graph.py 12.4 KB
Newer Older
matthmey's avatar
matthmey committed
1
import dask
2
from dask.core import get_dependencies, flatten
matthmey's avatar
matthmey committed
3
4
5
import numpy as np
import copy

6
7
8
9
10

class Node(object):
    def __init__(self):
        pass

matthmey's avatar
matthmey committed
11
    def configure(self, requests):
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
        """ Before a task graph is executed each node is configured.
            The request is propagated from the end to the beginning 
            of the DAG and each nodes "configure" routine is called.
            The request can be updated to reflect additional requirements,
            The return value gets passed to predecessors.

            Essentially the following question must be answered:
            What do I need to fulfil the request of my successor?

            Here, you must not configure the internal parameters of the
            Node otherwise it would not be thread-safe. You can however
            introduce a new key 'requires_request' in the request being 
            returned. This request will then be passed as an argument
            to the __call__ function.

            Best practice is to configure the Node on initialization with
            runtime independent configurations and define all runtime
            dependant configurations here.
        
        Arguments:
            requests {List} -- List of requests (i.e. dictionaries).

        
        Returns:
            dict -- The (updated) request. If updated modifications
                    must be made on a copy of the input. The return value
                    must be a dictionary. 
                    If multiple requests are input to this function they 
                    must be merged.
                    If nothing needs to be requested an empty dictionary
                    can be return. This removes all dependencies of this
                    node from the task graph.

        """
matthmey's avatar
matthmey committed
46
47
        if not isinstance(requests, list):
            raise RuntimeError("Please provide a **list** of request")
48
        if len(requests) > 1:
matthmey's avatar
matthmey committed
49
50
51
52
53
            raise RuntimeError(
                "Default configuration function cannot handle "
                "multiple requests. Please provide a custom "
                "configuration implementation"
            )
54
55
56
        return requests

    @dask.delayed
matthmey's avatar
matthmey committed
57
    def __call__(self, x, request=None):
58
59
60
61
62
63
64
        raise NotImplementedError()

    def get_config(self):
        """ returns a dictionary of configurations to recreate the state
        """
        raise NotImplementedError()

matthmey's avatar
matthmey committed
65
66
67

class StuettNode(Node):  # TODO: define where this class should be (maybe not here)
    def configure(self, requests):
68
69
70
71
72
73
74
75
76
77
78
79
        """ Default configure for stuett nodes
            Expects two keys per request (*start_time* and *tend*)
            If multiple requests are passed, they will be merged
            start_time = minimum of all requests' start_time
            end_time = maximum of all requests' end_time
        
        Arguments:
            request {list} -- List of requests

        Returns:
            dict -- Original request or merged requests 
        """
matthmey's avatar
matthmey committed
80
81
        if not isinstance(requests, list):
            raise RuntimeError("Please provide a list of request")
82
83
84
85

        # For time requests we just use the union of both time segments
        new_request = requests[0].copy()

matthmey's avatar
matthmey committed
86
        key_func = {"start_time": np.minimum, "end_time": np.maximum}
87
        for r in requests[1:]:
matthmey's avatar
matthmey committed
88
            for key in ["start_time", "end_time"]:
89
90
                if key in r:
                    if key in new_request:
matthmey's avatar
matthmey committed
91
                        new_request[key] = key_func[key](new_request[key], r[key])
92
                    else:
matthmey's avatar
matthmey committed
93
94
                        new_request[key] = r[key]

95
96
        return new_request

matthmey's avatar
matthmey committed
97
98

def configuration(delayed, request, keys=None, default_merge=None):
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
    """ Configures each node of the graph by propagating the request from outputs
        to inputs.
        Each node checks if it can fulfil the request and what it needs to fulfil the request.
        If a node requires additional configurations to fulfil the request it can set the
        'requires_request' flag in the returned request and this function will add the 
        return request as a a new input to the node's __call__(). See also Node.configure()
    
    Arguments:
        delayed {dask.delayed or list}  -- Delayed object or list of delayed objects
        request {dict or list}           -- request (dict), list of requests
        default_merge {callable}        -- request merge function 
    
    Keyword Arguments:
        keys {[type]} -- [description] (default: {None})
    
    Raises:
        RuntimeError: [description]
        RuntimeError: [description]
    
    Returns:
        dask.delayed or list -- Config-optimized delayed object or list of delayed objects
    """

matthmey's avatar
matthmey committed
122
    if not isinstance(delayed, list):
123
124
125
126
        collections = [delayed]

    # dsk = dask.base.collections_to_dsk(collections)
    dsk, dsk_keys = dask.base._extract_graph_and_keys(collections)
matthmey's avatar
matthmey committed
127
    dependencies, dependants = dask.core.get_deps(dsk)
128
129
130
131
132
133
134
135
136
137
138

    if keys is None:
        keys = dsk_keys

    if not isinstance(keys, (list, set)):
        keys = [keys]
    out_keys = []
    seen = set()

    work = list(set(flatten(keys)))

matthmey's avatar
matthmey committed
139
    if isinstance(request, list):
140
        if len(request) != len(work):
matthmey's avatar
matthmey committed
141
142
143
144
145
146
147
            raise RuntimeError(
                "When passing multiple request items "
                "The number of request items must be same "
                "as the number of keys"
            )

        requests = {work[i]: [request[i]] for i in range(len(request))}
148
    else:
matthmey's avatar
matthmey committed
149
        requests = {k: [request] for k in work}
150

matthmey's avatar
matthmey committed
151
    remove = {k: False for k in work}
152
153
154
155
156
157
158
159
160
161
162
163
164
    input_requests = {}
    while work:
        new_work = []
        out_keys += work
        deps = []
        for k in work:
            # if k not in requests:
            #     # there wasn't any request stored use initial config
            #     requests[k] = [config]

            # check if we have collected all dependencies so far
            # we will come back to this node another time
            # TODO: make a better check for the case when dependants[k] is a set, also: why is it a set in the first place..?
matthmey's avatar
matthmey committed
165
166
167
168
169
            if (
                k in dependants
                and len(dependants[k]) != len(requests[k])
                and not isinstance(dependants[k], set)
            ):
170
171
172
173
174
175
176
                # print(f'Waiting at {k}', dependants[k], requests[k])
                continue

            # print(f"configuring {k}",requests[k])
            # set configuration for this node k

            # If we create a delayed object from a class, `self` will be dsk[k][1]
matthmey's avatar
matthmey committed
177
178
179
            if isinstance(dsk[k], tuple) and isinstance(
                dsk[k][1], Node
            ):  # Check if we get a node of type Node class
180
181
                # current_requests = [r for r in requests[k] if r]                    # get all requests belonging to this node
                current_requests = requests[k]
matthmey's avatar
matthmey committed
182
183
184
185
186
187
                new_request = dsk[k][1].configure(
                    current_requests
                )  # Call the class configuration function
                if not isinstance(
                    new_request, list
                ):  # prepare the request return value
188
                    new_request = [new_request]
matthmey's avatar
matthmey committed
189
190
            else:  # We didn't get a Node class so there is no
                # custom configuration function: pass through
191
192
193
194
                if len(requests[k]) > 1:
                    if callable(default_merge):
                        new_request = default_merge(requests[k])
                    else:
matthmey's avatar
matthmey committed
195
196
197
198
199
                        raise RuntimeError(
                            "No valid default merger supplied. Cannot merge requests. "
                            "Either convert your function to a class Node or provide "
                            "a default merger"
                        )
200
201
                else:
                    new_request = requests[k]
matthmey's avatar
matthmey committed
202
203
204
205
206
207
208
209
210

            if (
                "requires_request" in new_request[0]
                and new_request[0]["requires_request"] == True
            ):
                del new_request[0]["requires_request"]
                input_requests[k] = copy.deepcopy(
                    new_request[0]
                )  # TODO: check if we need a deepcopy here!
211
212
213
214
215
216
217
218
219

            # update dependencies
            current_deps = get_dependencies(dsk, k, as_list=True)
            for i, d in enumerate(current_deps):
                if d in requests:
                    requests[d] += new_request
                    remove[d] = remove[d] and (not new_request[0])
                else:
                    requests[d] = new_request
matthmey's avatar
matthmey committed
220
221
222
                    remove[d] = not new_request[
                        0
                    ]  # if we received an empty dictionary flag deps for removal
223
224

                # only configure each node once in a round!
matthmey's avatar
matthmey committed
225
226
227
228
229
                if d not in new_work and d not in work:  # TODO: verify this
                    new_work.append(
                        d
                    )  # TODO: Do we need to configure dependency if we'll remove it?

230
231
232
233
234
235
236
237
238
239
        work = new_work

    # Assembling the configured new graph
    out = {k: dsk[k] for k in out_keys if not remove[k]}
    # After we have aquired all requests we can input the required_requests as a input node to the requiring node
    for k in input_requests:
        out[k] += (input_requests[k],)

    # convert to delayed object
    from dask.delayed import Delayed
matthmey's avatar
matthmey committed
240

241
    in_keys = list(flatten(keys))
matthmey's avatar
matthmey committed
242
    # print(in_keys)
243
    if len(in_keys) > 1:
matthmey's avatar
matthmey committed
244
        collection = [Delayed(key=key, dsk=out) for key in in_keys]
245
    else:
matthmey's avatar
matthmey committed
246
247
        collection = Delayed(key=in_keys[0], dsk=out)
        if isinstance(collection, list):
248
249
250
251
252
253
            collection = [collection]

    return collection


class Freezer(Node):
matthmey's avatar
matthmey committed
254
    def __init__(self, caching=True):
255
256
257
258
259
260
261
262
263
264
265
266
267
268
        self.caching = caching

    @dask.delayed
    def __call__(self, x):
        """If caching is enabled load a cached result or stores the input data and returns it
        
        Arguments:
            x {xarray or dict} -- Either the xarray data to be passed through (and cached)
                                  or request dictionary containing information about the data
                                  to be loaded
        
        Returns:
            xarray -- Data loaded from cache or input data passed through
        """
matthmey's avatar
matthmey committed
269
270

        if isinstance(x, dict):
271
272
273
274
            if self.is_cached(x) and self.caching:
                # TODO: load from cache and return it
                pass
            elif not self.caching:
matthmey's avatar
matthmey committed
275
                raise RuntimeError(f"If caching is disabled cannot perform request {x}")
276
            else:
matthmey's avatar
matthmey committed
277
278
279
280
                raise RuntimeError(
                    f"Result is not cached but cached result is requested with {x}"
                )

281
        if self.caching:
matthmey's avatar
matthmey committed
282
283
            # TODO: store the input data
            pass
284
285
286

        return x

matthmey's avatar
matthmey committed
287
    def configure(self, requests):
288
289
290
291
292
        if self.caching:
            return [{}]
        return config_conflict(requests)


matthmey's avatar
matthmey committed
293
def optimize_freeze(dsk, keys, request_key="request"):
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
    """ Return new dask with tasks removed which are unnecessary because a later stage 
    reads from cache
    ``keys`` may be a single key or list of keys.
    Examples
    --------

    Returns
    -------
    dsk: culled dask graph
    dependencies: Dict mapping {key: [deps]}.  Useful side effect to accelerate
        other optimizations, notably fuse.
    """
    if not isinstance(keys, (list, set)):
        keys = [keys]
    out_keys = []
    seen = set()
    dependencies = dict()

matthmey's avatar
matthmey committed
312
313
314
315
316
    if request_key not in dsk:
        raise RuntimeError(
            f"Please provide a task graph which includes '{request_key}'"
        )

317
318
    request = dsk[request_key]

matthmey's avatar
matthmey committed
319
320
321
    def is_cached(task, request):
        if isinstance(task, tuple):
            if isinstance(task[0], Freezer):
322
                return task[0].is_cached(request)
matthmey's avatar
matthmey committed
323
        return False
324
325
326
327
328
329
330
331

    work = list(set(flatten(keys)))
    cached_keys = []
    while work:
        new_work = []
        out_keys += work
        deps = []
        for k in work:
matthmey's avatar
matthmey committed
332
            if is_cached(dsk[k], request):
333
334
335
                cached_keys.append(k)
            else:
                deps.append((k, get_dependencies(dsk, k, as_list=True)))
matthmey's avatar
matthmey committed
336

337
338
339
340
341
342
343
344
345
346
347
        dependencies.update(deps)
        for _, deplist in deps:
            for d in deplist:
                if d not in seen:
                    seen.add(d)
                    new_work.append(d)
        work = new_work

    out = {k: dsk[k] for k in out_keys}

    # finally we need to replace the input of the caching nodes with the request
matthmey's avatar
matthmey committed
348
    cached = {k: (out[k][0], request_key) for k in cached_keys}
349
350
351
    out.update(cached)

    return out, dependencies