from __future__ import annotations
import contextlib
import functools
import inspect
import os
import sys
import threading
import warnings
from collections.abc import Iterator
from typing import Dict, Optional, Union
import dask.config
import dask.distributed
import toolz
from dask.base import tokenize
from dask.system import CPU_COUNT
from dask.utils import parse_timedelta
from typing_extensions import Literal
import coiled
from coiled.spans import span
from coiled.utils import error_info_for_tracking
_clients = {}
_lock = threading.RLock()
_clusters = {}
@contextlib.contextmanager
def _set_local_environ(environ: Dict[str, str]):
"""Temporarily set local environment variables"""
original_environ = dict(os.environ)
os.environ.update(environ)
try:
yield
finally:
os.environ.clear()
os.environ.update(original_environ)
class Function:
"""A function that you can run remotely"""
def __init__(self, function, cluster_kwargs, keepalive, environ, local: bool = False, name: Optional[str] = None):
self.function = function
self._cluster_kwargs = cluster_kwargs
self.keepalive = parse_timedelta(keepalive)
self._local = local
self._environ = environ
if isinstance(function, functools.partial): # Special case partial functions
_code = inspect.getsource(function.func)
else:
try:
_code = inspect.getsource(function)
except Exception:
_code = ""
self._code = _code
if name is not None:
self._name = name
else:
token = tokenize(
sys.executable,
local,
environ,
keepalive,
# TODO: include something about the software environment
**cluster_kwargs,
)
self._name = f"function-{token[:8]}"
@property
def __signature__(self):
return inspect.signature(self.function)
@property
def cluster_status(self) -> Optional[dask.distributed.Status]:
if self._name in _clusters:
return _clusters[self._name].status
return None
def _clear_terminal_cluster(self):
# we can't use a cluster that's stopping or stopped, so check if we need to remake cluster
if not self._local:
# note that status will be `None` if there's no cluster
if self.cluster_status in coiled.v2.cluster.TERMINATING_STATES:
_clusters.pop(self._name, None)
_clients.pop(self._name, None)
@property
def cluster(self) -> Union[coiled.Cluster, dask.distributed.LocalCluster]:
with _lock:
self._clear_terminal_cluster()
try:
return _clusters[self._name]
except KeyError:
success = True
exception = None
info = {}
info["keepalive"] = self.keepalive
info["local"] = self._local
try:
# Setting to use the local threaded scheduler avoids implicit tasks in tasks.
# This relies on `send_dask_config=True` (default value).
with dask.config.set({
"scheduler": "threads",
"distributed.worker.daemon": False,
"distributed.worker.memory.target": 0.9,
"distributed.worker.memory.spill": 0.9,
"distributed.worker.memory.pause": False,
"distributed.worker.memory.terminate": False,
}):
n_workers = self._cluster_kwargs["n_workers"]
if self._local:
n_workers = max(n_workers, 1)
with _set_local_environ(self._environ or {}):
cluster = dask.distributed.LocalCluster(
name=self._name,
n_workers=n_workers,
threads_per_worker=self._cluster_kwargs["worker_options"]["nthreads"],
)
cluster.adapt(minimum=0, maximum=CPU_COUNT or 10)
else:
cluster = coiled.Cluster(name=self._name, **self._cluster_kwargs)
info["account"] = cluster.account
info["cluster_id"] = cluster.cluster_id
if self._environ:
cluster.send_private_envs(self._environ)
# some env vars are set in dask config early in the dask loading process
# we set these to `""` since that's what we can do in the relevant dask config
# but some libraries don't accept empty string, so we'll unset them here.
# this won't affect things loaded by dask, but should affect user-code.
user_specified_env = {
**(self._environ or {}),
**(self._cluster_kwargs.get("environ") or {}),
}
single_thread_unsets = [
key
for key in coiled.utils.unset_single_thread_defaults().keys()
if not user_specified_env.get(key)
]
cluster.unset_env_vars(single_thread_unsets)
cluster.adapt(minimum=0, maximum=500)
_clusters[self._name] = cluster
return cluster
except Exception as e:
success = False
exception = e
raise
finally:
coiled.add_interaction(
"coiled-function",
success=success,
**info,
**error_info_for_tracking(exception),
)
@property
def client(self) -> dask.distributed.Client:
with _lock:
self._clear_terminal_cluster()
try:
return _clients[self._name]
except KeyError:
client = dask.distributed.Client(self.cluster, set_as_default=False)
if self.cluster.shutdown_on_close and isinstance(self.cluster, coiled.Cluster):
self.cluster.set_keepalive(keepalive=self.keepalive)
_clients[self._name] = client
return client
def __call__(self, *args, **kwargs):
# If this is being called from on the desired cluster, then run locally.
# This allows one Function (with same desired cluster specs) to call another without overhead.
if os.environ.get("COILED_CLUSTER_NAME", None) == self._name:
return self.local(*args, **kwargs)
# Otherwise, submit to cluster.
return self.submit(*args, **kwargs).result()
@property
def local(self):
return Function(
self.function,
cluster_kwargs=self._cluster_kwargs,
keepalive=self.keepalive,
environ=self._environ,
local=True,
)
def submit(self, *args, **kwargs) -> dask.distributed.Future:
"""Submit function call for asynchronous execution
This immediately returns a Dask Future, allowing for the submission of
many tasks in parallel.
Example
-------
>>> @coiled.function()
... def f(x):
... return x + 1
>>> f(10) # calling the function blocks until finished
11
>>> f.submit(10) # immediately returns a future
<Future: pending, key=f-1234>
>>> f.submit(10).result() # Call .result to get result
11
>>> futures = [f(i) for i in range(1000)] # parallelize with a for loop
>>> [future.result() for future in futures]
...
Returns
-------
future: dask.distributed.Future
See Also
--------
Function.map
"""
with span(self.cluster, callstack=[{"code": self._code, "relative_line": 0}]):
result = self.client.submit(self.function, *args, **kwargs)
return result
def map(self, *args, errors: Literal["raise", "skip"] = "raise", **kwargs) -> Iterator:
"""Map function across many inputs
This runs your function many times in parallel across all of the items
in an input list. Coiled will auto-scale your cluster to meet demand.
Parameters
----------
errors
Either 'raise' or 'skip' if we should raise if a function call has erred
or include ``None`` in the output collection.
Example
-------
>>> @coiled.function()
... def process(filename: str):
... " Convert CSV file to Parquet "
... df = pd.read_csv(filename)
... outfile = filename[:-4] + ".parquet"
... df.to_parquet(outfile)
... return outfile
>>> process("s3://my-bucket/data.csv") # calling the function blocks until finished
11
>>> filenames = process.map(filenames)
>>> print(list(filenames)) # print out all output filenames
Returns
-------
results: Iterator
See Also
--------
Function.submit
"""
kwargs.setdefault("pure", False)
kwargs.setdefault("batch_size", 100)
with span(self.cluster, callstack=[{"code": self._code, "relative_line": 0}]):
futures = self.client.map(self.function, *args, **kwargs)
batchsize = max(int(len(futures) / 50), 1) # type: ignore
batches = toolz.partition_all(batchsize, futures)
return (result for batch in batches for result in self.client.gather(batch, errors=errors)) # type: ignore
[docs]
def function(
*,
software: Optional[str] = None,
container: Optional[str] = None,
vm_type: Optional[Union[str, list[str]]] = None,
cpu: Optional[Union[int, list[int]]] = None,
memory: Optional[Union[str, list[str]]] = None,
gpu: Optional[bool] = None,
account: Optional[str] = None,
workspace: Optional[str] = None,
region: Optional[str] = None,
arm: Optional[bool] = None,
disk_size: Optional[Union[str, int]] = None,
allow_ingress_from: Optional[str] = None,
shutdown_on_close: bool = True,
spot_policy: Optional[str] = None,
idle_timeout: str = "24 hours",
keepalive="30 seconds",
package_sync_ignore: Optional[list[str]] = None,
environ: Optional[Dict[str, str]] = None,
threads_per_worker: Union[int, None] = 1,
local: bool = False,
name: Optional[str] = None,
tags: Optional[Dict[str, str]] = None,
):
"""
Decorate a function to run on cloud infrastructure
This creates a ``Function`` object that executes its code on a remote cluster
with the hardware and software specified in the arguments to the decorator.
It can run either as a normal function, or it can return Dask Futures for
parallel computing.
Parameters
----------
software
Name of the software environment to use; this allows you to use and re-use existing
Coiled software environments, and should not be used with package sync or when specifying
a container to use for this specific cluster.
container
Name or URI of container image to use; when using a pre-made container image with Coiled,
this allows you to skip the step of explicitly creating a Coiled software environment
from that image. Note that this should not be used with package sync or when specifying
an existing Coiled software environment.
vm_type
Instance type, or list of instance types, that you would like to use.
You can use ``coiled.list_instance_types()`` to see a list of allowed types.
cpu
Number, or range, of CPUs requested. Specify a range by
using a list of two elements, for example: ``cpu=[2, 8]``.
memory
Amount of memory to request for each VM, Coiled will use a +/- 10% buffer
from the memory that you specify. You may specify a range of memory by using a
list of two elements, for example: ``memory=["2GiB", "4GiB"]``.
disk_size
Size of persistent disk attached to each VM instance, specified as string with units
or integer for GiB.
gpu
Whether to attach a GPU; this would be a single NVIDIA T4.
account
**DEPRECATED**. Use ``workspace`` instead.
workspace
The Coiled workspace (previously "account") to use. If not specified,
will check the ``coiled.workspace`` or ``coiled.account`` configuration values,
or will use your default workspace if those aren't set.
region
The cloud provider region in which to run the cluster.
arm
Whether to use ARM instances for cluster; default is x86 (Intel) instances.
keepalive
Keep your cluster running for the specified time, even if your Python session closes.
Default is "30 seconds".
spot_policy
Purchase option to use for workers in your cluster, options are "on-demand", "spot", and
"spot_with_fallback"; by default this is "spot_with_fallback" for Coiled Functions.
(Google Cloud refers to this as "provisioning model" for your instances.)
Note that even with this option, the first VM is always on-demand. This only applies to any
additional VMs when running Coiled Functions in parallel across multiple
VMs with the ``.map()`` and ``.submit()`` methods. When running on a single VM, an on-demand
instance will be used.
Spot instances are much cheaper, but can have more limited availability and may be terminated
while you're still using them if the cloud provider needs more capacity for other customers.
On-demand instances have the best availability and are almost never
terminated while still in use, but they're significantly more expensive than spot instances.
For most workloads, "spot_with_fallback" is likely to be a good choice: Coiled will try to get as
many spot instances as we can, and if we get less than you requested, we'll try to get the remaining
instances as on-demand.
For AWS, when we're notified that an active spot instance is going to be terminated,
we'll attempt to get a replacement instance (spot if available, but could be on-demand if you've
enabled "fallback"). Dask on the active instance will attempt a graceful shutdown before the
instance is terminated so that computed results won't be lost.
idle_timeout
Shut down the cluster after this duration if no activity has occurred. Default is "24 hours".
package_sync_ignore
A list of package names to exclude from the cloud VM environment. This is useful when you have
large libraries installed locally that aren't needed for the function being run.
Note the packages listed here may still be installed by another package that depends on them.
environ
Dictionary of environment variables to securely pass to the cloud VM environment.
threads_per_worker
Number of threads to run concurrent tasks in for each VM. -1 can be used to run as many concurrent
tasks as there are CPU cores. Default is 1.
allow_ingress_from
Control the CIDR from which cluster firewall allows ingress to scheduler; by default this is open
to any source address (0.0.0.0/0). You can specify CIDR, or "me" for just your IP address.
local
Whether or not to run this function locally or on cloud VMs. If ``True``, this function will be
run on your local machine, which can be useful for debugging or during development.
Default is ``False``.
name
Name for the Coiled cluster on which this function will run. If not specified,
VM specification parameters like ``vm_type``, ``disk_size``, etc. will be used to produce
a unique, deterministic name. Note that ``name`` is used for sharing cloud VMs among
Coiled Functions with the same hardware and software specification, so please use
this parameter with care. Default to ``None``.
tags
Dictionary of tags.
See the :class:`coiled.Cluster` docstring for additional parameter descriptions.
Examples
--------
>>> import coiled
>>> @coiled.function()
... def f(x):
... return x + 1
>>> f(10) # calling the function blocks until finished
11
>>> f.submit(10) # immediately returns a future
<Future: pending, key=f-1234>
>>> f.submit(10).result() # Call .result to get result
11
>>> futures = [f(i) for i in range(1000)] # parallelize with a for loop
>>> [future.result() for future in futures]
...
"""
if workspace and account and workspace != account:
raise ValueError(
f"You specified both workspace='{workspace}' and account='{account}'. "
"The `account` kwarg is being deprecated, use `workspace` instead."
)
if account and not workspace:
warnings.warn("The `account` kwarg is deprecated, use `workspace` instead.", DeprecationWarning, stacklevel=2)
account = account or workspace
def decorator(func) -> Function:
nonlocal cpu, threads_per_worker, environ
default_environ = coiled.utils.unset_single_thread_defaults()
if container and "rapidsai" in container:
default_environ = {"DISABLE_JUPYTER": "true", **default_environ} # needed for "stable" RAPIDS image
if os.environ.get("COILED_REUSE_ENVIRON_KEYS"):
keys = os.environ["COILED_REUSE_ENVIRON_KEYS"].split(",")
default_environ.update({key: os.environ[key] for key in keys})
if memory is None and cpu is None and not vm_type and not gpu:
cpu = 2
if threads_per_worker == -1:
# Have `-1` mean the same as CPU count (Dask's default behavior)
threads_per_worker = None
cluster_kwargs = dict(
account=account,
n_workers=0,
scheduler_cpu=cpu,
scheduler_memory=memory,
worker_cpu=cpu,
worker_memory=memory,
software=software,
container=container,
idle_timeout=idle_timeout,
scheduler_vm_types=vm_type,
allow_ingress_from=allow_ingress_from,
worker_vm_types=vm_type,
allow_ssh_from="me",
environ=default_environ,
scheduler_gpu=gpu,
worker_gpu=gpu,
region=region,
arm=arm,
shutdown_on_close=shutdown_on_close,
spot_policy="spot_with_fallback" if spot_policy is None else spot_policy,
extra_worker_on_scheduler=True,
tags={**(tags or {}), **{"coiled-cluster-type": "function"}},
worker_options={"nthreads": threads_per_worker},
scheduler_disk_size=disk_size,
worker_disk_size=disk_size,
package_sync_ignore=package_sync_ignore,
)
return functools.wraps(func)(
Function(func, cluster_kwargs, keepalive=keepalive, environ=environ, local=local, name=name)
)
return decorator
# Small backwards compatibility shim
def run(*args, **kwargs):
warnings.warn(
"coiled.run has been renamed to coiled.function. "
"Please use coiled.function as coiled.run will be removed in a future release.",
FutureWarning,
stacklevel=2,
)
return function(*args, **kwargs)