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.

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 --nprocs 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. Here are the available options:

$ dask-worker --help
Usage: dask-worker [OPTIONS] SCHEDULER

  --worker-port INTEGER  Serving worker port, defaults to randomly assigned
  --http-port INTEGER    Serving http port, defaults to randomly assigned
  --nanny-port INTEGER   Serving nanny port, defaults to randomly assigned
  --port INTEGER         Deprecated, see --nanny-port
  --host TEXT            Serving host. Defaults to an ip address that can
                         hopefully be visible from the scheduler network.
  --nthreads INTEGER     Number of threads per process. Defaults to number of
  --nprocs INTEGER       Number of worker processes to launch.  Defaults to one.
  --name TEXT            Alias
  --memory-limit TEXT    Maximum bytes of memory that this worker should use.
                         Use 0 for unlimited, or 'auto' for
                         TOTAL_MEMORY * min(1, ncores / total_cores)
  --help                 Show this message and exit.

Internal Scheduling

Internally tasks that come to the scheduler proceed through the following pipeline:

Dask worker task states

The worker also tracks data dependencies that are required to run the tasks above. These follow through a simpler pipeline:

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. 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.

Memory Management

Workers are given a target memory limit to stay under with the command line --memory-limit keyword or the memory_limit= Python keyword argument, which sets the memory limit per worker processes launched by dask-worker

$ dask-worker tcp://scheduler:port --memory-limit=auto  # TOTAL_MEMORY * min(1, ncores / total_cores)
$ dask-worker tcp://scheduler:port --memory-limit=4e9  # four gigabytes per worker process.

Workers use a few different heuristics to keep memory use beneath this limit:

  1. At 60% of memory load (as estimated by sizeof), spill least recently used data to disk
  2. At 70% of memory load, spill least recently used data to disk regardless of what is reported by sizeof
  3. At 80% of memory load, stop accepting new work on local thread pool
  4. At 95% of memory load, terminate and restart the worker

These values can be configured by modifying the ~/.config/dask/distributed.yaml file

    # Fractions of worker memory at which we take action to avoid memory blowup
    # Set any of the lower three values to False to turn off the behavior entirely
      target: 0.60  # target fraction to stay below
      spill: 0.70  # fraction at which we spill to disk
      pause: 0.80  # fraction at which we pause worker threads
      terminate: 0.95  # fraction at which we terminate the worker

Spill data to Disk

Every time the worker finishes a task it estimates the size in bytes that the result costs to keep in memory using the sizeof function. This function defaults to sys.getsizeof for arbitrary objects which uses the standard Python __sizeof__ protocol, but also has special-cased implementations for common data types like NumPy arrays and Pandas dataframes.

When the sum of the number of bytes of the data in memory exceeds 60% of the available threshold the worker will begin to dump the least recently used data to disk. You can control this location with the --local-directory keyword.:

$ dask-worker tcp://scheduler:port --memory-limit 4e9 --local-directory /scratch

That data is still available and will be read back from disk when necessary. On the diagnostic dashboard status page disk I/O will show up in the task stream plot as orange blocks. Additionally the memory plot in the upper left will become orange and then red.

Monitor process memory load

The approach above can fail for a few reasons

  1. Custom objects may not report their memory size accurately
  2. User functions may take up more RAM than expected
  3. Significant amounts of data may accumulate in network I/O buffers

To address this we periodically monitor the memory of the worker process every 200 ms. If the system reported memory use is above 70% of the target memory usage then the worker will start dumping unused data to disk, even if internal sizeof recording hasn’t yet reached the normal 60% limit.

Halt worker threads

At 80% load the worker’s thread pool will stop accepting new tasks. This gives time for the write-to-disk functionality to take effect even in the face of rapidly accumulating data.

Kill Worker

At 95% memory load a worker’s nanny process will terminate it. This is to avoid having our worker job being terminated by an external job scheduler (like YARN, Mesos, SGE, etc..). After termination the nanny will restart the worker in a fresh state.

API Documentation

class distributed.worker.Worker(scheduler_ip=None, scheduler_port=None, scheduler_file=None, ncores=None, loop=None, local_dir='dask-worker-space', services=None, service_ports=None, name=None, reconnect=True, memory_limit='auto', executor=None, resources=None, silence_logs=None, death_timeout=None, preload=None, preload_argv=None, security=None, contact_address=None, memory_monitor_interval='200ms', extensions=None, metrics=None, data=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 the the worker uses to manage and track internal computations.


Informational State

These attributes don’t change significantly during execution.

  • ncores: int:
    Number of cores used by this worker process
  • executor: concurrent.futures.ThreadPoolExecutor:
    Executor used to perform computation
  • local_dir: 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_out_connections: int
    The maximum number of concurrent outgoing requests for data
  • total_in_connections: int
    The maximum number of concurrent incoming requests for data
  • total_comm_nbytes: int
  • 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

This 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.

  • data: {key: object}:
    Dictionary mapping keys to actual values
  • task_state: {key: string}:
    The state of all tasks that the scheduler has asked us to compute. Valid states include waiting, constrained, executing, memory, erred
  • tasks: {key: dict}
    The function, args, kwargs of a task. We run this when appropriate
  • dependencies: {key: {deps}}
    The data needed by this key to run
  • dependents: {dep: {keys}}
    The keys that use this dependency
  • data_needed: deque(keys)
    The keys whose data we still lack, arranged in a deque
  • waiting_for_data: {kep: {deps}}
    A dynamic verion of dependencies. All dependencies that we still don’t have for a particular key.
  • ready: [keys]
    Keys that are ready to run. Stored in a LIFO stack
  • constrained: [keys]
    Keys for which we have the data to run, but are waiting on abstract resources like GPUs. Stored in a FIFO deque
  • executing: {keys}
    Keys that are currently executing
  • executed_count: int
    A number of tasks that this worker has run in its lifetime
  • long_running: {keys}
    A set of keys of tasks that are running and have started their own long-running clients.
  • dep_state: {dep: string}:
    The state of all dependencies required by our tasks Valid states include waiting, flight, and memory
  • who_has: {dep: {worker}}
    Workers that we believe have this data
  • has_what: {worker: {deps}}
    The data that we care about that we think a worker has
  • pending_data_per_worker: {worker: [dep]}
    The data on each worker that we still want, prioritized as a deque
  • in_flight_tasks: {task: worker}
    All dependencies that are coming to us in current peer-to-peer connections and the workers from which they are coming.
  • in_flight_workers: {worker: {task}}
    The workers from which we are currently gathering data and the dependencies we expect from those connections
  • comm_bytes: int
    The total number of bytes in flight
  • suspicious_deps: {dep: int}
    The number of times a dependency has not been where we expected it
  • nbytes: {key: int}
    The size of a particular piece of data
  • types: {key: type}
    The type of a particular piece of data
  • threads: {key: int}
    The ID of the thread on which the task ran
  • active_threads: {int: key}
    The keys currently running on active threads
  • exceptions: {key: exception}
    The exception caused by running a task if it erred
  • tracebacks: {key: traceback}
    The exception caused by running a task if it erred
  • startstops: {key: [(str, float, float)]}
    Log of transfer, load, and compute times for a task
  • priorities: {key: tuple}
    The priority of a key given by the scheduler. Determines run order.
  • durations: {key: float}
    Expected duration of a task
  • resource_restrictions: {key: {str: number}}
    Abstract resources required to run a task
scheduler_ip: str
scheduler_port: int
ip: str, optional
data: MutableMapping, type, None

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

ncores: int, optional
loop: tornado.ioloop.IOLoop
local_dir: 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 TOTAL_MEMORY * min(1, ncores / total_cores) Use strings or numbers like 5GB or 5e9

memory_target_fraction: float

Fraction of memory to try to stay beneath

memory_spill_fraction: float

Fraction of memory at which we start spilling to disk

memory_pause_fraction: float

Fraction of memory at which we stop running new tasks

executor: concurrent.futures.Executor
resources: dict

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

See also

distributed.scheduler.Scheduler, distributed.nanny.Nanny


Use the command line to start a worker:

$ dask-scheduler
Start scheduler at

$ dask-worker
Start worker at:     
Registered with scheduler at: