graph.py 12.3 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
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
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
import dask 
from dask.core import get_dependencies, flatten
import numpy as np 
import copy 

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

    def configure(self,requests):
        """ 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.

        """
        if not isinstance(requests,list):
            raise RuntimeError('Please provide a **list** of request')
        if len(requests) > 1:
            raise RuntimeError('Default configuration function cannot handle '
                               'multiple requests. Please provide a custom '
                               'configuration implementation')
        return requests

    @dask.delayed
    def __call__(self,x,request=None):
        raise NotImplementedError()

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

class StuettNode(Node):             # TODO: define where this class should be (maybe not here)
    def configure(self,requests):
        """ 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 
        """
        if not isinstance(requests,list):
            raise RuntimeError('Please provide a list of request')

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

        key_func = {'start_time':np.minimum, 'end_time':np.maximum}
        for r in requests[1:]:
            for key in ['start_time', 'end_time']: 
                if key in r:
                    if key in new_request:
                        new_request[key] = key_func[key](new_request[key],r[key])
                    else:
                        new_request[key] = r[key]                   
        
        return new_request

def configuration(delayed,request,keys=None,default_merge=None):
    """ 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
    """


    if not isinstance(delayed,list):   
        collections = [delayed]

    # dsk = dask.base.collections_to_dsk(collections)
    dsk, dsk_keys = dask.base._extract_graph_and_keys(collections)
    dependencies,dependants = dask.core.get_deps(dsk)

    if keys is None:
        keys = dsk_keys

    print('dsk',dsk.layers)
    # print('keys',keys)

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

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

    if isinstance(request,list):
        if len(request) != len(work):
            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)) }
    else:
        requests = {k: [request] for k in work }

    remove = {k:False for k in work}
    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..?
            if k in dependants and len(dependants[k]) != len(requests[k]) and not isinstance(dependants[k],set): 
                # 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]
            if isinstance(dsk[k],tuple) and isinstance(dsk[k][1],Node):             # Check if we get a node of type Node class
                # current_requests = [r for r in requests[k] if r]                    # get all requests belonging to this node
                current_requests = requests[k]
                new_request = dsk[k][1].configure(current_requests)                 # Call the class configuration function
                if not isinstance(new_request,list):                                # prepare the request return value
                    new_request = [new_request]
            else:                                                                   # We didn't get a Node class so there is no 
                                                                                    # custom configuration function: pass through
                if len(requests[k]) > 1:
                    if callable(default_merge):
                        new_request = default_merge(requests[k])
                    else:
                        raise RuntimeError("No valid default merger supplied. Cannot merge requests. " 
                                           "Either convert your function to a class Node or provide " 
                                           "a default merger")
                else:
                    new_request = requests[k]
               
            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!

            # 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
                    remove[d] = (not new_request[0]) # if we received an empty dictionary flag deps for removal

                # only configure each node once in a round!
                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?
            
        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
    in_keys = list(flatten(keys))
    print(in_keys)
    if len(in_keys) > 1:
        collection = [Delayed(key=key,dsk=out) for key in in_keys]
    else:
        collection = Delayed(key=in_keys[0],dsk=out)
        if isinstance(collection,list):
            collection = [collection]


    return collection



class Freezer(Node):
    def __init__(self,caching=True):
        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
        """
        
        if isinstance(x,dict):
            if self.is_cached(x) and self.caching:
                # TODO: load from cache and return it
                pass
            elif not self.caching:
                raise RuntimeError(f'If caching is disabled cannot perform request {x}')
            else:
                raise RuntimeError(f'Result is not cached but cached result is requested with {x}')
        
        if self.caching:
             # TODO: store the input data
             pass

        return x

    def configure(self,requests):
        if self.caching:
            return [{}]
        return config_conflict(requests)


def optimize_freeze(dsk, keys, request_key='request'):
    """ 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()

    if (request_key not in dsk):
            raise RuntimeError(f"Please provide a task graph which includes '{request_key}'")
    
    request = dsk[request_key]

    def is_cached(task,request):
        if isinstance(task,tuple):
            if isinstance(task[0],Freezer):
                return task[0].is_cached(request)
        return False 

    work = list(set(flatten(keys)))
    cached_keys = []
    while work:
        new_work = []
        out_keys += work
        deps = []
        for k in work:
            if is_cached(dsk[k],request):
                cached_keys.append(k)
            else:
                deps.append((k, get_dependencies(dsk, k, as_list=True)))
                
        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
    cached = {k: (out[k][0],request_key) for k in cached_keys}
    out.update(cached)

    return out, dependencies