from collections import OrderedDict
from contextlib import contextmanager
from functools import partial
from hashlib import md5
from numbers import Number
from operator import getitem
from typing import Iterator, Mapping, Set
import inspect
import pickle
import os
import threading
import uuid
from tlz import merge, groupby, curry, identity
from tlz.functoolz import Compose
from .compatibility import is_dataclass, dataclass_fields
from .context import thread_state
from .core import flatten, quote, get as simple_get, literal
from .hashing import hash_buffer_hex
from .utils import Dispatch, ensure_dict, apply, key_split
from . import config, local, threaded
__all__ = (
"DaskMethodsMixin",
"annotate",
"is_dask_collection",
"compute",
"persist",
"optimize",
"visualize",
"tokenize",
"normalize_token",
"get_collection_names",
"get_name_from_key",
"replace_name_in_key",
"clone_key",
)
[docs]@contextmanager
def annotate(**annotations):
"""Context Manager for setting HighLevelGraph Layer annotations.
Annotations are metadata or soft constraints associated with
tasks that dask schedulers may choose to respect: They signal intent
without enforcing hard constraints. As such, they are
primarily designed for use with the distributed scheduler.
Almost any object can serve as an annotation, but small Python objects
are preferred, while large objects such as NumPy arrays are discouraged.
Callables supplied as an annotation should take a single *key* argument and
produce the appropriate annotation. Individual task keys in the annotated collection
are supplied to the callable.
Parameters
----------
**annotations : key-value pairs
Examples
--------
All tasks within array A should have priority 100 and be retried 3 times
on failure.
>>> import dask
>>> import dask.array as da
>>> with dask.annotate(priority=100, retries=3):
... A = da.ones((10000, 10000))
Prioritise tasks within Array A on flattened block ID.
>>> nblocks = (10, 10)
>>> with dask.annotate(priority=lambda k: k[1]*nblocks[1] + k[2]):
... A = da.ones((1000, 1000), chunks=(100, 100))
Annotations may be nested.
>>> with dask.annotate(priority=1):
... with dask.annotate(retries=3):
... A = da.ones((1000, 1000))
... B = A + 1
"""
# Sanity check annotations used in place of
# legacy distributed Client.{submit, persist, compute} keywords
if "workers" in annotations:
if isinstance(annotations["workers"], (list, set, tuple)):
annotations["workers"] = list(annotations["workers"])
elif isinstance(annotations["workers"], str):
annotations["workers"] = [annotations["workers"]]
elif callable(annotations["workers"]):
pass
else:
raise TypeError(
"'workers' annotation must be a sequence of str, a str or a callable, but got %s."
% annotations["workers"]
)
if (
"priority" in annotations
and not isinstance(annotations["priority"], Number)
and not callable(annotations["priority"])
):
raise TypeError(
"'priority' annotation must be a Number or a callable, but got %s"
% annotations["priority"]
)
if (
"retries" in annotations
and not isinstance(annotations["retries"], Number)
and not callable(annotations["retries"])
):
raise TypeError(
"'retries' annotation must be a Number or a callable, but got %s"
% annotations["retries"]
)
if (
"resources" in annotations
and not isinstance(annotations["resources"], dict)
and not callable(annotations["resources"])
):
raise TypeError(
"'resources' annotation must be a dict, but got %s"
% annotations["resources"]
)
if (
"allow_other_workers" in annotations
and not isinstance(annotations["allow_other_workers"], bool)
and not callable(annotations["allow_other_workers"])
):
raise TypeError(
"'allow_other_workers' annotations must be a bool or a callable, but got %s"
% annotations["allow_other_workers"]
)
prev_annotations = config.get("annotations", {})
new_annotations = {
**prev_annotations,
**{f"annotations.{k}": v for k, v in annotations.items()},
}
with config.set(new_annotations):
yield
[docs]def is_dask_collection(x):
"""Returns ``True`` if ``x`` is a dask collection"""
try:
return x.__dask_graph__() is not None
except (AttributeError, TypeError):
return False
class DaskMethodsMixin:
"""A mixin adding standard dask collection methods"""
__slots__ = ()
def visualize(self, filename="mydask", format=None, optimize_graph=False, **kwargs):
"""Render the computation of this object's task graph using graphviz.
Requires ``graphviz`` to be installed.
Parameters
----------
filename : str or None, optional
The name of the file to write to disk. If the provided `filename`
doesn't include an extension, '.png' will be used by default.
If `filename` is None, no file will be written, and we communicate
with dot using only pipes.
format : {'png', 'pdf', 'dot', 'svg', 'jpeg', 'jpg'}, optional
Format in which to write output file. Default is 'png'.
optimize_graph : bool, optional
If True, the graph is optimized before rendering. Otherwise,
the graph is displayed as is. Default is False.
color: {None, 'order'}, optional
Options to color nodes. Provide ``cmap=`` keyword for additional
colormap
**kwargs
Additional keyword arguments to forward to ``to_graphviz``.
Examples
--------
>>> x.visualize(filename='dask.pdf') # doctest: +SKIP
>>> x.visualize(filename='dask.pdf', color='order') # doctest: +SKIP
Returns
-------
result : IPython.diplay.Image, IPython.display.SVG, or None
See dask.dot.dot_graph for more information.
See Also
--------
dask.base.visualize
dask.dot.dot_graph
Notes
-----
For more information on optimization see here:
https://docs.dask.org/en/latest/optimize.html
"""
return visualize(
self,
filename=filename,
format=format,
optimize_graph=optimize_graph,
**kwargs,
)
def persist(self, **kwargs):
"""Persist this dask collection into memory
This turns a lazy Dask collection into a Dask collection with the same
metadata, but now with the results fully computed or actively computing
in the background.
The action of function differs significantly depending on the active
task scheduler. If the task scheduler supports asynchronous computing,
such as is the case of the dask.distributed scheduler, then persist
will return *immediately* and the return value's task graph will
contain Dask Future objects. However if the task scheduler only
supports blocking computation then the call to persist will *block*
and the return value's task graph will contain concrete Python results.
This function is particularly useful when using distributed systems,
because the results will be kept in distributed memory, rather than
returned to the local process as with compute.
Parameters
----------
scheduler : string, optional
Which scheduler to use like "threads", "synchronous" or "processes".
If not provided, the default is to check the global settings first,
and then fall back to the collection defaults.
optimize_graph : bool, optional
If True [default], the graph is optimized before computation.
Otherwise the graph is run as is. This can be useful for debugging.
**kwargs
Extra keywords to forward to the scheduler function.
Returns
-------
New dask collections backed by in-memory data
See Also
--------
dask.base.persist
"""
(result,) = persist(self, traverse=False, **kwargs)
return result
def compute(self, **kwargs):
"""Compute this dask collection
This turns a lazy Dask collection into its in-memory equivalent.
For example a Dask array turns into a NumPy array and a Dask dataframe
turns into a Pandas dataframe. The entire dataset must fit into memory
before calling this operation.
Parameters
----------
scheduler : string, optional
Which scheduler to use like "threads", "synchronous" or "processes".
If not provided, the default is to check the global settings first,
and then fall back to the collection defaults.
optimize_graph : bool, optional
If True [default], the graph is optimized before computation.
Otherwise the graph is run as is. This can be useful for debugging.
kwargs
Extra keywords to forward to the scheduler function.
See Also
--------
dask.base.compute
"""
(result,) = compute(self, traverse=False, **kwargs)
return result
def __await__(self):
try:
from distributed import wait, futures_of
except ImportError as e:
raise ImportError(
"Using async/await with dask requires the `distributed` package"
) from e
from tornado import gen
@gen.coroutine
def f():
if futures_of(self):
yield wait(self)
raise gen.Return(self)
return f().__await__()
def compute_as_if_collection(cls, dsk, keys, scheduler=None, get=None, **kwargs):
"""Compute a graph as if it were of type cls.
Allows for applying the same optimizations and default scheduler."""
schedule = get_scheduler(scheduler=scheduler, cls=cls, get=get)
dsk2 = optimization_function(cls)(ensure_dict(dsk), keys, **kwargs)
return schedule(dsk2, keys, **kwargs)
def dont_optimize(dsk, keys, **kwargs):
return dsk
def optimization_function(x):
return getattr(x, "__dask_optimize__", dont_optimize)
def collections_to_dsk(collections, optimize_graph=True, optimizations=(), **kwargs):
"""
Convert many collections into a single dask graph, after optimization
"""
from .highlevelgraph import HighLevelGraph
optimizations = tuple(optimizations) + tuple(config.get("optimizations", ()))
if optimize_graph:
groups = groupby(optimization_function, collections)
graphs = []
for opt, val in groups.items():
dsk, keys = _extract_graph_and_keys(val)
dsk = opt(dsk, keys, **kwargs)
for opt in optimizations:
dsk = opt(dsk, keys, **kwargs)
graphs.append(dsk)
# Merge all graphs
if any(isinstance(graph, HighLevelGraph) for graph in graphs):
dsk = HighLevelGraph.merge(*graphs)
else:
dsk = merge(*map(ensure_dict, graphs))
else:
dsk, _ = _extract_graph_and_keys(collections)
return dsk
def _extract_graph_and_keys(vals):
"""Given a list of dask vals, return a single graph and a list of keys such
that ``get(dsk, keys)`` is equivalent to ``[v.compute() for v in vals]``."""
from .highlevelgraph import HighLevelGraph
graphs, keys = [], []
for v in vals:
graphs.append(v.__dask_graph__())
keys.append(v.__dask_keys__())
if any(isinstance(graph, HighLevelGraph) for graph in graphs):
graph = HighLevelGraph.merge(*graphs)
else:
graph = merge(*map(ensure_dict, graphs))
return graph, keys
def unpack_collections(*args, **kwargs):
"""Extract collections in preparation for compute/persist/etc...
Intended use is to find all collections in a set of (possibly nested)
python objects, do something to them (compute, etc...), then repackage them
in equivalent python objects.
Parameters
----------
*args
Any number of objects. If it is a dask collection, it's extracted and
added to the list of collections returned. By default, python builtin
collections are also traversed to look for dask collections (for more
information see the ``traverse`` keyword).
traverse : bool, optional
If True (default), builtin python collections are traversed looking for
any dask collections they might contain.
Returns
-------
collections : list
A list of all dask collections contained in ``args``
repack : callable
A function to call on the transformed collections to repackage them as
they were in the original ``args``.
"""
traverse = kwargs.pop("traverse", True)
collections = []
repack_dsk = {}
collections_token = uuid.uuid4().hex
def _unpack(expr):
if is_dask_collection(expr):
tok = tokenize(expr)
if tok not in repack_dsk:
repack_dsk[tok] = (getitem, collections_token, len(collections))
collections.append(expr)
return tok
tok = uuid.uuid4().hex
if not traverse:
tsk = quote(expr)
else:
# Treat iterators like lists
typ = list if isinstance(expr, Iterator) else type(expr)
if typ in (list, tuple, set):
tsk = (typ, [_unpack(i) for i in expr])
elif typ in (dict, OrderedDict):
tsk = (typ, [[_unpack(k), _unpack(v)] for k, v in expr.items()])
elif is_dataclass(expr) and not isinstance(expr, type):
tsk = (
apply,
typ,
(),
(
dict,
[
[f.name, _unpack(getattr(expr, f.name))]
for f in dataclass_fields(expr)
],
),
)
else:
return expr
repack_dsk[tok] = tsk
return tok
out = uuid.uuid4().hex
repack_dsk[out] = (tuple, [_unpack(i) for i in args])
def repack(results):
dsk = repack_dsk.copy()
dsk[collections_token] = quote(results)
return simple_get(dsk, out)
return collections, repack
[docs]def optimize(*args, **kwargs):
"""Optimize several dask collections at once.
Returns equivalent dask collections that all share the same merged and
optimized underlying graph. This can be useful if converting multiple
collections to delayed objects, or to manually apply the optimizations at
strategic points.
Note that in most cases you shouldn't need to call this method directly.
Parameters
----------
*args : objects
Any number of objects. If a dask object, its graph is optimized and
merged with all those of all other dask objects before returning an
equivalent dask collection. Non-dask arguments are passed through
unchanged.
traverse : bool, optional
By default dask traverses builtin python collections looking for dask
objects passed to ``optimize``. For large collections this can be
expensive. If none of the arguments contain any dask objects, set
``traverse=False`` to avoid doing this traversal.
optimizations : list of callables, optional
Additional optimization passes to perform.
**kwargs
Extra keyword arguments to forward to the optimization passes.
Examples
--------
>>> import dask as d
>>> import dask.array as da
>>> a = da.arange(10, chunks=2).sum()
>>> b = da.arange(10, chunks=2).mean()
>>> a2, b2 = d.optimize(a, b)
>>> a2.compute() == a.compute()
True
>>> b2.compute() == b.compute()
True
"""
collections, repack = unpack_collections(*args, **kwargs)
if not collections:
return args
dsk = collections_to_dsk(collections, **kwargs)
postpersists = []
for a in collections:
r, s = a.__dask_postpersist__()
postpersists.append(r(dsk, *s))
return repack(postpersists)
[docs]def compute(*args, **kwargs):
"""Compute several dask collections at once.
Parameters
----------
args : object
Any number of objects. If it is a dask object, it's computed and the
result is returned. By default, python builtin collections are also
traversed to look for dask objects (for more information see the
``traverse`` keyword). Non-dask arguments are passed through unchanged.
traverse : bool, optional
By default dask traverses builtin python collections looking for dask
objects passed to ``compute``. For large collections this can be
expensive. If none of the arguments contain any dask objects, set
``traverse=False`` to avoid doing this traversal.
scheduler : string, optional
Which scheduler to use like "threads", "synchronous" or "processes".
If not provided, the default is to check the global settings first,
and then fall back to the collection defaults.
optimize_graph : bool, optional
If True [default], the optimizations for each collection are applied
before computation. Otherwise the graph is run as is. This can be
useful for debugging.
kwargs
Extra keywords to forward to the scheduler function.
Examples
--------
>>> import dask as d
>>> import dask.array as da
>>> a = da.arange(10, chunks=2).sum()
>>> b = da.arange(10, chunks=2).mean()
>>> d.compute(a, b)
(45, 4.5)
By default, dask objects inside python collections will also be computed:
>>> d.compute({'a': a, 'b': b, 'c': 1})
({'a': 45, 'b': 4.5, 'c': 1},)
"""
traverse = kwargs.pop("traverse", True)
optimize_graph = kwargs.pop("optimize_graph", True)
collections, repack = unpack_collections(*args, traverse=traverse)
if not collections:
return args
schedule = get_scheduler(
scheduler=kwargs.pop("scheduler", None),
collections=collections,
get=kwargs.pop("get", None),
)
dsk = collections_to_dsk(collections, optimize_graph, **kwargs)
keys, postcomputes = [], []
for x in collections:
keys.append(x.__dask_keys__())
postcomputes.append(x.__dask_postcompute__())
results = schedule(dsk, keys, **kwargs)
return repack([f(r, *a) for r, (f, a) in zip(results, postcomputes)])
[docs]def visualize(*args, **kwargs):
"""
Visualize several dask graphs at once.
Requires ``graphviz`` to be installed. All options that are not the dask
graph(s) should be passed as keyword arguments.
Parameters
----------
dsk : dict(s) or collection(s)
The dask graph(s) to visualize.
filename : str or None, optional
The name of the file to write to disk. If the provided `filename`
doesn't include an extension, '.png' will be used by default.
If `filename` is None, no file will be written, and we communicate
with dot using only pipes.
format : {'png', 'pdf', 'dot', 'svg', 'jpeg', 'jpg'}, optional
Format in which to write output file. Default is 'png'.
optimize_graph : bool, optional
If True, the graph is optimized before rendering. Otherwise,
the graph is displayed as is. Default is False.
color : {None, 'order'}, optional
Options to color nodes. Provide ``cmap=`` keyword for additional
colormap
collapse_outputs : bool, optional
Whether to collapse output boxes, which often have empty labels.
Default is False.
verbose : bool, optional
Whether to label output and input boxes even if the data aren't chunked.
Beware: these labels can get very long. Default is False.
**kwargs
Additional keyword arguments to forward to ``to_graphviz``.
Examples
--------
>>> x.visualize(filename='dask.pdf') # doctest: +SKIP
>>> x.visualize(filename='dask.pdf', color='order') # doctest: +SKIP
Returns
-------
result : IPython.diplay.Image, IPython.display.SVG, or None
See dask.dot.dot_graph for more information.
See Also
--------
dask.dot.dot_graph
Notes
-----
For more information on optimization see here:
https://docs.dask.org/en/latest/optimize.html
"""
from dask.dot import dot_graph
filename = kwargs.pop("filename", "mydask")
optimize_graph = kwargs.pop("optimize_graph", False)
dsks = []
args3 = []
for arg in args:
if isinstance(arg, (list, tuple, set)):
for a in arg:
if isinstance(a, Mapping):
dsks.append(a)
if is_dask_collection(a):
args3.append(a)
else:
if isinstance(arg, Mapping):
dsks.append(arg)
if is_dask_collection(arg):
args3.append(arg)
dsk = dict(collections_to_dsk(args3, optimize_graph=optimize_graph))
for d in dsks:
dsk.update(d)
color = kwargs.get("color")
if color == "order":
from .order import order
import matplotlib.pyplot as plt
o = order(dsk)
try:
cmap = kwargs.pop("cmap")
except KeyError:
cmap = plt.cm.RdBu
if isinstance(cmap, str):
import matplotlib.pyplot as plt
cmap = getattr(plt.cm, cmap)
mx = max(o.values()) + 1
colors = {k: _colorize(cmap(v / mx, bytes=True)) for k, v in o.items()}
kwargs["function_attributes"] = {
k: {"color": v, "label": str(o[k])} for k, v in colors.items()
}
kwargs["data_attributes"] = {k: {"color": v} for k, v in colors.items()}
elif color:
raise NotImplementedError("Unknown value color=%s" % color)
return dot_graph(dsk, filename=filename, **kwargs)
[docs]def persist(*args, **kwargs):
"""Persist multiple Dask collections into memory
This turns lazy Dask collections into Dask collections with the same
metadata, but now with their results fully computed or actively computing
in the background.
For example a lazy dask.array built up from many lazy calls will now be a
dask.array of the same shape, dtype, chunks, etc., but now with all of
those previously lazy tasks either computed in memory as many small :class:`numpy.array`
(in the single-machine case) or asynchronously running in the
background on a cluster (in the distributed case).
This function operates differently if a ``dask.distributed.Client`` exists
and is connected to a distributed scheduler. In this case this function
will return as soon as the task graph has been submitted to the cluster,
but before the computations have completed. Computations will continue
asynchronously in the background. When using this function with the single
machine scheduler it blocks until the computations have finished.
When using Dask on a single machine you should ensure that the dataset fits
entirely within memory.
Examples
--------
>>> df = dd.read_csv('/path/to/*.csv') # doctest: +SKIP
>>> df = df[df.name == 'Alice'] # doctest: +SKIP
>>> df['in-debt'] = df.balance < 0 # doctest: +SKIP
>>> df = df.persist() # triggers computation # doctest: +SKIP
>>> df.value().min() # future computations are now fast # doctest: +SKIP
-10
>>> df.value().max() # doctest: +SKIP
100
>>> from dask import persist # use persist function on multiple collections
>>> a, b = persist(a, b) # doctest: +SKIP
Parameters
----------
*args: Dask collections
scheduler : string, optional
Which scheduler to use like "threads", "synchronous" or "processes".
If not provided, the default is to check the global settings first,
and then fall back to the collection defaults.
traverse : bool, optional
By default dask traverses builtin python collections looking for dask
objects passed to ``persist``. For large collections this can be
expensive. If none of the arguments contain any dask objects, set
``traverse=False`` to avoid doing this traversal.
optimize_graph : bool, optional
If True [default], the graph is optimized before computation.
Otherwise the graph is run as is. This can be useful for debugging.
**kwargs
Extra keywords to forward to the scheduler function.
Returns
-------
New dask collections backed by in-memory data
"""
traverse = kwargs.pop("traverse", True)
optimize_graph = kwargs.pop("optimize_graph", True)
collections, repack = unpack_collections(*args, traverse=traverse)
if not collections:
return args
schedule = get_scheduler(
scheduler=kwargs.pop("scheduler", None), collections=collections
)
if inspect.ismethod(schedule):
try:
from distributed.client import default_client
except ImportError:
pass
else:
try:
client = default_client()
except ValueError:
pass
else:
if client.get == schedule:
results = client.persist(
collections, optimize_graph=optimize_graph, **kwargs
)
return repack(results)
dsk = collections_to_dsk(collections, optimize_graph, **kwargs)
keys, postpersists = [], []
for a in collections:
a_keys = list(flatten(a.__dask_keys__()))
rebuild, state = a.__dask_postpersist__()
keys.extend(a_keys)
postpersists.append((rebuild, a_keys, state))
results = schedule(dsk, keys, **kwargs)
d = dict(zip(keys, results))
results2 = [r({k: d[k] for k in ks}, *s) for r, ks, s in postpersists]
return repack(results2)
############
# Tokenize #
############
def tokenize(*args, **kwargs):
"""Deterministic token
>>> tokenize([1, 2, '3'])
'7d6a880cd9ec03506eee6973ff551339'
>>> tokenize('Hello') == tokenize('Hello')
True
"""
if kwargs:
args = args + (kwargs,)
return md5(str(tuple(map(normalize_token, args))).encode()).hexdigest()
normalize_token = Dispatch()
normalize_token.register(
(int, float, str, bytes, type(None), type, slice, complex, type(Ellipsis)), identity
)
@normalize_token.register(dict)
def normalize_dict(d):
return normalize_token(sorted(d.items(), key=str))
@normalize_token.register(OrderedDict)
def normalize_ordered_dict(d):
return type(d).__name__, normalize_token(list(d.items()))
@normalize_token.register(set)
def normalize_set(s):
return normalize_token(sorted(s, key=str))
@normalize_token.register((tuple, list))
def normalize_seq(seq):
def func(seq):
try:
return list(map(normalize_token, seq))
except RecursionError:
return str(uuid.uuid4())
return type(seq).__name__, func(seq)
@normalize_token.register(literal)
def normalize_literal(lit):
return "literal", normalize_token(lit())
@normalize_token.register(range)
def normalize_range(r):
return list(map(normalize_token, [r.start, r.stop, r.step]))
@normalize_token.register(object)
def normalize_object(o):
method = getattr(o, "__dask_tokenize__", None)
if method is not None:
return method()
return normalize_function(o) if callable(o) else uuid.uuid4().hex
function_cache = {}
function_cache_lock = threading.Lock()
def normalize_function(func):
try:
return function_cache[func]
except KeyError:
result = _normalize_function(func)
if len(function_cache) >= 500: # clear half of cache if full
with function_cache_lock:
if len(function_cache) >= 500:
for k in list(function_cache)[::2]:
del function_cache[k]
function_cache[func] = result
return result
except TypeError: # not hashable
return _normalize_function(func)
def _normalize_function(func):
if isinstance(func, Compose):
first = getattr(func, "first", None)
funcs = reversed((first,) + func.funcs) if first else func.funcs
return tuple(normalize_function(f) for f in funcs)
elif isinstance(func, (partial, curry)):
args = tuple(normalize_token(i) for i in func.args)
if func.keywords:
kws = tuple(
(k, normalize_token(v)) for k, v in sorted(func.keywords.items())
)
else:
kws = None
return (normalize_function(func.func), args, kws)
else:
try:
result = pickle.dumps(func, protocol=0)
if b"__main__" not in result: # abort on dynamic functions
return result
except Exception:
pass
try:
import cloudpickle
return cloudpickle.dumps(func, protocol=0)
except Exception:
return str(func)
@normalize_token.register_lazy("pandas")
def register_pandas():
import pandas as pd
@normalize_token.register(pd.Index)
def normalize_index(ind):
values = ind.array
return [ind.name, normalize_token(values)]
@normalize_token.register(pd.MultiIndex)
def normalize_index(ind):
codes = ind.codes
return (
[ind.name]
+ [normalize_token(x) for x in ind.levels]
+ [normalize_token(x) for x in codes]
)
@normalize_token.register(pd.Categorical)
def normalize_categorical(cat):
return [normalize_token(cat.codes), normalize_token(cat.dtype)]
@normalize_token.register(pd.arrays.PeriodArray)
@normalize_token.register(pd.arrays.DatetimeArray)
@normalize_token.register(pd.arrays.TimedeltaArray)
def normalize_period_array(arr):
return [normalize_token(arr.asi8), normalize_token(arr.dtype)]
@normalize_token.register(pd.arrays.IntervalArray)
def normalize_interval_array(arr):
return [
normalize_token(arr.left),
normalize_token(arr.right),
normalize_token(arr.closed),
]
@normalize_token.register(pd.Series)
def normalize_series(s):
return [
s.name,
s.dtype,
normalize_token(s._data.blocks[0].values),
normalize_token(s.index),
]
@normalize_token.register(pd.DataFrame)
def normalize_dataframe(df):
data = [block.values for block in df._data.blocks]
data.extend([df.columns, df.index])
return list(map(normalize_token, data))
@normalize_token.register(pd.api.extensions.ExtensionArray)
def normalize_extension_array(arr):
import numpy as np
return normalize_token(np.asarray(arr))
# Dtypes
@normalize_token.register(pd.api.types.CategoricalDtype)
def normalize_categorical_dtype(dtype):
return [normalize_token(dtype.categories), normalize_token(dtype.ordered)]
@normalize_token.register(pd.api.extensions.ExtensionDtype)
def normalize_period_dtype(dtype):
return normalize_token(dtype.name)
@normalize_token.register_lazy("numpy")
def register_numpy():
import numpy as np
@normalize_token.register(np.ndarray)
def normalize_array(x):
if not x.shape:
return (x.item(), x.dtype)
if hasattr(x, "mode") and getattr(x, "filename", None):
if hasattr(x.base, "ctypes"):
offset = (
x.ctypes.get_as_parameter().value
- x.base.ctypes.get_as_parameter().value
)
else:
offset = 0 # root memmap's have mmap object as base
if hasattr(
x, "offset"
): # offset numpy used while opening, and not the offset to the beginning of the file
offset += getattr(x, "offset")
return (
x.filename,
os.path.getmtime(x.filename),
x.dtype,
x.shape,
x.strides,
offset,
)
if x.dtype.hasobject:
try:
try:
# string fast-path
data = hash_buffer_hex(
"-".join(x.flat).encode(
encoding="utf-8", errors="surrogatepass"
)
)
except UnicodeDecodeError:
# bytes fast-path
data = hash_buffer_hex(b"-".join(x.flat))
except (TypeError, UnicodeDecodeError):
try:
data = hash_buffer_hex(pickle.dumps(x, pickle.HIGHEST_PROTOCOL))
except Exception:
# pickling not supported, use UUID4-based fallback
data = uuid.uuid4().hex
else:
try:
data = hash_buffer_hex(x.ravel(order="K").view("i1"))
except (BufferError, AttributeError, ValueError):
data = hash_buffer_hex(x.copy().ravel(order="K").view("i1"))
return (data, x.dtype, x.shape, x.strides)
@normalize_token.register(np.matrix)
def normalize_matrix(x):
return type(x).__name__, normalize_array(x.view(type=np.ndarray))
normalize_token.register(np.dtype, repr)
normalize_token.register(np.generic, repr)
@normalize_token.register(np.ufunc)
def normalize_ufunc(x):
try:
name = x.__name__
if getattr(np, name) is x:
return "np." + name
except AttributeError:
return normalize_function(x)
@normalize_token.register_lazy("scipy")
def register_scipy():
import scipy.sparse as sp
def normalize_sparse_matrix(x, attrs):
return (
type(x).__name__,
normalize_seq((normalize_token(getattr(x, key)) for key in attrs)),
)
for cls, attrs in [
(sp.dia_matrix, ("data", "offsets", "shape")),
(sp.bsr_matrix, ("data", "indices", "indptr", "blocksize", "shape")),
(sp.coo_matrix, ("data", "row", "col", "shape")),
(sp.csr_matrix, ("data", "indices", "indptr", "shape")),
(sp.csc_matrix, ("data", "indices", "indptr", "shape")),
(sp.lil_matrix, ("data", "rows", "shape")),
]:
normalize_token.register(cls, partial(normalize_sparse_matrix, attrs=attrs))
@normalize_token.register(sp.dok_matrix)
def normalize_dok_matrix(x):
return type(x).__name__, normalize_token(sorted(x.items()))
def _colorize(t):
"""Convert (r, g, b) triple to "#RRGGBB" string
For use with ``visualize(color=...)``
Examples
--------
>>> _colorize((255, 255, 255))
'#FFFFFF'
>>> _colorize((0, 32, 128))
'#002080'
"""
t = t[:3]
i = sum(v * 256 ** (len(t) - i - 1) for i, v in enumerate(t))
h = hex(int(i))[2:].upper()
h = "0" * (6 - len(h)) + h
return "#" + h
named_schedulers = {
"sync": local.get_sync,
"synchronous": local.get_sync,
"single-threaded": local.get_sync,
"threads": threaded.get,
"threading": threaded.get,
}
try:
from dask import multiprocessing as dask_multiprocessing
except ImportError:
pass
else:
named_schedulers.update(
{
"processes": dask_multiprocessing.get,
"multiprocessing": dask_multiprocessing.get,
}
)
get_err_msg = """
The get= keyword has been removed.
Please use the scheduler= keyword instead with the name of
the desired scheduler like 'threads' or 'processes'
x.compute(scheduler='single-threaded')
x.compute(scheduler='threads')
x.compute(scheduler='processes')
or with a function that takes the graph and keys
x.compute(scheduler=my_scheduler_function)
or with a Dask client
x.compute(scheduler=client)
""".strip()
def get_scheduler(get=None, scheduler=None, collections=None, cls=None):
"""Get scheduler function
There are various ways to specify the scheduler to use:
1. Passing in scheduler= parameters
2. Passing these into global configuration
3. Using defaults of a dask collection
This function centralizes the logic to determine the right scheduler to use
from those many options
"""
if get:
raise TypeError(get_err_msg)
if scheduler is not None:
if callable(scheduler):
return scheduler
elif "Client" in type(scheduler).__name__ and hasattr(scheduler, "get"):
return scheduler.get
elif scheduler.lower() in named_schedulers:
return named_schedulers[scheduler.lower()]
elif scheduler.lower() in ("dask.distributed", "distributed"):
from distributed.worker import get_client
return get_client().get
else:
raise ValueError(
"Expected one of [distributed, %s]"
% ", ".join(sorted(named_schedulers))
)
# else: # try to connect to remote scheduler with this name
# return get_client(scheduler).get
if config.get("scheduler", None):
return get_scheduler(scheduler=config.get("scheduler", None))
if config.get("get", None):
raise ValueError(get_err_msg)
if getattr(thread_state, "key", False):
from distributed.worker import get_worker
return get_worker().client.get
if cls is not None:
return cls.__dask_scheduler__
if collections:
collections = [c for c in collections if c is not None]
if collections:
get = collections[0].__dask_scheduler__
if not all(c.__dask_scheduler__ == get for c in collections):
raise ValueError(
"Compute called on multiple collections with "
"differing default schedulers. Please specify a "
"scheduler=` parameter explicitly in compute or "
"globally with `dask.config.set`."
)
return get
return None
def wait(x, timeout=None, return_when="ALL_COMPLETED"):
"""Wait until computation has finished
This is a compatibility alias for ``dask.distributed.wait``.
If it is applied onto Dask collections without Dask Futures or if Dask
distributed is not installed then it is a no-op
"""
try:
from distributed import wait
return wait(x, timeout=timeout, return_when=return_when)
except (ImportError, ValueError):
return x
def get_collection_names(collection) -> Set[str]:
"""Infer the collection names from the dask keys, under the assumption that all keys
are either tuples with matching first element, and that element is a string, or
there is exactly one key and it is a string.
Examples
--------
>>> a.__dask_keys__() # doctest: +SKIP
["foo", "bar"] # doctest: +SKIP
>>> get_collection_names(a) # doctest: +SKIP
{"foo", "bar"} # doctest: +SKIP
>>> b.__dask_keys__() # doctest: +SKIP
[[("foo-123", 0, 0), ("foo-123", 0, 1)], [("foo-123", 1, 0), ("foo-123", 1, 1)]] # doctest: +SKIP
>>> get_collection_names(b) # doctest: +SKIP
{"foo-123"} # doctest: +SKIP
"""
if not is_dask_collection(collection):
raise TypeError(f"Expected Dask collection; got {type(collection)}")
return {get_name_from_key(k) for k in flatten(collection.__dask_keys__())}
def get_name_from_key(key) -> str:
"""Given a dask collection's key, extract the collection name.
Parameters
----------
key: string or tuple
Dask collection's key, which must be either a single string or a tuple whose
first element is a string (commonly referred to as a collection's 'name'),
Examples
--------
>>> get_name_from_key("foo")
'foo'
>>> get_name_from_key(("foo-123", 1, 2))
'foo-123'
"""
if isinstance(key, tuple) and key and isinstance(key[0], str):
return key[0]
if isinstance(key, str):
return key
raise TypeError(f"Expected str or tuple[str, Hashable, ...]; got {key}")
def replace_name_in_key(key, rename: Mapping[str, str]):
"""Given a dask collection's key, replace the collection name with a new one.
Parameters
----------
key: string or tuple
Dask collection's key, which must be either a single string or a tuple whose
first element is a string (commonly referred to as a collection's 'name'),
rename:
Mapping of zero or more names from : to. Extraneous names will be ignored.
Names not found in this mapping won't be replaced.
Examples
--------
>>> replace_name_in_key("foo", {})
'foo'
>>> replace_name_in_key("foo", {"foo": "bar"})
'bar'
>>> replace_name_in_key(("foo-123", 1, 2), {"foo-123": "bar-456"})
('bar-456', 1, 2)
"""
if isinstance(key, tuple) and key and isinstance(key[0], str):
return (rename.get(key[0], key[0]),) + key[1:]
if isinstance(key, str):
return rename.get(key, key)
raise TypeError(f"Expected str or tuple[str, Hashable, ...]; got {key}")
def clone_key(key, seed):
"""Clone a key from a Dask collection, producing a new key with the same prefix and
indices and a token which a deterministic function of the previous token and seed.
Examples
--------
>>> clone_key("inc-cbb1eca3bafafbb3e8b2419c4eebb387", 123) # doctest: +SKIP
'inc-1d291de52f5045f8a969743daea271fd' # doctest: +SKIP
>>> clone_key(("sum-cbb1eca3bafafbb3e8b2419c4eebb387", 4, 3), 123) # doctest: +SKIP
('sum-f0962cc58ef4415689a86cc1d4cc1723', 4, 3) # doctest: +SKIP
"""
if isinstance(key, tuple) and key and isinstance(key[0], str):
return (clone_key(key[0], seed),) + key[1:]
if isinstance(key, str):
prefix = key_split(key)
token = key[len(prefix) + 1 :]
if token:
return prefix + "-" + tokenize(token, seed)
else:
return tokenize(key, seed)
raise TypeError(f"Expected str or tuple[str, Hashable, ...]; got {key}")