Coverage for python/lsst/pipe/base/quantum_graph_builder.py: 25%
372 statements
« prev ^ index » next coverage.py v7.4.2, created at 2024-02-22 11:04 +0000
« prev ^ index » next coverage.py v7.4.2, created at 2024-02-22 11:04 +0000
1# This file is part of pipe_base.
2#
3# Developed for the LSST Data Management System.
4# This product includes software developed by the LSST Project
5# (http://www.lsst.org).
6# See the COPYRIGHT file at the top-level directory of this distribution
7# for details of code ownership.
8#
9# This software is dual licensed under the GNU General Public License and also
10# under a 3-clause BSD license. Recipients may choose which of these licenses
11# to use; please see the files gpl-3.0.txt and/or bsd_license.txt,
12# respectively. If you choose the GPL option then the following text applies
13# (but note that there is still no warranty even if you opt for BSD instead):
14#
15# This program is free software: you can redistribute it and/or modify
16# it under the terms of the GNU General Public License as published by
17# the Free Software Foundation, either version 3 of the License, or
18# (at your option) any later version.
19#
20# This program is distributed in the hope that it will be useful,
21# but WITHOUT ANY WARRANTY; without even the implied warranty of
22# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
23# GNU General Public License for more details.
24#
25# You should have received a copy of the GNU General Public License
26# along with this program. If not, see <http://www.gnu.org/licenses/>.
28"""The base class for the QuantumGraph-generation algorithm and various
29helper classes.
30"""
32from __future__ import annotations
34__all__ = (
35 "QuantumGraphBuilder",
36 "ExistingDatasets",
37 "QuantumGraphBuilderError",
38 "OutputExistsError",
39 "PrerequisiteMissingError",
40)
42import dataclasses
43from abc import ABC, abstractmethod
44from collections.abc import Iterable, Mapping, Sequence
45from typing import TYPE_CHECKING, Any, final
47from lsst.daf.butler import (
48 Butler,
49 CollectionType,
50 DataCoordinate,
51 DatasetRef,
52 DatasetType,
53 DimensionUniverse,
54 NamedKeyDict,
55 NamedKeyMapping,
56 Quantum,
57)
58from lsst.daf.butler.registry import MissingCollectionError, MissingDatasetTypeError
59from lsst.utils.logging import LsstLogAdapter, getLogger
60from lsst.utils.timer import timeMethod
62from . import automatic_connection_constants as acc
63from ._status import NoWorkFound
64from ._task_metadata import TaskMetadata
65from .connections import AdjustQuantumHelper
66from .graph import QuantumGraph
67from .pipeline_graph import PipelineGraph, TaskNode
68from .prerequisite_helpers import PrerequisiteInfo, SkyPixBoundsBuilder, TimespanBuilder
69from .quantum_graph_skeleton import (
70 DatasetKey,
71 PrerequisiteDatasetKey,
72 QuantumGraphSkeleton,
73 QuantumKey,
74 TaskInitKey,
75)
77if TYPE_CHECKING:
78 from .pipeline import TaskDef
81class QuantumGraphBuilderError(Exception):
82 """Base class for exceptions generated by QuantumGraphBuilder."""
84 pass
87class GraphBuilderError(QuantumGraphBuilderError):
88 """Backwards-compatibility near-alias for QuantumGraphBuilderError."""
90 pass
93# Inherit from backwards-compatibility alias for backwards-compatibility.
94class OutputExistsError(GraphBuilderError):
95 """Exception generated when output datasets already exist."""
97 pass
100# Inherit from backwards-compatibility alias for backwards-compatibility.
101class PrerequisiteMissingError(GraphBuilderError):
102 """Exception generated when a prerequisite dataset does not exist."""
104 pass
107class InitInputMissingError(QuantumGraphBuilderError):
108 """Exception generated when an init-input dataset does not exist."""
110 pass
113class QuantumGraphBuilder(ABC):
114 """An abstract base class for building `QuantumGraph` objects from a
115 pipeline.
117 Parameters
118 ----------
119 pipeline_graph : `.pipeline_graph.PipelineGraph`
120 Pipeline to build a `QuantumGraph` from, as a graph. Will be resolved
121 in-place with the given butler (any existing resolution is ignored).
122 butler : `lsst.daf.butler.Butler`
123 Client for the data repository. Should be read-only.
124 input_collections : `~collections.abc.Sequence` [ `str` ], optional
125 Collections to search for overall-input datasets. If not provided,
126 ``butler.collections`` is used (and must not be empty).
127 output_run : `str`, optional
128 Output `~lsst.daf.butler.CollectionType.RUN` collection. If not
129 provided, ``butler.run`` is used (and must not be `None`).
130 skip_existing_in : `~collections.abc.Sequence` [ `str` ], optional
131 Collections to search for outputs that already exist for the purpose of
132 skipping quanta that have already been run.
133 clobber : `bool`, optional
134 Whether to raise if predicted outputs already exist in ``output_run``
135 (not including those quanta that would be skipped because they've
136 already been run). This never actually clobbers outputs; it just
137 informs the graph generation algorithm whether execution will run with
138 clobbering enabled. This is ignored if ``output_run`` does not exist.
140 Notes
141 -----
142 Constructing a `QuantumGraphBuilder` will run queries for existing datasets
143 with empty data IDs (including but not limited to init inputs and outputs),
144 in addition to resolving the given pipeline graph and testing for existence
145 of the ``output`` run collection.
147 The `build` method splits the pipeline graph into independent subgraphs,
148 then calls the abstract method `process_subgraph` on each, to allow
149 concrete implementations to populate the rough graph structure (the
150 `~quantum_graph_skeleton.QuantumGraphSkeleton` class) and search for
151 existing datasets (further populating the builder's `existing_datasets`
152 struct). The `build` method then:
154 - assembles `lsst.daf.butler.Quantum` instances from all data IDs in the
155 skeleton;
156 - looks for existing outputs found in ``skip_existing_in`` to see if any
157 quanta should be skipped;
158 - calls `PipelineTaskConnections.adjustQuantum` on all quanta, adjusting
159 downstream quanta appropriately when preliminary predicted outputs are
160 rejected (pruning nodes that will not have the inputs they need to run);
161 - attaches datastore records and registry dataset types to the graph.
163 In addition to implementing `process_subgraph`, derived classes are
164 generally expected to add new construction keyword-only arguments to
165 control the data IDs of the quantum graph, while forwarding all of the
166 arguments defined in the base class to `super`.
167 """
169 def __init__(
170 self,
171 pipeline_graph: PipelineGraph,
172 butler: Butler,
173 *,
174 input_collections: Sequence[str] | None = None,
175 output_run: str | None = None,
176 skip_existing_in: Sequence[str] = (),
177 clobber: bool = False,
178 ):
179 self.log = getLogger(__name__)
180 self.metadata = TaskMetadata()
181 self._pipeline_graph = pipeline_graph
182 self.butler = butler
183 if input_collections is None:
184 input_collections = butler.collections
185 if not input_collections:
186 raise ValueError("No input collections provided.")
187 self.input_collections = input_collections
188 if output_run is None:
189 output_run = butler.run
190 if not output_run:
191 raise ValueError("No output RUN collection provided.")
192 self.output_run = output_run
193 self.skip_existing_in = skip_existing_in
194 self.empty_data_id = DataCoordinate.make_empty(butler.dimensions)
195 self.clobber = clobber
196 # See whether the output run already exists.
197 self.output_run_exists = False
198 try:
199 if self.butler.registry.getCollectionType(self.output_run) is not CollectionType.RUN:
200 raise RuntimeError(f"{self.output_run!r} is not a RUN collection.")
201 self.output_run_exists = True
202 except MissingCollectionError:
203 # If the run doesn't exist we never need to clobber. This is not
204 # an error so you can run with clobber=True the first time you
205 # attempt some processing as well as all subsequent times, instead
206 # of forcing the user to make the first attempt different.
207 self.clobber = False
208 # We need to know whether the skip_existing_in collection sequence
209 # starts with the output run collection, as an optimization to avoid
210 # queries later.
211 try:
212 skip_existing_in_flat = self.butler.registry.queryCollections(
213 self.skip_existing_in, flattenChains=True
214 )
215 except MissingCollectionError:
216 skip_existing_in_flat = []
217 if not skip_existing_in_flat:
218 self.skip_existing_in = []
219 if self.skip_existing_in and self.output_run_exists:
220 self.skip_existing_starts_with_output_run = self.output_run == skip_existing_in_flat[0]
221 else:
222 self.skip_existing_starts_with_output_run = False
223 self.existing_datasets = ExistingDatasets()
224 try:
225 packages_storage_class = butler.get_dataset_type(acc.PACKAGES_INIT_OUTPUT_NAME).storageClass_name
226 except MissingDatasetTypeError:
227 packages_storage_class = acc.PACKAGES_INIT_OUTPUT_STORAGE_CLASS
228 self._global_init_output_types = {
229 acc.PACKAGES_INIT_OUTPUT_NAME: DatasetType(
230 acc.PACKAGES_INIT_OUTPUT_NAME,
231 self.universe.empty,
232 packages_storage_class,
233 )
234 }
235 with self.butler.registry.caching_context():
236 self._pipeline_graph.resolve(self.butler.registry)
237 self._find_empty_dimension_datasets()
238 self.prerequisite_info = {
239 task_node.label: PrerequisiteInfo(task_node, self._pipeline_graph)
240 for task_node in pipeline_graph.tasks.values()
241 }
243 log: LsstLogAdapter
244 """Logger to use for all quantum-graph generation messages.
246 General and per-task status messages should be logged at `~logging.INFO`
247 level or higher, per-dataset-type status messages should be logged at
248 `~lsst.utils.logging.VERBOSE` or higher, and per-data-ID status messages
249 should be logged at `logging.DEBUG` or higher.
250 """
252 metadata: TaskMetadata
253 """Metadata to store in the QuantumGraph.
255 The `TaskMetadata` class is used here primarily in order to enable
256 resource-usage collection with the `lsst.utils.timer.timeMethod` decorator.
257 """
259 butler: Butler
260 """Client for the data repository.
262 Should be read-only.
263 """
265 input_collections: Sequence[str]
266 """Collections to search for overall-input datasets.
267 """
269 output_run: str
270 """Output `~lsst.daf.butler.CollectionType.RUN` collection.
271 """
273 skip_existing_in: Sequence[str]
274 """Collections to search for outputs that already exist for the purpose
275 of skipping quanta that have already been run.
276 """
278 clobber: bool
279 """Whether to raise if predicted outputs already exist in ``output_run``
281 This never actually clobbers outputs; it just informs the graph generation
282 algorithm whether execution will run with clobbering enabled. This is
283 always `False` if `output_run_exists` is `False`.
284 """
286 empty_data_id: DataCoordinate
287 """An empty data ID in the data repository's dimension universe.
288 """
290 output_run_exists: bool
291 """Whether the output run exists in the data repository already.
292 """
294 skip_existing_starts_with_output_run: bool
295 """Whether the `skip_existing_in` sequence begins with `output_run`.
297 If this is true, any dataset found in `output_run` can be used to
298 short-circuit queries in `skip_existing_in`.
299 """
301 existing_datasets: ExistingDatasets
302 """Struct holding datasets that have already been found in the data
303 repository.
305 This is updated in-place as the `QuantumGraph` generation algorithm
306 proceeds.
307 """
309 prerequisite_info: Mapping[str, PrerequisiteInfo]
310 """Helper objects for finding prerequisite inputs, organized by task label.
312 Subclasses that find prerequisites should remove the
313 covered `~prerequisite_helpers.PrerequisiteFinder` objects from this
314 attribute.
315 """
317 @property
318 def universe(self) -> DimensionUniverse:
319 """Definitions of all data dimensions."""
320 return self.butler.dimensions
322 @final
323 @timeMethod
324 def build(self, metadata: Mapping[str, Any] | None = None) -> QuantumGraph:
325 """Build the quantum graph.
327 Parameters
328 ----------
329 metadata : `~collections.abc.Mapping`, optional
330 Flexible metadata to add to the quantum graph.
332 Returns
333 -------
334 quantum_graph : `QuantumGraph`
335 DAG describing processing to be performed.
337 Notes
338 -----
339 External code is expected to construct a `QuantumGraphBuilder` and then
340 call this method exactly once. See class documentation for details on
341 what it does.
342 """
343 with self.butler.registry.caching_context():
344 full_skeleton = QuantumGraphSkeleton(self._pipeline_graph.tasks)
345 subgraphs = list(self._pipeline_graph.split_independent())
346 for i, subgraph in enumerate(subgraphs):
347 self.log.info(
348 "Processing pipeline subgraph %d of %d with %d task(s).",
349 i + 1,
350 len(subgraphs),
351 len(subgraph.tasks),
352 )
353 self.log.verbose("Subgraph tasks: [%s]", ", ".join(label for label in subgraph.tasks))
354 subgraph_skeleton = self.process_subgraph(subgraph)
355 full_skeleton.update(subgraph_skeleton)
356 # Loop over tasks. The pipeline graph must be topologically
357 # sorted, so a quantum is only processed after any quantum that
358 # provides its inputs has been processed.
359 for task_node in self._pipeline_graph.tasks.values():
360 self._resolve_task_quanta(task_node, full_skeleton)
361 # Add global init-outputs to the skeleton.
362 for dataset_type in self._global_init_output_types.values():
363 dataset_key = full_skeleton.add_dataset_node(
364 dataset_type.name, self.empty_data_id, is_global_init_output=True
365 )
366 ref = self.existing_datasets.outputs_in_the_way.get(dataset_key)
367 if ref is None:
368 ref = DatasetRef(dataset_type, self.empty_data_id, run=self.output_run)
369 full_skeleton[dataset_key]["ref"] = ref
370 # Remove dataset nodes with no edges that are not global init
371 # outputs, which are generally overall-inputs whose original quanta
372 # end up skipped or with no work to do (we can't remove these along
373 # with the quanta because no quantum knows if its the only
374 # consumer).
375 full_skeleton.remove_orphan_datasets()
376 self._attach_datastore_records(full_skeleton)
377 # TODO initialize most metadata here instead of in ctrl_mpexec.
378 if metadata is None:
379 metadata = {}
380 return self._construct_quantum_graph(full_skeleton, metadata)
382 @abstractmethod
383 def process_subgraph(self, subgraph: PipelineGraph) -> QuantumGraphSkeleton:
384 """Build the rough structure for an independent subset of the
385 `QuantumGraph` and query for relevant existing datasets.
387 Parameters
388 ----------
389 subgraph : `.pipeline_graph.PipelineGraph`
390 Subset of the pipeline graph that should be processed by this call.
391 This is always resolved and topologically sorted. It should not be
392 modified.
394 Returns
395 -------
396 skeleton : `quantum_graph_skeleton.QuantumGraphSkeleton`
397 Class representing an initial quantum graph. See
398 `quantum_graph_skeleton.QuantumGraphSkeleton` docs for details.
399 After this is returned, the object may be modified in-place in
400 unspecified ways.
402 Notes
403 -----
404 In addition to returning a
405 `quantum_graph_skeleton.QuantumGraphSkeleton`, this method should
406 populate the `existing_datasets` structure by querying for all relevant
407 datasets with non-empty data IDs (those with empty data IDs will
408 already be present). In particular:
410 - `~ExistingDatasets.inputs` must always be populated with all
411 overall-input datasets (but not prerequisites), by querying
412 `input_collections`;
413 - `~ExistingDatasets.outputs_for_skip` must be populated with any
414 intermediate our output datasets present in `skip_existing_in` (it
415 can be ignored if `skip_existing_in` is empty);
416 - `~ExistingDatasets.outputs_in_the_way` must be populated with any
417 intermediate or output datasets present in `output_run`, if
418 `output_run_exists` (it can be ignored if `output_run_exists` is
419 `False`). Note that the presence of such datasets is not
420 automatically an error, even if `clobber is `False`, as these may be
421 quanta that will be skipped.
422 - `~ExistingDatasets.inputs` must be populated with all
423 prerequisite-input datasets that were included in the skeleton, by
424 querying `input_collections` (not all prerequisite inputs need to be
425 included in the skeleton, but the base class can only use per-quantum
426 queries to find them, and that can be slow when there are many
427 quanta).
429 Dataset types should never be components and should always use the
430 "common" storage class definition in `pipeline_graph.DatasetTypeNode`
431 (which is the data repository definition when the dataset type is
432 registered).
433 """
434 raise NotImplementedError()
436 @final
437 @timeMethod
438 def _resolve_task_quanta(self, task_node: TaskNode, skeleton: QuantumGraphSkeleton) -> None:
439 """Process the quanta for one task in a skeleton graph to skip those
440 that have already completed and adjust those that request it.
442 Parameters
443 ----------
444 task_node : `pipeline_graph.TaskNode`
445 Node for this task in the pipeline graph.
446 skeleton : `quantum_graph_skeleton.QuantumGraphSkeleton`
447 Preliminary quantum graph, to be modified in-place.
449 Notes
450 -----
451 This method modifies ``skeleton`` in-place in several ways:
453 - It adds a "ref" attribute to dataset nodes, using the contents of
454 `existing_datasets`. This ensures producing and consuming tasks
455 start from the same `DatasetRef`.
456 - It adds "inputs", "outputs", and "init_inputs" attributes to the
457 quantum nodes, holding the same `NamedValueMapping` objects needed to
458 construct an actual `Quantum` instances.
459 - It removes quantum nodes that are to be skipped because their outputs
460 already exist in `skip_existing_in`. It also removes their outputs
461 from `ExistingDatasets.outputs_in_the_way`.
462 - It adds prerequisite dataset nodes and edges that connect them to the
463 quanta that consume them.
464 - It removes quantum nodes whose
465 `~PipelineTaskConnections.adjustQuantum` calls raise `NoWorkFound` or
466 predict no outputs;
467 - It removes the nodes of output datasets that are "adjusted away".
468 - It removes the edges of input datasets that are "adjusted away".
470 The difference between how adjusted inputs and outputs are handled
471 reflects the fact that many quanta can share the same input, but only
472 one produces each output. This can lead to the graph having
473 superfluous isolated nodes after processing is complete, but these
474 should only be removed after all the quanta from all tasks have been
475 processed.
476 """
477 # Extract the helper object for the prerequisite inputs of this task,
478 # and tell it to prepare to construct skypix bounds and timespans for
479 # each quantum (these will automatically do nothing if nothing needs
480 # those bounds).
481 task_prerequisite_info = self.prerequisite_info[task_node.label]
482 task_prerequisite_info.update_bounds()
483 # Loop over all quanta for this task, remembering the ones we've
484 # gotten rid of.
485 skipped_quanta = []
486 no_work_quanta = []
487 for quantum_key in skeleton.get_quanta(task_node.label):
488 if self._skip_quantum_if_metadata_exists(task_node, quantum_key, skeleton):
489 skipped_quanta.append(quantum_key)
490 continue
491 quantum_data_id = skeleton[quantum_key]["data_id"]
492 skypix_bounds_builder = task_prerequisite_info.bounds.make_skypix_bounds_builder(quantum_data_id)
493 timespan_builder = task_prerequisite_info.bounds.make_timespan_builder(quantum_data_id)
494 adjusted_outputs = self._gather_quantum_outputs(
495 task_node, quantum_key, skeleton, skypix_bounds_builder, timespan_builder
496 )
497 adjusted_inputs = self._gather_quantum_inputs(
498 task_node,
499 quantum_key,
500 skeleton,
501 task_prerequisite_info,
502 skypix_bounds_builder,
503 timespan_builder,
504 )
505 # Give the task's Connections class an opportunity to remove
506 # some inputs, or complain if they are unacceptable. This will
507 # raise if one of the check conditions is not met, which is the
508 # intended behavior.
509 helper = AdjustQuantumHelper(inputs=adjusted_inputs, outputs=adjusted_outputs)
510 try:
511 helper.adjust_in_place(
512 task_node._get_imported_data().connections, task_node.label, quantum_data_id
513 )
514 except NoWorkFound as err:
515 # Do not generate this quantum; it would not produce any
516 # outputs. Remove it and all of the outputs it might have
517 # produced from the skeleton.
518 try:
519 _, connection_name, _ = err.args
520 details = f"not enough datasets for connection {connection_name}."
521 except ValueError:
522 details = str(err)
523 self.log.debug(
524 "No work found for quantum %s of task %s: %s",
525 quantum_key.data_id_values,
526 quantum_key.task_label,
527 details,
528 )
529 no_work_quanta.append(quantum_key)
530 continue
531 if helper.outputs_adjusted:
532 if not any(adjusted_refs for adjusted_refs in helper.outputs.values()):
533 # No outputs also means we don't generate this quantum.
534 self.log.debug(
535 "No outputs predicted for quantum %s of task %s.",
536 quantum_key.data_id_values,
537 quantum_key.task_label,
538 )
539 no_work_quanta.append(quantum_key)
540 continue
541 # Remove output nodes that were not retained by
542 # adjustQuantum.
543 skeleton.remove_dataset_nodes(
544 self._find_removed(skeleton.iter_outputs_of(quantum_key), helper.outputs)
545 )
546 if helper.inputs_adjusted:
547 if not any(bool(adjusted_refs) for adjusted_refs in helper.inputs.values()):
548 raise QuantumGraphBuilderError(
549 f"adjustQuantum implementation for {task_node.label}@{quantum_key.data_id_values} "
550 "returned outputs but no inputs."
551 )
552 # Remove input dataset edges that were not retained by
553 # adjustQuantum. We can't remove the input dataset nodes
554 # because some other quantum might still want them.
555 skeleton.remove_input_edges(
556 quantum_key, self._find_removed(skeleton.iter_inputs_of(quantum_key), helper.inputs)
557 )
558 # Save the adjusted inputs and outputs to the quantum node's
559 # state so we don't have to regenerate those data structures
560 # from the graph.
561 skeleton[quantum_key]["inputs"] = helper.inputs
562 skeleton[quantum_key]["outputs"] = helper.outputs
563 for no_work_quantum in no_work_quanta:
564 skeleton.remove_quantum_node(no_work_quantum, remove_outputs=True)
565 for skipped_quantum in skipped_quanta:
566 skeleton.remove_quantum_node(skipped_quantum, remove_outputs=False)
567 remaining_quanta = skeleton.get_quanta(task_node.label)
568 self._resolve_task_init(task_node, skeleton, bool(skipped_quanta))
569 message_terms = []
570 if no_work_quanta:
571 message_terms.append(f"{len(no_work_quanta)} had no work to do")
572 if skipped_quanta:
573 message_terms.append(f"{len(skipped_quanta)} previously succeeded")
574 message_parenthetical = f" ({', '.join(message_terms)})" if message_terms else ""
575 if remaining_quanta:
576 self.log.info(
577 "Generated %s for task %s%s.",
578 _quantum_or_quanta(len(remaining_quanta)),
579 task_node.label,
580 message_parenthetical,
581 )
582 else:
583 self.log.info(
584 "Dropping task %s because no quanta remain%s.", task_node.label, message_parenthetical
585 )
586 skeleton.remove_task(task_node.label)
588 def _skip_quantum_if_metadata_exists(
589 self, task_node: TaskNode, quantum_key: QuantumKey, skeleton: QuantumGraphSkeleton
590 ) -> bool:
591 """Identify and drop quanta that should be skipped because their
592 metadata datasets already exist.
594 Parameters
595 ----------
596 task_node : `pipeline_graph.TaskNode`
597 Node for this task in the pipeline graph.
598 quantum_key : `QuantumKey`
599 Identifier for this quantum in the graph.
600 skeleton : `quantum_graph_skeleton.QuantumGraphSkeleton`
601 Preliminary quantum graph, to be modified in-place.
603 Returns
604 -------
605 skipped : `bool`
606 `True` if the quantum is being skipped and has been removed from
607 the graph, `False` otherwise.
609 Notes
610 -----
611 If the metadata dataset for this quantum exists in
612 `ExistingDatasets.outputs_for_skip`, the quantum will be skipped. This
613 causes the quantum node to be removed from the graph. Dataset nodes
614 that were previously the outputs of this quantum will have their "ref"
615 attribute set from `ExistingDatasets.outputs_for_skip`, or will be
616 removed if there is no such dataset there. Any output dataset in
617 `ExistingDatasets.outputs_in_the_way` will be removed.
618 """
619 metadata_dataset_key = DatasetKey(
620 task_node.metadata_output.parent_dataset_type_name, quantum_key.data_id_values
621 )
622 if metadata_dataset_key in self.existing_datasets.outputs_for_skip:
623 # This quantum's metadata is already present in the the
624 # skip_existing_in collections; we'll skip it. But the presence of
625 # the metadata dataset doesn't guarantee that all of the other
626 # outputs we predicted are present; we have to check.
627 for output_dataset_key in list(skeleton.iter_outputs_of(quantum_key)):
628 if (
629 output_ref := self.existing_datasets.outputs_for_skip.get(output_dataset_key)
630 ) is not None:
631 # Populate the skeleton graph's node attributes
632 # with the existing DatasetRef, just like a
633 # predicted output of a non-skipped quantum.
634 skeleton[output_dataset_key]["ref"] = output_ref
635 else:
636 # Remove this dataset from the skeleton graph,
637 # because the quantum that would have produced it
638 # is being skipped and it doesn't already exist.
639 skeleton.remove_dataset_nodes([output_dataset_key])
640 # If this dataset was "in the way" (i.e. already in the
641 # output run), it isn't anymore.
642 self.existing_datasets.outputs_in_the_way.pop(output_dataset_key, None)
643 # Removing the quantum node from the graph will happen outside this
644 # function.
645 return True
646 return False
648 @final
649 def _gather_quantum_outputs(
650 self,
651 task_node: TaskNode,
652 quantum_key: QuantumKey,
653 skeleton: QuantumGraphSkeleton,
654 skypix_bounds_builder: SkyPixBoundsBuilder,
655 timespan_builder: TimespanBuilder,
656 ) -> NamedKeyDict[DatasetType, list[DatasetRef]]:
657 """Collect outputs or generate datasets for a preliminary quantum and
658 put them in the form used by `~lsst.daf.butler.Quantum` and
659 `~PipelineTaskConnections.adjustQuantum`.
661 Parameters
662 ----------
663 task_node : `pipeline_graph.TaskNode`
664 Node for this task in the pipeline graph.
665 quantum_key : `QuantumKey`
666 Identifier for this quantum in the graph.
667 skeleton : `quantum_graph_skeleton.QuantumGraphSkeleton`
668 Preliminary quantum graph, to be modified in-place.
669 skypix_bounds_builder : `~prerequisite_helpers.SkyPixBoundsBuilder`
670 An object that accumulates the appropriate spatial bounds for a
671 quantum.
672 timespan_builder : `~prerequisite_helpers.TimespanBuilder`
673 An object that accumulates the appropriate timespan for a quantum.
675 Returns
676 -------
677 outputs : `~lsst.daf.butler.NamedKeyDict` [ \
678 `~lsst.daf.butler.DatasetType`, `list` [ \
679 `~lsst.daf.butler.DatasetRef` ] ]
680 All outputs to the task, using the storage class and components
681 defined by the task's own connections.
683 Notes
684 -----
685 This first looks for outputs already present in the `output_run` by
686 looking in `ExistingDatasets.outputs_in_the_way`; if it finds something
687 and `clobber` is `True`, it uses that ref (it's not ideal that both the
688 original dataset and its replacement will have the same UUID, but we
689 don't have space in the quantum graph for two UUIDs, and we need the
690 datastore records of the original there). If `clobber` is `False`,
691 `RuntimeError` is raised. If there is no output already present, a new
692 one with a random UUID is generated. In all cases the "ref" attribute
693 of the dataset node in the skeleton is set.
694 """
695 outputs_by_type: dict[str, list[DatasetRef]] = {}
696 dataset_key: DatasetKey
697 for dataset_key in skeleton.iter_outputs_of(quantum_key):
698 dataset_data_id = skeleton[dataset_key]["data_id"]
699 dataset_type_node = self._pipeline_graph.dataset_types[dataset_key.parent_dataset_type_name]
700 if (ref := self.existing_datasets.outputs_in_the_way.get(dataset_key)) is None:
701 ref = DatasetRef(dataset_type_node.dataset_type, dataset_data_id, run=self.output_run)
702 elif not self.clobber:
703 # We intentionally raise here, before running adjustQuantum,
704 # because it'd be weird if we left an old potential output of a
705 # task sitting there in the output collection, just because the
706 # task happened to not actually produce it.
707 raise OutputExistsError(
708 f"Potential output dataset {ref} already exists in the output run "
709 f"{self.output_run}, but clobbering outputs was not expected to be necessary."
710 )
711 skypix_bounds_builder.handle_dataset(dataset_key.parent_dataset_type_name, dataset_data_id)
712 timespan_builder.handle_dataset(dataset_key.parent_dataset_type_name, dataset_data_id)
713 skeleton[dataset_key]["ref"] = ref
714 outputs_by_type.setdefault(dataset_key.parent_dataset_type_name, []).append(ref)
715 adapted_outputs: NamedKeyDict[DatasetType, list[DatasetRef]] = NamedKeyDict()
716 for write_edge in task_node.iter_all_outputs():
717 dataset_type_node = self._pipeline_graph.dataset_types[write_edge.parent_dataset_type_name]
718 edge_dataset_type = write_edge.adapt_dataset_type(dataset_type_node.dataset_type)
719 adapted_outputs[edge_dataset_type] = [
720 write_edge.adapt_dataset_ref(ref)
721 for ref in sorted(outputs_by_type.get(write_edge.parent_dataset_type_name, []))
722 ]
723 return adapted_outputs
725 @final
726 def _gather_quantum_inputs(
727 self,
728 task_node: TaskNode,
729 quantum_key: QuantumKey,
730 skeleton: QuantumGraphSkeleton,
731 task_prerequisite_info: PrerequisiteInfo,
732 skypix_bounds_builder: SkyPixBoundsBuilder,
733 timespan_builder: TimespanBuilder,
734 ) -> NamedKeyDict[DatasetType, list[DatasetRef]]:
735 """Collect input datasets for a preliminary quantum and put them in the
736 form used by `~lsst.daf.butler.Quantum` and
737 `~PipelineTaskConnections.adjustQuantum`.
739 Parameters
740 ----------
741 task_node : `pipeline_graph.TaskNode`
742 Node for this task in the pipeline graph.
743 quantum_key : `QuantumKey`
744 Identifier for this quantum in the graph.
745 skeleton : `quantum_graph_skeleton.QuantumGraphSkeleton`
746 Preliminary quantum graph, to be modified in-place.
747 skypix_bounds_builder : `~prerequisite_helpers.SkyPixBoundsBuilder`
748 An object that accumulates the appropriate spatial bounds for a
749 quantum.
750 timespan_builder : `~prerequisite_helpers.TimespanBuilder`
751 An object that accumulates the appropriate timespan for a quantum.
753 Returns
754 -------
755 inputs : `~lsst.daf.butler.NamedKeyDict` [ \
756 `~lsst.daf.butler.DatasetType`, `list` [ \
757 `~lsst.daf.butler.DatasetRef` ] ]
758 All regular and prerequisite inputs to the task, using the storage
759 class and components defined by the task's own connections.
761 Notes
762 -----
763 On return, the dataset nodes that represent inputs to this quantum will
764 either have their "ref" attribute set (using the common dataset type,
765 not the task-specific one) or will be removed from the graph.
767 For regular inputs, usually an existing "ref" (corresponding to an
768 output of another quantum) will be found and left unchanged. When
769 there is no existing "ref" attribute, `ExistingDatasets.inputs` is
770 searched next; if there is nothing there, the input will be removed.
772 Prerequisite inputs are always queried for directly here (delegating to
773 `_find_prerequisite_inputs`). They are never produced by other tasks,
774 and cannot in general be queried for in advance when
775 `ExistingDatasets.inputs` is populated.
776 """
777 quantum_data_id = skeleton[quantum_key]["data_id"]
778 inputs_by_type: dict[str, set[DatasetRef]] = {}
779 dataset_key: DatasetKey | PrerequisiteDatasetKey
780 # Process inputs already present in the skeleton - this should include
781 # all regular inputs (including intermediates) and may include some
782 # prerequisites.
783 for dataset_key in list(skeleton.iter_inputs_of(quantum_key)):
784 if (ref := skeleton[dataset_key].get("ref")) is None:
785 # This dataset is an overall input - if it was an intermediate,
786 # we would have already either removed the node or set the
787 # "ref" attribute when processing its producing quantum - and
788 # this is the first time we're trying to resolve it.
789 if (ref := self.existing_datasets.inputs.get(dataset_key)) is None:
790 # It also doesn't exist in the input collections, so we
791 # remove its node in the skeleton graph (so other consumers
792 # won't have to check for it).
793 skeleton.remove_dataset_nodes([dataset_key])
794 continue
795 skeleton[dataset_key]["ref"] = ref
796 inputs_by_type.setdefault(dataset_key.parent_dataset_type_name, set()).add(ref)
797 skypix_bounds_builder.handle_dataset(dataset_key.parent_dataset_type_name, ref.dataId)
798 timespan_builder.handle_dataset(dataset_key.parent_dataset_type_name, ref.dataId)
799 # Query for any prerequisites not handled by process_subgraph. Note
800 # that these were not already in the skeleton graph, so we add them
801 # now.
802 skypix_bounds = skypix_bounds_builder.finish()
803 timespan = timespan_builder.finish()
804 for finder in task_prerequisite_info.finders.values():
805 inputs_for_type = inputs_by_type.setdefault(finder.dataset_type_node.name, set())
806 dataset_keys = []
807 for ref in finder.find(
808 self.butler, self.input_collections, quantum_data_id, skypix_bounds, timespan
809 ):
810 dataset_key = skeleton.add_prerequisite_node(ref.datasetType.name, ref=ref)
811 dataset_keys.append(dataset_key)
812 inputs_for_type.add(ref)
813 skeleton.add_input_edges(quantum_key, dataset_keys)
814 adapted_inputs: NamedKeyDict[DatasetType, list[DatasetRef]] = NamedKeyDict()
815 for read_edge in task_node.iter_all_inputs():
816 dataset_type_node = self._pipeline_graph.dataset_types[read_edge.parent_dataset_type_name]
817 edge_dataset_type = read_edge.adapt_dataset_type(dataset_type_node.dataset_type)
818 if (current_dataset_type := adapted_inputs.keys().get(edge_dataset_type.name)) is None:
819 adapted_inputs[edge_dataset_type] = [
820 read_edge.adapt_dataset_ref(ref)
821 for ref in sorted(inputs_by_type.get(read_edge.parent_dataset_type_name, frozenset()))
822 ]
823 elif current_dataset_type != edge_dataset_type:
824 raise NotImplementedError(
825 f"Task {task_node.label!r} has {edge_dataset_type.name!r} as an input via "
826 "two different connections, with two different storage class overrides. "
827 "This is not yet supported due to limitations in the Quantum data structure."
828 )
829 # If neither the `if` nor the `elif` above match, it means
830 # multiple input connections have exactly the same dataset
831 # type, and hence nothing to do after the first one.
832 return adapted_inputs
834 @final
835 def _resolve_task_init(
836 self, task_node: TaskNode, skeleton: QuantumGraphSkeleton, has_skipped_quanta: bool
837 ) -> None:
838 """Add init-input and init-output dataset nodes and edges for a task to
839 the skeleton.
841 Parameters
842 ----------
843 task_node : `pipeline_graph.TaskNode`
844 Pipeline graph description of the task.
845 skeleton : `QuantumGraphSkeleton`
846 In-progress quantum graph data structure to update in-place.
847 has_skipped_quanta : `bool`
848 Whether any of this task's quanta were skipped because they had
849 already succeeded.
850 """
851 quanta = skeleton.get_quanta(task_node.label)
852 task_init_key = TaskInitKey(task_node.label)
853 if quanta:
854 adapted_inputs: NamedKeyDict[DatasetType, DatasetRef] = NamedKeyDict()
855 # Process init-inputs.
856 input_keys: list[DatasetKey] = []
857 for read_edge in task_node.init.iter_all_inputs():
858 dataset_key = skeleton.add_dataset_node(
859 read_edge.parent_dataset_type_name, self.empty_data_id
860 )
861 skeleton.add_input_edge(task_init_key, dataset_key)
862 if (ref := skeleton[dataset_key].get("ref")) is None:
863 try:
864 ref = self.existing_datasets.inputs[dataset_key]
865 except KeyError:
866 raise InitInputMissingError(
867 f"Overall init-input dataset {read_edge.parent_dataset_type_name!r} "
868 f"needed by task {task_node.label!r} not found in input collection(s) "
869 f"{self.input_collections}."
870 ) from None
871 skeleton[dataset_key]["ref"] = ref
872 for quantum_key in skeleton.get_quanta(task_node.label):
873 skeleton.add_input_edge(quantum_key, dataset_key)
874 input_keys.append(dataset_key)
875 adapted_ref = read_edge.adapt_dataset_ref(ref)
876 adapted_inputs[adapted_ref.datasetType] = adapted_ref
877 # Save the quantum-adapted init inputs to each quantum, and add
878 # skeleton edges connecting the init inputs to each quantum.
879 for quantum_key in skeleton.get_quanta(task_node.label):
880 skeleton[quantum_key]["init_inputs"] = adapted_inputs
881 # Process init-outputs.
882 adapted_outputs: NamedKeyDict[DatasetType, DatasetRef] = NamedKeyDict()
883 for write_edge in task_node.init.iter_all_outputs():
884 dataset_key = skeleton.add_dataset_node(
885 write_edge.parent_dataset_type_name, self.empty_data_id
886 )
887 if (ref := self.existing_datasets.outputs_in_the_way.get(dataset_key)) is None:
888 ref = DatasetRef(
889 self._pipeline_graph.dataset_types[write_edge.parent_dataset_type_name].dataset_type,
890 self.empty_data_id,
891 run=self.output_run,
892 )
893 skeleton[dataset_key]["ref"] = ref
894 skeleton.add_output_edge(task_init_key, dataset_key)
895 adapted_ref = write_edge.adapt_dataset_ref(ref)
896 adapted_outputs[adapted_ref.datasetType] = adapted_ref
897 skeleton[task_init_key]["inputs"] = adapted_inputs
898 skeleton[task_init_key]["outputs"] = adapted_outputs
899 elif has_skipped_quanta:
900 # No quanta remain for this task, but at least one quantum was
901 # skipped because its outputs were present in the skip_existing_in
902 # collections. This means all init outputs should be present in
903 # the skip_existing_in collections, too, and we need to put those
904 # refs in the graph.
905 for write_edge in task_node.init.iter_all_outputs():
906 dataset_key = skeleton.add_dataset_node(
907 write_edge.parent_dataset_type_name, self.empty_data_id
908 )
909 if (ref := self.existing_datasets.outputs_for_skip.get(dataset_key)) is None:
910 raise InitInputMissingError(
911 f"Init-output dataset {write_edge.parent_dataset_type_name!r} of skipped task "
912 f"{task_node.label!r} not found in skip-existing-in collection(s) "
913 f"{self.skip_existing_in}."
914 ) from None
915 skeleton[dataset_key]["ref"] = ref
916 # If this dataset was "in the way" (i.e. already in the output
917 # run), it isn't anymore.
918 self.existing_datasets.outputs_in_the_way.pop(dataset_key, None)
919 # No quanta remain in this task, but none were skipped; this means
920 # they all got pruned because of NoWorkFound conditions. This
921 # dooms all downstream quanta to the same fate, so we don't bother
922 # doing anything with the task's init-outputs, since nothing is
923 # going to consume them.
925 @final
926 @timeMethod
927 def _find_empty_dimension_datasets(self) -> None:
928 """Query for all dataset types with no dimensions, updating
929 `existing_datasets` in-place.
931 This includes but is not limited to init inputs and init outputs.
932 """
933 _, dataset_type_nodes = self._pipeline_graph.group_by_dimensions()[self.universe.empty.as_group()]
934 dataset_types = [node.dataset_type for node in dataset_type_nodes.values()]
935 dataset_types.extend(self._global_init_output_types.values())
936 for dataset_type in dataset_types:
937 key = DatasetKey(dataset_type.name, self.empty_data_id.required_values)
938 if (
939 self._pipeline_graph.producer_of(dataset_type.name) is None
940 and dataset_type.name not in self._global_init_output_types
941 ):
942 # Dataset type is an overall input; we always need to try to
943 # find these.
944 try:
945 ref = self.butler.find_dataset(dataset_type.name, collections=self.input_collections)
946 except MissingDatasetTypeError:
947 ref = None
948 if ref is not None:
949 self.existing_datasets.inputs[key] = ref
950 elif self.skip_existing_in:
951 # Dataset type is an intermediate or output; need to find these
952 # if only they're from previously executed quanta that we might
953 # skip...
954 try:
955 ref = self.butler.find_dataset(dataset_type.name, collections=self.skip_existing_in)
956 except MissingDatasetTypeError:
957 ref = None
958 if ref is not None:
959 self.existing_datasets.outputs_for_skip[key] = ref
960 if ref.run == self.output_run:
961 self.existing_datasets.outputs_in_the_way[key] = ref
962 if self.output_run_exists and not self.skip_existing_starts_with_output_run:
963 # ...or if they're in the way and would need to be clobbered
964 # (and we haven't already found them in the previous block).
965 try:
966 ref = self.butler.find_dataset(dataset_type.name, collections=[self.output_run])
967 except MissingDatasetTypeError:
968 ref = None
969 if ref is not None:
970 self.existing_datasets.outputs_in_the_way[key] = ref
972 @final
973 @timeMethod
974 def _attach_datastore_records(self, skeleton: QuantumGraphSkeleton) -> None:
975 """Add datastore records for all overall inputs to a preliminary
976 quantum graph.
978 Parameters
979 ----------
980 skeleton : `quantum_graph_skeleton.QuantumGraphSkeleton`
981 Preliminary quantum graph to update in place.
983 Notes
984 -----
985 On return, all quantum nodes in the skeleton graph will have a
986 "datastore_records" attribute that is a mapping from datastore name
987 to `lsst.daf.butler.DatastoreRecordData`, as used by
988 `lsst.daf.butler.Quantum`.
989 """
990 overall_inputs = skeleton.extract_overall_inputs()
991 exported_records = self.butler._datastore.export_records(overall_inputs.values())
992 for quantum_key in skeleton.iter_all_quanta():
993 quantum_records = {}
994 input_ids = {
995 ref.id
996 for dataset_key in skeleton.iter_inputs_of(quantum_key)
997 if (ref := overall_inputs.get(dataset_key)) is not None
998 }
999 if input_ids:
1000 for datastore_name, records in exported_records.items():
1001 matching_records = records.subset(input_ids)
1002 if matching_records is not None:
1003 quantum_records[datastore_name] = matching_records
1004 skeleton[quantum_key]["datastore_records"] = quantum_records
1006 @final
1007 @timeMethod
1008 def _construct_quantum_graph(
1009 self, skeleton: QuantumGraphSkeleton, metadata: Mapping[str, Any]
1010 ) -> QuantumGraph:
1011 """Construct a `QuantumGraph` object from the contents of a
1012 fully-processed `quantum_graph_skeleton.QuantumGraphSkeleton`.
1014 Parameters
1015 ----------
1016 skeleton : `quantum_graph_skeleton.QuantumGraphSkeleton`
1017 Preliminary quantum graph. Must have "init_inputs", "inputs", and
1018 "outputs" attributes on all quantum nodes, as added by
1019 `_resolve_task_quanta`, as well as a "datastore_records" attribute
1020 as added by `_attach_datastore_records`.
1021 metadata : `Mapping`
1022 Flexible metadata to add to the graph.
1024 Returns
1025 -------
1026 quantum_graph : `QuantumGraph`
1027 DAG describing processing to be performed.
1028 """
1029 quanta: dict[TaskDef, set[Quantum]] = {}
1030 init_inputs: dict[TaskDef, Iterable[DatasetRef]] = {}
1031 init_outputs: dict[TaskDef, Iterable[DatasetRef]] = {}
1032 for task_def in self._pipeline_graph._iter_task_defs():
1033 if not skeleton.has_task(task_def.label):
1034 continue
1035 task_node = self._pipeline_graph.tasks[task_def.label]
1036 task_init_key = skeleton.get_task_init_node(task_def.label)
1037 init_inputs[task_def] = skeleton[task_init_key]["inputs"].values()
1038 init_outputs[task_def] = skeleton[task_init_key]["outputs"].values()
1039 quanta_for_task: set[Quantum] = set()
1040 for quantum_key in skeleton.get_quanta(task_node.label):
1041 node_state = skeleton[quantum_key]
1042 quanta_for_task.add(
1043 Quantum(
1044 taskName=task_node.task_class_name,
1045 taskClass=task_node.task_class,
1046 dataId=node_state["data_id"],
1047 initInputs=node_state["init_inputs"],
1048 inputs=node_state["inputs"],
1049 outputs=node_state["outputs"],
1050 datastore_records=node_state.get("datastore_records"),
1051 )
1052 )
1053 quanta[task_def] = quanta_for_task
1055 registry_dataset_types: list[DatasetType] = [
1056 node.dataset_type for node in self._pipeline_graph.dataset_types.values()
1057 ]
1059 all_metadata = self.metadata.to_dict()
1060 all_metadata.update(metadata)
1061 return QuantumGraph(
1062 quanta,
1063 metadata=all_metadata,
1064 universe=self.universe,
1065 initInputs=init_inputs,
1066 initOutputs=init_outputs,
1067 globalInitOutputs=[skeleton[key]["ref"] for key in skeleton.global_init_outputs],
1068 registryDatasetTypes=registry_dataset_types,
1069 )
1071 @staticmethod
1072 @final
1073 def _find_removed(
1074 original: Iterable[DatasetKey | PrerequisiteDatasetKey],
1075 adjusted: NamedKeyMapping[DatasetType, Sequence[DatasetRef]],
1076 ) -> set[DatasetKey | PrerequisiteDatasetKey]:
1077 """Identify skeleton-graph dataset nodes that have been removed by
1078 `~PipelineTaskConnections.adjustQuantum`.
1080 Parameters
1081 ----------
1082 original : `~collections.abc.Iterable` [ `DatasetKey` or \
1083 `PrerequisiteDatasetKey` ]
1084 Identifiers for the dataset nodes that were the original neighbors
1085 (inputs or outputs) of a quantum.
1086 adjusted : `~lsst.daf.butler.NamedKeyMapping` [ \
1087 `~lsst.daf.butler.DatasetType`, \
1088 `~collections.abc.Sequence` [ `lsst.daf.butler.DatasetType` ] ]
1089 Adjusted neighbors, in the form used by `lsst.daf.butler.Quantum`.
1091 Returns
1092 -------
1093 removed : `set` [ `DatasetKey` ]
1094 Datasets in ``original`` that have no counterpart in ``adjusted``.
1095 """
1096 result = set(original)
1097 for dataset_type, kept_refs in adjusted.items():
1098 parent_dataset_type_name, _ = DatasetType.splitDatasetTypeName(dataset_type.name)
1099 for kept_ref in kept_refs:
1100 # We don't know if this was a DatasetKey or a
1101 # PrerequisiteDatasetKey; just try both.
1102 result.discard(DatasetKey(parent_dataset_type_name, kept_ref.dataId.required_values))
1103 result.discard(PrerequisiteDatasetKey(parent_dataset_type_name, kept_ref.id.bytes))
1104 return result
1107@dataclasses.dataclass(eq=False, order=False)
1108class ExistingDatasets:
1109 """Struct that holds the results of dataset queries for
1110 `QuantumGraphBuilder`.
1111 """
1113 inputs: dict[DatasetKey | PrerequisiteDatasetKey, DatasetRef] = dataclasses.field(default_factory=dict)
1114 """Overall-input datasets found in `QuantumGraphBuilder.input_collections`.
1116 This may include prerequisite inputs. It does include init-inputs.
1117 It does not include intermediates.
1118 """
1120 outputs_for_skip: dict[DatasetKey, DatasetRef] = dataclasses.field(default_factory=dict)
1121 """Output datasets found in `QuantumGraphBuilder.skip_existing_in`.
1123 It is unspecified whether this contains include init-outputs; there is
1124 no concept of skipping at the init stage, so this is not expected to
1125 matter.
1126 """
1128 outputs_in_the_way: dict[DatasetKey, DatasetRef] = dataclasses.field(default_factory=dict)
1129 """Output datasets found in `QuantumGraphBuilder.output_run`.
1131 This includes regular outputs and init-outputs.
1132 """
1135def _quantum_or_quanta(n: int) -> str:
1136 """Correctly pluralize 'quantum' if needed."""
1137 return f"{n} quanta" if n != 1 else "1 quantum"