"""``AbstractRunner`` is the base class for all ``Pipeline`` runner
implementations.
"""
from __future__ import annotations
import inspect
import itertools as it
import logging
from abc import ABC, abstractmethod
from collections import deque
from concurrent.futures import (
ALL_COMPLETED,
Future,
ThreadPoolExecutor,
as_completed,
wait,
)
from typing import Any, Collection, Iterable, Iterator
from more_itertools import interleave
from pluggy import PluginManager
from kedro.framework.hooks.manager import _NullPluginManager
from kedro.io import DataCatalog, MemoryDataset
from kedro.pipeline import Pipeline
from kedro.pipeline.node import Node
[docs]
class AbstractRunner(ABC):
"""``AbstractRunner`` is the base class for all ``Pipeline`` runner
implementations.
"""
[docs]
def __init__(
self,
is_async: bool = False,
extra_dataset_patterns: dict[str, dict[str, Any]] | None = None,
):
"""Instantiates the runner class.
Args:
is_async: If True, the node inputs and outputs are loaded and saved
asynchronously with threads. Defaults to False.
extra_dataset_patterns: Extra dataset factory patterns to be added to the DataCatalog
during the run. This is used to set the default datasets on the Runner instances.
"""
self._is_async = is_async
self._extra_dataset_patterns = extra_dataset_patterns
@property
def _logger(self) -> logging.Logger:
return logging.getLogger(self.__module__)
[docs]
def run(
self,
pipeline: Pipeline,
catalog: DataCatalog,
hook_manager: PluginManager | None = None,
session_id: str | None = None,
) -> dict[str, Any]:
"""Run the ``Pipeline`` using the datasets provided by ``catalog``
and save results back to the same objects.
Args:
pipeline: The ``Pipeline`` to run.
catalog: The ``DataCatalog`` from which to fetch data.
hook_manager: The ``PluginManager`` to activate hooks.
session_id: The id of the session.
Raises:
ValueError: Raised when ``Pipeline`` inputs cannot be satisfied.
Returns:
Any node outputs that cannot be processed by the ``DataCatalog``.
These are returned in a dictionary, where the keys are defined
by the node outputs.
"""
hook_or_null_manager = hook_manager or _NullPluginManager()
catalog = catalog.shallow_copy()
# Check which datasets used in the pipeline are in the catalog or match
# a pattern in the catalog
registered_ds = [ds for ds in pipeline.datasets() if ds in catalog]
# Check if there are any input datasets that aren't in the catalog and
# don't match a pattern in the catalog.
unsatisfied = pipeline.inputs() - set(registered_ds)
if unsatisfied:
raise ValueError(
f"Pipeline input(s) {unsatisfied} not found in the DataCatalog"
)
# Identify MemoryDataset in the catalog
memory_datasets = {
ds_name
for ds_name, ds in catalog._datasets.items()
if isinstance(ds, MemoryDataset)
}
# Check if there's any output datasets that aren't in the catalog and don't match a pattern
# in the catalog and include MemoryDataset.
free_outputs = pipeline.outputs() - (set(registered_ds) - memory_datasets)
# Register the default dataset pattern with the catalog
catalog = catalog.shallow_copy(
extra_dataset_patterns=self._extra_dataset_patterns
)
if self._is_async:
self._logger.info(
"Asynchronous mode is enabled for loading and saving data"
)
self._run(pipeline, catalog, hook_or_null_manager, session_id) # type: ignore[arg-type]
self._logger.info("Pipeline execution completed successfully.")
return {ds_name: catalog.load(ds_name) for ds_name in free_outputs}
[docs]
def run_only_missing(
self, pipeline: Pipeline, catalog: DataCatalog, hook_manager: PluginManager
) -> dict[str, Any]:
"""Run only the missing outputs from the ``Pipeline`` using the
datasets provided by ``catalog``, and save results back to the
same objects.
Args:
pipeline: The ``Pipeline`` to run.
catalog: The ``DataCatalog`` from which to fetch data.
hook_manager: The ``PluginManager`` to activate hooks.
Raises:
ValueError: Raised when ``Pipeline`` inputs cannot be
satisfied.
Returns:
Any node outputs that cannot be processed by the
``DataCatalog``. These are returned in a dictionary, where
the keys are defined by the node outputs.
"""
free_outputs = pipeline.outputs() - set(catalog.list())
missing = {ds for ds in catalog.list() if not catalog.exists(ds)}
to_build = free_outputs | missing
to_rerun = pipeline.only_nodes_with_outputs(*to_build) + pipeline.from_inputs(
*to_build
)
# We also need any missing datasets that are required to run the
# `to_rerun` pipeline, including any chains of missing datasets.
unregistered_ds = pipeline.datasets() - set(catalog.list())
output_to_unregistered = pipeline.only_nodes_with_outputs(*unregistered_ds)
input_from_unregistered = to_rerun.inputs() & unregistered_ds
to_rerun += output_to_unregistered.to_outputs(*input_from_unregistered)
return self.run(to_rerun, catalog, hook_manager)
@abstractmethod # pragma: no cover
def _run(
self,
pipeline: Pipeline,
catalog: DataCatalog,
hook_manager: PluginManager,
session_id: str | None = None,
) -> None:
"""The abstract interface for running pipelines, assuming that the
inputs have already been checked and normalized by run().
Args:
pipeline: The ``Pipeline`` to run.
catalog: The ``DataCatalog`` from which to fetch data.
hook_manager: The ``PluginManager`` to activate hooks.
session_id: The id of the session.
"""
pass
def _suggest_resume_scenario(
self,
pipeline: Pipeline,
done_nodes: Iterable[Node],
catalog: DataCatalog,
) -> None:
"""
Suggest a command to the user to resume a run after it fails.
The run should be started from the point closest to the failure
for which persisted input exists.
Args:
pipeline: the ``Pipeline`` of the run.
done_nodes: the ``Node``s that executed successfully.
catalog: the ``DataCatalog`` of the run.
"""
remaining_nodes = set(pipeline.nodes) - set(done_nodes)
postfix = ""
if done_nodes:
start_node_names = _find_nodes_to_resume_from(
pipeline=pipeline,
unfinished_nodes=remaining_nodes,
catalog=catalog,
)
start_nodes_str = ",".join(sorted(start_node_names))
postfix += f' --from-nodes "{start_nodes_str}"'
if not postfix:
self._logger.warning(
"No nodes ran. Repeat the previous command to attempt a new run."
)
else:
self._logger.warning(
f"There are {len(remaining_nodes)} nodes that have not run.\n"
"You can resume the pipeline run from the nearest nodes with "
"persisted inputs by adding the following "
f"argument to your previous command:\n{postfix}"
)
def _find_nodes_to_resume_from(
pipeline: Pipeline, unfinished_nodes: Collection[Node], catalog: DataCatalog
) -> set[str]:
"""Given a collection of unfinished nodes in a pipeline using
a certain catalog, find the node names to pass to pipeline.from_nodes()
to cover all unfinished nodes, including any additional nodes
that should be re-run if their outputs are not persisted.
Args:
pipeline: the ``Pipeline`` to find starting nodes for.
unfinished_nodes: collection of ``Node``s that have not finished yet
catalog: the ``DataCatalog`` of the run.
Returns:
Set of node names to pass to pipeline.from_nodes() to continue
the run.
"""
nodes_to_be_run = _find_all_nodes_for_resumed_pipeline(
pipeline, unfinished_nodes, catalog
)
# Find which of the remaining nodes would need to run first (in topo sort)
persistent_ancestors = _find_initial_node_group(pipeline, nodes_to_be_run)
return {n.name for n in persistent_ancestors}
def _find_all_nodes_for_resumed_pipeline(
pipeline: Pipeline, unfinished_nodes: Iterable[Node], catalog: DataCatalog
) -> set[Node]:
"""Breadth-first search approach to finding the complete set of
``Node``s which need to run to cover all unfinished nodes,
including any additional nodes that should be re-run if their outputs
are not persisted.
Args:
pipeline: the ``Pipeline`` to analyze.
unfinished_nodes: the iterable of ``Node``s which have not finished yet.
catalog: the ``DataCatalog`` of the run.
Returns:
A set containing all input unfinished ``Node``s and all remaining
``Node``s that need to run in case their outputs are not persisted.
"""
nodes_to_run = set(unfinished_nodes)
initial_nodes = _nodes_with_external_inputs(unfinished_nodes)
queue, visited = deque(initial_nodes), set(initial_nodes)
while queue:
current_node = queue.popleft()
nodes_to_run.add(current_node)
# Look for parent nodes which produce non-persistent inputs (if those exist)
non_persistent_inputs = _enumerate_non_persistent_inputs(current_node, catalog)
for node in _enumerate_nodes_with_outputs(pipeline, non_persistent_inputs):
if node in visited:
continue
visited.add(node)
queue.append(node)
# Make sure no downstream tasks are skipped
nodes_to_run = set(pipeline.from_nodes(*(n.name for n in nodes_to_run)).nodes)
return nodes_to_run
def _nodes_with_external_inputs(nodes_of_interest: Iterable[Node]) -> set[Node]:
"""For given ``Node``s , find their subset which depends on
external inputs of the ``Pipeline`` they constitute. External inputs
are pipeline inputs not produced by other ``Node``s in the ``Pipeline``.
Args:
nodes_of_interest: the ``Node``s to analyze.
Returns:
A set of ``Node``s that depend on external inputs
of nodes of interest.
"""
p_nodes_of_interest = Pipeline(nodes_of_interest)
p_nodes_with_external_inputs = p_nodes_of_interest.only_nodes_with_inputs(
*p_nodes_of_interest.inputs()
)
return set(p_nodes_with_external_inputs.nodes)
def _enumerate_non_persistent_inputs(node: Node, catalog: DataCatalog) -> set[str]:
"""Enumerate non-persistent input datasets of a ``Node``.
Args:
node: the ``Node`` to check the inputs of.
catalog: the ``DataCatalog`` of the run.
Returns:
Set of names of non-persistent inputs of given ``Node``.
"""
# We use _datasets because they pertain parameter name format
catalog_datasets = catalog._datasets
non_persistent_inputs: set[str] = set()
for node_input in node.inputs:
if node_input.startswith("params:"):
continue
if (
node_input not in catalog_datasets
or catalog_datasets[node_input]._EPHEMERAL
):
non_persistent_inputs.add(node_input)
return non_persistent_inputs
def _enumerate_nodes_with_outputs(
pipeline: Pipeline, outputs: Collection[str]
) -> list[Node]:
"""For given outputs, returns a list containing nodes that
generate them in the given ``Pipeline``.
Args:
pipeline: the ``Pipeline`` to search for nodes in.
outputs: the dataset names to find source nodes for.
Returns:
A list of all ``Node``s that are producing ``outputs``.
"""
parent_pipeline = pipeline.only_nodes_with_outputs(*outputs)
return parent_pipeline.nodes
def _find_initial_node_group(pipeline: Pipeline, nodes: Iterable[Node]) -> list[Node]:
"""Given a collection of ``Node``s in a ``Pipeline``,
find the initial group of ``Node``s to be run (in topological order).
This can be used to define a sub-pipeline with the smallest possible
set of nodes to pass to --from-nodes.
Args:
pipeline: the ``Pipeline`` to search for initial ``Node``s in.
nodes: the ``Node``s to find initial group for.
Returns:
A list of initial ``Node``s to run given inputs (in topological order).
"""
node_names = set(n.name for n in nodes)
if len(node_names) == 0:
return []
sub_pipeline = pipeline.only_nodes(*node_names)
initial_nodes = sub_pipeline.grouped_nodes[0]
return initial_nodes
[docs]
def run_node(
node: Node,
catalog: DataCatalog,
hook_manager: PluginManager,
is_async: bool = False,
session_id: str | None = None,
) -> Node:
"""Run a single `Node` with inputs from and outputs to the `catalog`.
Args:
node: The ``Node`` to run.
catalog: A ``DataCatalog`` containing the node's inputs and outputs.
hook_manager: The ``PluginManager`` to activate hooks.
is_async: If True, the node inputs and outputs are loaded and saved
asynchronously with threads. Defaults to False.
session_id: The session id of the pipeline run.
Raises:
ValueError: Raised if is_async is set to True for nodes wrapping
generator functions.
Returns:
The node argument.
"""
if is_async and inspect.isgeneratorfunction(node.func):
raise ValueError(
f"Async data loading and saving does not work with "
f"nodes wrapping generator functions. Please make "
f"sure you don't use `yield` anywhere "
f"in node {str(node)}."
)
if is_async:
node = _run_node_async(node, catalog, hook_manager, session_id)
else:
node = _run_node_sequential(node, catalog, hook_manager, session_id)
for name in node.confirms:
catalog.confirm(name)
return node
def _collect_inputs_from_hook( # noqa: PLR0913
node: Node,
catalog: DataCatalog,
inputs: dict[str, Any],
is_async: bool,
hook_manager: PluginManager,
session_id: str | None = None,
) -> dict[str, Any]:
inputs = inputs.copy() # shallow copy to prevent in-place modification by the hook
hook_response = hook_manager.hook.before_node_run(
node=node,
catalog=catalog,
inputs=inputs,
is_async=is_async,
session_id=session_id,
)
additional_inputs = {}
if (
hook_response is not None
): # all hooks on a _NullPluginManager will return None instead of a list
for response in hook_response:
if response is not None and not isinstance(response, dict):
response_type = type(response).__name__
raise TypeError(
f"'before_node_run' must return either None or a dictionary mapping "
f"dataset names to updated values, got '{response_type}' instead."
)
additional_inputs.update(response or {})
return additional_inputs
def _call_node_run( # noqa: PLR0913
node: Node,
catalog: DataCatalog,
inputs: dict[str, Any],
is_async: bool,
hook_manager: PluginManager,
session_id: str | None = None,
) -> dict[str, Any]:
try:
outputs = node.run(inputs)
except Exception as exc:
hook_manager.hook.on_node_error(
error=exc,
node=node,
catalog=catalog,
inputs=inputs,
is_async=is_async,
session_id=session_id,
)
raise exc
hook_manager.hook.after_node_run(
node=node,
catalog=catalog,
inputs=inputs,
outputs=outputs,
is_async=is_async,
session_id=session_id,
)
return outputs
def _run_node_sequential(
node: Node,
catalog: DataCatalog,
hook_manager: PluginManager,
session_id: str | None = None,
) -> Node:
inputs = {}
for name in node.inputs:
hook_manager.hook.before_dataset_loaded(dataset_name=name, node=node)
inputs[name] = catalog.load(name)
hook_manager.hook.after_dataset_loaded(
dataset_name=name, data=inputs[name], node=node
)
is_async = False
additional_inputs = _collect_inputs_from_hook(
node, catalog, inputs, is_async, hook_manager, session_id=session_id
)
inputs.update(additional_inputs)
outputs = _call_node_run(
node, catalog, inputs, is_async, hook_manager, session_id=session_id
)
items: Iterable = outputs.items()
# if all outputs are iterators, then the node is a generator node
if all(isinstance(d, Iterator) for d in outputs.values()):
# Python dictionaries are ordered, so we are sure
# the keys and the chunk streams are in the same order
# [a, b, c]
keys = list(outputs.keys())
# [Iterator[chunk_a], Iterator[chunk_b], Iterator[chunk_c]]
streams = list(outputs.values())
# zip an endless cycle of the keys
# with an interleaved iterator of the streams
# [(a, chunk_a), (b, chunk_b), ...] until all outputs complete
items = zip(it.cycle(keys), interleave(*streams))
for name, data in items:
hook_manager.hook.before_dataset_saved(dataset_name=name, data=data, node=node)
catalog.save(name, data)
hook_manager.hook.after_dataset_saved(dataset_name=name, data=data, node=node)
return node
def _run_node_async(
node: Node,
catalog: DataCatalog,
hook_manager: PluginManager,
session_id: str | None = None,
) -> Node:
def _synchronous_dataset_load(dataset_name: str) -> Any:
"""Minimal wrapper to ensure Hooks are run synchronously
within an asynchronous dataset load."""
hook_manager.hook.before_dataset_loaded(dataset_name=dataset_name, node=node)
return_ds = catalog.load(dataset_name)
hook_manager.hook.after_dataset_loaded(
dataset_name=dataset_name, data=return_ds, node=node
)
return return_ds
with ThreadPoolExecutor() as pool:
inputs: dict[str, Future] = {}
for name in node.inputs:
inputs[name] = pool.submit(_synchronous_dataset_load, name)
wait(inputs.values(), return_when=ALL_COMPLETED)
inputs = {key: value.result() for key, value in inputs.items()}
is_async = True
additional_inputs = _collect_inputs_from_hook(
node, catalog, inputs, is_async, hook_manager, session_id=session_id
)
inputs.update(additional_inputs)
outputs = _call_node_run(
node, catalog, inputs, is_async, hook_manager, session_id=session_id
)
future_dataset_mapping = {}
for name, data in outputs.items():
hook_manager.hook.before_dataset_saved(
dataset_name=name, data=data, node=node
)
future = pool.submit(catalog.save, name, data)
future_dataset_mapping[future] = (name, data)
for future in as_completed(future_dataset_mapping):
exception = future.exception()
if exception:
raise exception
name, data = future_dataset_mapping[future]
hook_manager.hook.after_dataset_saved(
dataset_name=name, data=data, node=node
)
return node
