tasks that are ready to run, but are waiting on abstract resources like GPUs. Mutually exclusive with ready. See available_resources and Worker Resources.

data: MutableMapping[str, object]

In-memory tasks data. This collection is shared by reference between Worker, WorkerMemoryManager, and this class.

data_needed: defaultdict[str, HeapSet[TaskState]]

The tasks which still require data in order to execute and are in memory on at least another worker, prioritized as per-worker heaps. All and only tasks with TaskState.state == 'fetch' are in this collection. A TaskState with multiple entries in who_has will appear multiple times here.

executed_count: int

A number of tasks that this worker has run in its lifetime; this includes failed and cancelled tasks. See also executing_count().

executing: set[TaskState]

Set of tasks that are currently running. See also executing_count() and long_running.

property executing_count: int

Count of tasks currently executing on this worker and counting towards the maximum number of threads.

It includes cancelled tasks, but does not include long running (a.k.a. seceded) tasks.

generation: int

Counter that decreases every time the compute-task handler is invoked by the Scheduler. It is appended to TaskState.priority and acts as a tie-breaker between tasks that have the same priority on the Scheduler, determining a last-in-first-out order between them.

handle_stimulus(*stims: distributed.worker_state_machine.StateMachineEvent) list[source]

Process one or more external events, transition relevant tasks to new states, and return a list of instructions to be executed as a consequence.

has_what: defaultdict[str, set[str]]

{worker address: {ts.key, ...}. The data that we care about that we think a worker has

in_flight_tasks: set[TaskState]

Which tasks that are coming to us in current peer-to-peer connections. All and only tasks with TaskState.state == ‘flight’. See also in_flight_tasks_count().

property in_flight_tasks_count: int

Count of tasks currently being replicated from other workers to this one.

in_flight_workers: dict[str, set[str]]

{worker address: {ts.key, ...}} The workers from which we are currently gathering data and the dependencies we expect from those connections. Workers in this dict won’t be asked for additional dependencies until the current query returns.

log: deque[tuple]

Transition log: [(..., stimulus_id: str | None, timestamp: float), ...] The number of stimuli logged is capped. See also story() and stimulus_log.

long_running: set[TaskState]

Set of tasks that are currently running and have called secede(), so they no longer count towards the maximum number of concurrent tasks (nthreads). These tasks do not appear in the executing set.

missing_dep_flight: set[TaskState]

All and only tasks with TaskState.state == 'missing'.

nthreads: int

Number of tasks that can be executing in parallel. At any given time, executing_count() <= nthreads.

plugins: dict[str, WorkerPlugin]

{name: worker plugin}. This collection is shared by reference between Worker and this class. The Worker managed adding and removing plugins, while the WorkerState invokes the WorkerPlugin.transition method, is available.

ready: HeapSet[TaskState]

Priority heap of tasks that are ready to run and have no resource constrains. Mutually exclusive with constrained.

rng: random.Random

Statically-seeded random state, used to guarantee determinism whenever a pseudo-random choice is required

running: bool

True if the state machine should start executing more tasks and fetch dependencies whenever a slot is available. This property must be kept aligned with the Worker: WorkerState.running == (Worker.status is Status.running).

stimulus_log: deque[StateMachineEvent]

Log of all stimuli received by handle_stimulus(). The number of events logged is capped. See also log and stimulus_story().

stimulus_story(*keys_or_tasks: str | TaskState) list[StateMachineEvent][source]

Return all state machine events involving one or more tasks

story(*keys_or_tasks_or_stimuli: str | TaskState) list[tuple][source]

Return all records from the transitions log involving one or more tasks or stimulus_id’s

target_message_size: int

Number of bytes to fetch from the same worker in a single call to BaseWorker.gather_dep(). Multiple small tasks that can be fetched from the same worker will be clustered in a single instruction as long as their combined size doesn’t exceed this value.

tasks: dict[str, TaskState]

{key: TaskState}. The tasks currently executing on this worker (and any dependencies of those tasks)

threads: dict[str, int]

{ts.key: thread ID}. This collection is shared by reference between Worker and this class. While the WorkerState is thread-agnostic, it still needs access to this information in some cases. This collection is populated by distributed.worker.Worker.execute(). It does not need to be populated for the WorkerState to work.

total_out_connections: int

The maximum number of concurrent incoming requests for data. See also distributed.worker.Worker.total_in_connections.

total_resources: dict[str, float]

{resource name: amount}. Total resources available for task execution. See :doc: resources.

transition_counter: int

Total number of state transitions so far. See also log and transition_counter_max.

transition_counter_max: int | Literal[False]

Raise an error if the transition_counter ever reaches this value. This is meant for debugging only, to catch infinite recursion loops. In production, it should always be set to False.

validate: bool

If True, enable expensive internal consistency check. Typically disabled in production.

waiting_for_data_count: int

A count of how many tasks are currently waiting for data

class distributed.worker_state_machine.BaseWorker(state: distributed.worker_state_machine.WorkerState)[source]

Wrapper around the WorkerState that implements instructions handling. This is an abstract class with several @abc.abstractmethod methods, to be subclassed by Worker and by unit test mock-ups.

abstract batched_send(msg: dict[str, Any]) None[source]

Send a fire-and-forget message to the scheduler through bulk comms.

msg: dict

msgpack-serializable message to send to the scheduler. Must have a ‘op’ key which is registered in Scheduler.stream_handlers.

async close(timeout: float = 30) None[source]

Cancel all asynchronous instructions

abstract async execute(key: str, *, stimulus_id: str) StateMachineEvent | None[source]

Execute a task

abstract async gather_dep(worker: str, to_gather: Collection[str], total_nbytes: int, *, stimulus_id: str) StateMachineEvent | None[source]

Gather dependencies for a task from a worker who has them


Address of worker to gather dependencies from


Keys of dependencies to gather from worker – this is not necessarily equivalent to the full list of dependencies of dep as some dependencies may already be present on this worker.


Total number of bytes for all the dependencies in to_gather combined

handle_stimulus(*stims: distributed.worker_state_machine.StateMachineEvent) None[source]

Forward one or more external stimuli to WorkerState.handle_stimulus() and process the returned instructions, invoking the relevant Worker callbacks (@abc.abstractmethod methods below).

Spawn asyncio tasks for all asynchronous instructions and start tracking them.

abstract async retry_busy_worker_later(worker: str) StateMachineEvent | None[source]

Wait some time, then take a peer worker out of busy state

class distributed.worker.Worker(scheduler_ip: str | None = None, scheduler_port: int | None = None, *, scheduler_file: str | None = None, nthreads: int | None = None, loop: IOLoop | None = None, local_directory: str | None = None, services: dict | None = None, name: Any | None = None, reconnect: bool | None = None, executor: Executor | dict[str, Executor] | Literal['offload'] | None = None, resources: dict[str, float] | None = None, silence_logs: int | None = None, death_timeout: Any | None = None, preload: list[str] | None = None, preload_argv: list[str] | list[list[str]] | None = None, security: Security | dict[str, Any] | None = None, contact_address: str | None = None, heartbeat_interval: Any = '1s', extensions: dict[str, type] | None = None, metrics: Mapping[str, Callable[[Worker], Any]] = {}, startup_information: Mapping[str, Callable[[Worker], Any]] = {}, interface: str | None = None, host: str | None = None, port: int | str | Collection[int] | None = None, protocol: str | None = None, dashboard_address: str | None = None, dashboard: bool = False, http_prefix: str = '/', nanny: Nanny | None = None, plugins: tuple[WorkerPlugin, ...] = (), low_level_profiler: bool | None = None, validate: bool | None = None, profile_cycle_interval=None, lifetime: Any | None = None, lifetime_stagger: Any | None = None, lifetime_restart: bool | None = None, transition_counter_max: int | Literal[False] = False, memory_limit: str | float = 'auto', data: MutableMapping[str, Any] | Callable[[], MutableMapping[str, Any]] | tuple[Callable[..., MutableMapping[str, Any]], dict[str, Any]] | None = None, memory_target_fraction: float | Literal[False] | None = None, memory_spill_fraction: float | Literal[False] | None = None, memory_pause_fraction: float | Literal[False] | None = None, **kwargs)[source]

Worker node in a Dask distributed cluster

Workers perform two functions:

  1. Serve data from a local dictionary

  2. Perform computation on that data and on data from peers

Workers keep the scheduler informed of their data and use that scheduler to gather data from other workers when necessary to perform a computation.

You can start a worker with the dask-worker command line application:

$ dask-worker scheduler-ip:port

Use the --help flag to see more options:

$ dask-worker --help

The rest of this docstring is about the internal state that the worker uses to manage and track internal computations.


Informational State

These attributes don’t change significantly during execution.

  • nthreads: int:

    Number of nthreads used by this worker process

  • executors: dict[str, concurrent.futures.Executor]:

    Executors used to perform computation. Always contains the default executor.

  • local_directory: path:

    Path on local machine to store temporary files

  • scheduler: rpc:

    Location of scheduler. See .ip/.port attributes.

  • name: string:


  • services: {str: Server}:

    Auxiliary web servers running on this worker

  • service_ports: {str: port}:

  • total_in_connections: int

    The maximum number of concurrent incoming requests for data. See also distributed.worker_state_machine.WorkerState.total_out_connections.

  • batched_stream: BatchedSend

    A batched stream along which we communicate to the scheduler

  • log: [(message)]

    A structured and queryable log. See Worker.story

Volatile State

These attributes track the progress of tasks that this worker is trying to complete. In the descriptions below a key is the name of a task that we want to compute and dep is the name of a piece of dependent data that we want to collect from others.

  • threads: {key: int}

    The ID of the thread on which the task ran

  • active_threads: {int: key}

    The keys currently running on active threads

  • state: WorkerState

    Encapsulated state machine. See BaseWorker and WorkerState

scheduler_ip: str, optional
scheduler_port: int, optional
scheduler_file: str, optional
host: str, optional
data: MutableMapping, type, None

The object to use for storage, builds a disk-backed LRU dict by default

nthreads: int, optional
local_directory: str, optional

Directory where we place local resources

name: str, optional
memory_limit: int, float, string

Number of bytes of memory that this worker should use. Set to zero for no limit. Set to ‘auto’ to calculate as system.MEMORY_LIMIT * min(1, nthreads / total_cores) Use strings or numbers like 5GB or 5e9

memory_target_fraction: float or False

Fraction of memory to try to stay beneath (default: read from config key

memory_spill_fraction: float or False

Fraction of memory at which we start spilling to disk (default: read from config key distributed.worker.memory.spill)

memory_pause_fraction: float or False

Fraction of memory at which we stop running new tasks (default: read from config key distributed.worker.memory.pause)

max_spill: int, string or False

Limit of number of bytes to be spilled on disk. (default: read from config key distributed.worker.memory.max-spill)

executor: concurrent.futures.Executor, dict[str, concurrent.futures.Executor], “offload”
The executor(s) to use. Depending on the type, it has the following meanings:
  • Executor instance: The default executor.

  • Dict[str, Executor]: mapping names to Executor instances. If the “default” key isn’t in the dict, a “default” executor will be created using ThreadPoolExecutor(nthreads).

  • Str: The string “offload”, which refer to the same thread pool used for offloading communications. This results in the same thread being used for deserialization and computation.

resources: dict

Resources that this worker has like {'GPU': 2}

nanny: str

Address on which to contact nanny, if it exists

lifetime: str

Amount of time like “1 hour” after which we gracefully shut down the worker. This defaults to None, meaning no explicit shutdown time.

lifetime_stagger: str

Amount of time like “5 minutes” to stagger the lifetime value The actual lifetime will be selected uniformly at random between lifetime +/- lifetime_stagger

lifetime_restart: bool

Whether or not to restart a worker after it has reached its lifetime Default False

kwargs: optional

Additional parameters to ServerNode constructor


Use the command line to start a worker:

$ dask-scheduler
Start scheduler at

$ dask-worker
Start worker at:     
Registered with scheduler at:
batched_send(msg: dict[str, Any]) None[source]

Implements BaseWorker abstract method.

Send a fire-and-forget message to the scheduler through bulk comms.

If we’re not currently connected to the scheduler, the message will be silently dropped!

async close(timeout: float = 30, executor_wait: bool = True, nanny: bool = True) str | None[source]

Close the worker

Close asynchronous operations running on the worker, stop all executors and comms. If requested, this also closes the nanny.

timeoutfloat, default 30

Timeout in seconds for shutting down individual instructions

executor_waitbool, default True

If True, shut down executors synchronously, otherwise asynchronously

nannybool, default True

If True, close the nanny

str | None

None if worker already in closing state or failed, “OK” otherwise

async close_gracefully(restart=None)[source]

Gracefully shut down a worker

This first informs the scheduler that we’re shutting down, and asks it to move our data elsewhere. Afterwards, we close as normal

property data:[str, Any]

{task key: task payload} of all completed tasks, whether they were computed on this Worker or computed somewhere else and then transferred here over the network.

When using the default configuration, this is a zict buffer that automatically spills to disk whenever the target threshold is exceeded. If spilling is disabled, it is a plain dict instead. It could also be a user-defined arbitrary dict-like passed when initialising the Worker or the Nanny. Worker logic should treat this opaquely and stick to the MutableMapping API.


This same collection is also available at and

async execute(key: str, *, stimulus_id: str) StateMachineEvent | None[source]

Execute a task. Implements BaseWorker abstract method.

async gather_dep(worker: str, to_gather: Collection[str], total_nbytes: int, *, stimulus_id: str) StateMachineEvent | None[source]

Implements BaseWorker abstract method

get_current_task() str[source]

Get the key of the task we are currently running

This only makes sense to run within a task

See also



>>> from dask.distributed import get_worker
>>> def f():
...     return get_worker().get_current_task()
>>> future = client.submit(f)  
>>> future.result()  
handle_stimulus(*stims: distributed.worker_state_machine.StateMachineEvent) None[source]

Override BaseWorker method for added validation

async retry_busy_worker_later(worker: str) StateMachineEvent | None[source]

Wait some time, then take a peer worker out of busy state. Implements BaseWorker abstract method.

async start_unsafe()[source]

Attempt to start the server. This is not idempotent and not protected against concurrent startup attempts.

This is intended to be overwritten or called by subclasses. For a safe startup, please use Server.start instead.

If death_timeout is configured, we will require this coroutine to finish before this timeout is reached. If the timeout is reached we will close the instance and raise an asyncio.TimeoutError

trigger_profile() None[source]

Get a frame from all actively computing threads

Merge these frames into existing profile counts

property worker_address

For API compatibility with Nanny