Futures
Dask supports a real-time task framework that extends Python’s concurrent.futures interface. This interface is good for arbitrary task scheduling like dask.delayed, but is immediate rather than lazy, which provides some more flexibility in situations where the computations may evolve over time.
These features depend on the second generation task scheduler found in dask.distributed (which, despite its name, runs very well on a single machine).
Examples
Visit https://examples.dask.org/futures.html to see and run examples using futures with Dask.
Start Dask Client
You must start a Client
to use the futures interface. This tracks state
among the various worker processes or threads:
from dask.distributed import Client
client = Client() # start local workers as processes
# or
client = Client(processes=False) # start local workers as threads
If you have Bokeh installed, then this starts up a
diagnostic dashboard at http://localhost:8787
.
Submit Tasks
You can submit individual tasks using the submit
method:
def inc(x):
return x + 1
def add(x, y):
return x + y
a = client.submit(inc, 10) # calls inc(10) in background thread or process
b = client.submit(inc, 20) # calls inc(20) in background thread or process
The submit
function returns a Future
, which refers to a remote result. This result may
not yet be completed:
>>> a
<Future: status: pending, key: inc-b8aaf26b99466a7a1980efa1ade6701d>
Eventually it will complete. The result stays in the remote thread/process/worker until you ask for it back explicitly:
>>> a
<Future: status: finished, type: int, key: inc-b8aaf26b99466a7a1980efa1ade6701d>
>>> a.result() # blocks until task completes and data arrives
11
You can pass futures as inputs to submit. Dask automatically handles dependency tracking; once all input futures have completed, they will be moved onto a single worker (if necessary), and then the computation that depends on them will be started. You do not need to wait for inputs to finish before submitting a new task; Dask will handle this automatically:
c = client.submit(add, a, b) # calls add on the results of a and b
Similar to Python’s map
, you can use Client.map
to call the same
function and many inputs:
futures = client.map(inc, range(1000))
However, note that each task comes with about 1ms of overhead. If you want to map a function over a large number of inputs, then you might consider dask.bag or dask.dataframe instead.
Move Data
Given any future, you can call the .result
method to gather the result.
This will block until the future is done computing and then transfer the result
back to your local process if necessary:
>>> c.result()
32
You can gather many results concurrently using the Client.gather
method.
This can be more efficient than calling .result()
on each future
sequentially:
>>> # results = [future.result() for future in futures]
>>> results = client.gather(futures) # this can be faster
If you have important local data that you want to include in your computation, you can either include it as a normal input to a submit or map call:
>>> df = pd.read_csv('training-data.csv')
>>> future = client.submit(my_function, df)
Or you can scatter
it explicitly. Scattering moves your data to a worker
and returns a future pointing to that data:
>>> remote_df = client.scatter(df)
>>> remote_df
<Future: status: finished, type: DataFrame, key: bbd0ca93589c56ea14af49cba470006e>
>>> future = client.submit(my_function, remote_df)
Both of these accomplish the same result, but using scatter can sometimes be
faster. This is especially true if you use processes or distributed workers
(where data transfer is necessary) and you want to use df
in many
computations. Scattering the data beforehand avoids excessive data movement.
Calling scatter on a list scatters all elements individually. Dask will spread these elements evenly throughout workers in a round-robin fashion:
>>> client.scatter([1, 2, 3])
[<Future: status: finished, type: int, key: c0a8a20f903a4915b94db8de3ea63195>,
<Future: status: finished, type: int, key: 58e78e1b34eb49a68c65b54815d1b158>,
<Future: status: finished, type: int, key: d3395e15f605bc35ab1bac6341a285e2>]
References, Cancellation, and Exceptions
Dask will only compute and hold onto results for which there are active futures. In this way, your local variables define what is active in Dask. When a future is garbage collected by your local Python session, Dask will feel free to delete that data or stop ongoing computations that were trying to produce it:
>>> del future # deletes remote data once future is garbage collected
You can also explicitly cancel a task using the Future.cancel
or
Client.cancel
methods:
>>> future.cancel() # deletes data even if other futures point to it
If a future fails, then Dask will raise the remote exceptions and tracebacks if you try to get the result:
def div(x, y):
return x / y
>>> a = client.submit(div, 1, 0) # 1 / 0 raises a ZeroDivisionError
>>> a
<Future: status: error, key: div-3601743182196fb56339e584a2bf1039>
>>> a.result()
1 def div(x, y):
----> 2 return x / y
ZeroDivisionError: division by zero
All futures that depend on an erred future also err with the same exception:
>>> b = client.submit(inc, a)
>>> b
<Future: status: error, key: inc-15e2e4450a0227fa38ede4d6b1a952db>
You can collect the exception or traceback explicitly with the
Future.exception
or Future.traceback
methods.
Waiting on Futures
You can wait on a future or collection of futures using the wait
function:
from dask.distributed import wait
>>> wait(futures)
This blocks until all futures are finished or have erred.
You can also iterate over the futures as they complete using the
as_completed
function:
from dask.distributed import as_completed
futures = client.map(score, x_values)
best = -1
for future in as_completed(futures):
y = future.result()
if y > best:
best = y
For greater efficiency, you can also ask as_completed
to gather the results
in the background:
for future, result in as_completed(futures, with_results=True):
# y = future.result() # don't need this
...
Or collect all futures in batches that had arrived since the last iteration:
for batch in as_completed(futures, with_results=True).batches():
for future, result in batch:
...
Additionally, for iterative algorithms, you can add more futures into the
as_completed
iterator during iteration:
seq = as_completed(futures)
for future in seq:
y = future.result()
if condition(y):
new_future = client.submit(...)
seq.add(new_future) # add back into the loop
or use seq.update(futures)
to add multiple futures at once.
Fire and Forget
Sometimes we don’t care about gathering the result of a task, and only care about side effects that it might have like writing a result to a file:
>>> a = client.submit(load, filename)
>>> b = client.submit(process, a)
>>> c = client.submit(write, b, out_filename)
As noted above, Dask will stop work that doesn’t have any active futures. It
thinks that because no one has a pointer to this data that no one cares. You
can tell Dask to compute a task anyway, even if there are no active futures,
using the fire_and_forget
function:
from dask.distributed import fire_and_forget
>>> fire_and_forget(c)
This is particularly useful when a future may go out of scope, for example, as part of a function:
def process(filename):
out_filename = 'out-' + filename
a = client.submit(load, filename)
b = client.submit(process, a)
c = client.submit(write, b, out_filename)
fire_and_forget(c)
return # here we lose the reference to c, but that's now ok
for filename in filenames:
process(filename)
Submit Tasks from Tasks
This is an advanced feature and is rarely necessary in the common case.
Tasks can launch other tasks by getting their own client. This enables complex and highly dynamic workloads:
from dask.distributed import get_client
def my_function(x):
...
# Get locally created client
client = get_client()
# Do normal client operations, asking cluster for computation
a = client.submit(...)
b = client.submit(...)
a, b = client.gather([a, b])
return a + b
It also allows you to set up long running tasks that watch other resources like sockets or physical sensors:
def monitor(device):
client = get_client()
while True:
data = device.read_data()
future = client.submit(process, data)
fire_and_forget(future)
for device in devices:
fire_and_forget(client.submit(monitor))
However, each running task takes up a single thread, and so if you launch many
tasks that launch other tasks, then it is possible to deadlock the system if you
are not careful. You can call the secede
function from within a task to
have it remove itself from the dedicated thread pool into an administrative
thread that does not take up a slot within the Dask worker:
from dask.distributed import get_client, secede
def monitor(device):
client = get_client()
secede() # remove this task from the thread pool
while True:
data = device.read_data()
future = client.submit(process, data)
fire_and_forget(future)
If you intend to do more work in the same thread after waiting on client work, you may want to explicitly block until the thread is able to rejoin the thread pool. This allows some control over the number of threads that are created and stops too many threads from being active at once, over-saturating your hardware:
def f(n): # assume that this runs as a task
client = get_client()
secede() # secede while we wait for results to come back
futures = client.map(func, range(n))
results = client.gather(futures)
rejoin() # block until a slot is open in the thread pool
result = analyze(results)
return result
Alternatively, you can just use the normal compute
function within a
task. This will automatically call secede
and rejoin
appropriately:
def f(name, fn):
df = dd.read_csv(fn) # note that this is a dask collection
result = df[df.name == name].count()
# This calls secede
# Then runs the computation on the cluster (including this worker)
# Then blocks on rejoin, and finally delivers the answer
result = result.compute()
return result
Coordination Primitives
Sometimes situations arise where tasks, workers, or clients need to coordinate with each other in ways beyond normal task scheduling with futures. In these cases Dask provides additional primitives to help in complex situations.
Dask provides distributed versions of coordination primitives like locks, events, queues, global variables, and pub-sub systems that, where appropriate, match their in-memory counterparts. These can be used to control access to external resources, track progress of ongoing computations, or share data in side-channels between many workers, clients, and tasks sensibly.
These features are rarely necessary for common use of Dask. We recommend that
beginning users stick with using the simpler futures found above (like
Client.submit
and Client.gather
) rather than embracing needlessly
complex techniques.
Queues
Dask queues follow the API for the standard Python Queue, but now move futures or small messages between clients. Queues serialize sensibly and reconnect themselves on remote clients if necessary:
from dask.distributed import Queue
def load_and_submit(filename):
data = load(filename)
client = get_client()
future = client.submit(process, data)
queue.put(future)
client = Client()
queue = Queue()
for filename in filenames:
future = client.submit(load_and_submit, filename)
fire_and_forget(future)
while True:
future = queue.get()
print(future.result())
Queues can also send small pieces of information, anything that is msgpack encodable (ints, strings, bools, lists, dicts, etc.). This can be useful to send back small scores or administrative messages:
def func(x):
try:
...
except Exception as e:
error_queue.put(str(e))
error_queue = Queue()
Queues are mediated by the central scheduler, and so they are not ideal for sending large amounts of data (everything you send will be routed through a central point). They are well suited to move around small bits of metadata, or futures. These futures may point to much larger pieces of data safely:
>>> x = ... # my large numpy array
# Don't do this!
>>> q.put(x)
# Do this instead
>>> future = client.scatter(x)
>>> q.put(future)
# Or use futures for metadata
>>> q.put({'status': 'OK', 'stage=': 1234})
If you’re looking to move large amounts of data between workers, then you might also want to consider the Pub/Sub system described a few sections below.
Global Variables
Variables are like Queues in that they communicate futures and small data between clients. However, variables hold only a single value. You can get or set that value at any time:
>>> var = Variable('stopping-criterion')
>>> var.set(False)
>>> var.get()
False
This is often used to signal stopping criteria or current parameters between clients.
If you want to share large pieces of information, then scatter the data first:
>>> parameters = np.array(...)
>>> future = client.scatter(parameters)
>>> var.set(future)
Locks
You can also hold onto cluster-wide locks using the Lock
object.
Dask Locks have the same API as normal threading.Lock
objects, except that
they work across the cluster:
from dask.distributed import Lock
lock = Lock()
with lock:
# access protected resource
You can manage several locks at the same time. Lock can either be given a consistent name or you can pass the lock object around itself.
Using a consistent name is convenient when you want to lock some known named resource:
from dask.distributed import Lock
def load(fn):
with Lock('the-production-database'):
# read data from filename using some sensitive source
return ...
futures = client.map(load, filenames)
Passing around a lock works as well and is easier when you want to create short-term locks for a particular situation:
from dask.distributed import Lock
lock = Lock()
def load(fn, lock=None):
with lock:
# read data from filename using some sensitive source
return ...
futures = client.map(load, filenames, lock=lock)
This can be useful if you want to control concurrent access to some external resource like a database or un-thread-safe library.
Events
Dask Events mimic asyncio.Event
objects, but on a cluster scope.
They hold a single flag which can be set or cleared.
Clients can wait until the event flag is set.
Different from a Lock
, every client can set or clear the flag and there
is no “ownership” of an event.
You can use events to e.g. synchronize multiple clients:
# One one client
from dask.distributed import Event
event = Event("my-event-1")
event.wait()
The call to wait will block until the event is set, e.g. in another client
# In another client
from dask.distributed import Event
event = Event("my-event-1")
# do some work
event.set()
Events can be set, cleared and waited on multiple times. Every waiter referencing the same event name will be notified on event set (and not only the first one as in the case of a lock):
from dask.distributed import Event
def wait_for_event(x):
event = Event("my-event")
event.wait()
# at this point, all function calls
# are in sync once the event is set
futures = client.map(wait_for_event, range(10))
Event("my-event").set()
client.gather(futures)
Semaphore
Similar to the single-valued Lock
it is also possible to use a cluster-wide
semaphore to coordinate and limit access to a sensitive resource like a
database.
from dask.distributed import Semaphore
sem = Semaphore(max_leases=2, name="database")
def access_limited(val, sem):
with sem:
# Interact with the DB
return
futures = client.map(access_limited, range(10), sem=sem)
client.gather(futures)
sem.close()
Publish-Subscribe
Dask implements the Publish Subscribe pattern, providing an additional channel of communication between ongoing tasks.
Actors
Actors allow workers to manage rapidly changing state without coordinating with the central scheduler. This has the advantage of reducing latency (worker-to-worker roundtrip latency is around 1ms), reducing pressure on the centralized scheduler (workers can coordinate actors entirely among each other), and also enabling workflows that require stateful or in-place memory manipulation.
However, these benefits come at a cost. The scheduler is unaware of actors and so they don’t benefit from diagnostics, load balancing, or resilience. Once an actor is running on a worker it is forever tied to that worker. If that worker becomes overburdened or dies, then there is no opportunity to recover the workload.
Because Actors avoid the central scheduler they can be high-performing, but not resilient.
Example: Counter
An actor is a class containing both state and methods that is submitted to a worker:
class Counter:
n = 0
def __init__(self):
self.n = 0
def increment(self):
self.n += 1
return self.n
from dask.distributed import Client
client = Client()
future = client.submit(Counter, actor=True)
counter = future.result()
>>> counter
<Actor: Counter, key=Counter-afa1cdfb6b4761e616fa2cfab42398c8>
Method calls on this object produce ActorFutures
, which are similar to
normal Futures, but interact only with the worker holding the Actor:
>>> future = counter.increment()
>>> future
<ActorFuture>
>>> future.result()
1
Attribute access is synchronous and blocking:
>>> counter.n
1
Example: Parameter Server
This example will perform the following minimization with a parameter server:
This is a simple minimization that will serve as an illustrative example.
The Dask Actor will serve as the parameter server that will hold the model. The client will calculate the gradient of the loss function above.
import numpy as np
from dask.distributed import Client
client = Client(processes=False)
class ParameterServer:
def __init__(self):
self.data = dict()
def put(self, key, value):
self.data[key] = value
def get(self, key):
return self.data[key]
def train(params, lr=0.1):
grad = 2 * (params - 1) # gradient of (params - 1)**2
new_params = params - lr * grad
return new_params
ps_future = client.submit(ParameterServer, actor=True)
ps = ps_future.result()
ps.put('parameters', np.random.random(1000))
for k in range(20):
params = ps.get('parameters').result()
new_params = train(params)
ps.put('parameters', new_params)
print(new_params.mean())
# k=0: "0.5988202981316124"
# k=10: "0.9569236575164062"
This example works, and the loss function is minimized. The (simple) equation above is minimize, so each \(p_i\) converges to 1. If desired, this example could be adapted to machine learning with a more complex function to minimize.
Asynchronous Operation
All operations that require talking to the remote worker are awaitable:
async def f():
future = client.submit(Counter, actor=True)
counter = await future # gather actor object locally
counter.increment() # send off a request asynchronously
await counter.increment() # or wait until it was received
n = await counter.n # attribute access also must be awaited