feat(sinks): add a SplitSink

Add `sghi.etl.commons.sink.SplitSink`, a `Sink` that splits aggregated
processed data into its consitituent data parts, before draining each data
part to multiple other sinks concurrently.

src/sghi/etl/commons/__init__.py

@@ -8,7 +8,7 @@
     pipe_processors,
     processor,
 )
-from .sinks import NullSink, sink
+from .sinks import NullSink, SplitSink, sink
 from .sources import GatherSource, source
 from .utils import fail_fast, fail_fast_factory, ignored_failed
 from .workflow_definitions import SimpleWorkflowDefinition
@@ -21,6 +21,7 @@
     "SimpleWorkflowDefinition",
     "ScatterGatherProcessor",
     "SplitGatherProcessor",
+    "SplitSink",
     "fail_fast",
     "fail_fast_factory",
     "ignored_failed",

src/sghi/etl/commons/sinks.py

@@ -3,7 +3,9 @@
 from __future__ import annotations
 
 import logging
-from collections.abc import Callable
+from collections.abc import Callable, Iterable, Sequence
+from concurrent.futures import Executor, Future, ThreadPoolExecutor
+from contextlib import ExitStack
 from functools import update_wrapper
 from logging import Logger
 from typing import Final, Generic, Self, TypeVar, final
@@ -12,7 +14,17 @@
 
 from sghi.disposable import not_disposed
 from sghi.etl.core import Sink
-from sghi.utils import ensure_callable, type_fqn
+from sghi.retry import Retry, noop_retry
+from sghi.task import ConcurrentExecutor, Task, task
+from sghi.utils import (
+    ensure_callable,
+    ensure_instance_of,
+    ensure_not_none_nor_empty,
+    ensure_predicate,
+    type_fqn,
+)
+
+from .utils import fail_fast
 
 # =============================================================================
 # TYPES
@@ -22,6 +34,10 @@
 _PDT = TypeVar("_PDT")
 """Type variable representing the data type after processing."""
 
+_T = TypeVar("_T")
+
+_ResultGatherer = Callable[[Iterable[Future[_T]]], Iterable[_T]]
+
 _SinkCallable = Callable[[_PDT], None]
 
 
@@ -153,6 +169,236 @@ def dispose(self) -> None:
         self._logger.info("Disposal complete.")
 
 
+@final
+class SplitSink(Sink[Sequence[_PDT]], Generic[_PDT]):
+    """A :class:`Sink` that splits processed data and drains the split data to
+    multiple sinks.
+
+    This ``Sink`` implementation takes aggregated processed data, splits it
+    into its constituent data parts, and then drains each data part to each
+    embedded sink concurrently. That is, each data part is mapped to each
+    embedded sink before executing all the embedded sinks concurrently. As
+    such, the supplied processed data's size must equal the number of embedded
+    sinks contained by a sink of this type. A result gatherer function can be
+    provided to specify how to handle errors while draining. A :class:`retry
+    policy<Retry>` to handle transient draining errors to the embedded sinks
+    can also be provided. A suitable :class:`Executor` can be specified at
+    instantiation to control the concurrent draining to the embedded sinks.
+
+    Instances of this class are **NOT SAFE** to retry and **SHOULD NEVER** be
+    retried. However, they do support retrying their embedded sinks. This is
+    disabled by default but can be enabled by providing a suitable value to
+    the ``retry_policy_factory`` constructor parameter when creating new
+    instances. When enabled, each embedded sink will be retried individually
+    per the specified retry policy in case it fails.
+
+    Disposing instances of this class also disposes of their embedded sinks.
+
+    .. admonition:: Regarding retry safety
+        :class: tip
+
+        Instances of this ``Sink`` are **NOT SAFE** to retry.
+    """  # noqa: D205
+
+    __slots__ = (
+        "_sinks",
+        "_retry_policy_factory",
+        "_executor_factory",
+        "_result_gatherer",
+        "_is_disposed",
+        "_logger",
+        "_exit_stack",
+        "_prepped_sinks",
+        "_executor",
+    )
+
+    def __init__(
+        self,
+        sinks: Sequence[Sink[_PDT]],
+        retry_policy_factory: Callable[[], Retry] = noop_retry,
+        executor_factory: Callable[[], Executor] = ThreadPoolExecutor,
+        result_gatherer: _ResultGatherer[None] = fail_fast,
+    ) -> None:
+        """Create a new ``SplitSink`` of the given properties.
+
+        :param sinks: A ``Sequence`` of sinks to drain each processed data part
+            to. These sinks are also referred to as the embedded sinks. The
+            given ``Sequence`` *MUST NOT* be empty.
+        :param retry_policy_factory: A callable that supplies retry policy
+            instance(s) to apply to each embedded sink. This *MUST* be a valid
+            callable object. Defaults to a factory that returns retry policies
+            that do nothing.
+        :param executor_factory: A callable that suppliers suitable
+            ``Executor`` instance(s) to use for the concurrent draining. This
+            *MUST* be a valid callable object. Defaults to a factory that
+            returns ``ThreadPoolExecutor`` instances.
+        :param result_gatherer: A function that specifies how to handle
+            draining errors. This *MUST* be a valid callable object. Defaults
+            to a gatherer that fails if draining to any of the embedded sinks
+            failed, or returns silently otherwise.
+
+        :raise TypeError: If ``sinks`` is NOT a ``Sequence``.
+        :raise ValueError: If ``sinks`` is empty or if
+            ``retry_policy_factory``, ``executor_factory`` and
+            ``result_gatherer`` are NOT callable objects.
+        """
+        super().__init__()
+        ensure_not_none_nor_empty(
+            value=ensure_instance_of(
+                value=sinks,
+                message="'sinks' MUST be a collections.abc.Sequence object.",
+                klass=Sequence,
+            ),
+            message="'sinks' MUST NOT be empty.",
+        )
+        self._sinks: Sequence[Sink[_PDT]] = tuple(sinks)
+        self._retry_policy_factory: Callable[[], Retry] = ensure_callable(
+            value=retry_policy_factory,
+            message="'retry_policy_factory' MUST be a callable.",
+        )
+        self._executor_factory: Callable[[], Executor] = ensure_callable(
+            value=executor_factory,
+            message="'executor_factory' MUST be a callable.",
+        )
+        self._result_gatherer: _ResultGatherer[None] = ensure_callable(
+            value=result_gatherer,
+            message="'result_gatherer' MUST be a callable.",
+        )
+        self._is_disposed: bool = False
+        self._logger: Logger = logging.getLogger(type_fqn(self.__class__))
+        self._exit_stack: ExitStack = ExitStack()
+
+        # Prepare embedded sinks for execution by ensuring that they are all
+        # disposed of properly once this object is disposed.
+        self._prepped_sinks: Sequence[Task[Sequence[_PDT], None]] = tuple(
+            self._sink_to_task(self._exit_stack.push(_sink), _i)
+            for _i, _sink in enumerate(self._sinks)
+        )
+        self._executor: ConcurrentExecutor[Sequence[_PDT], None]
+        self._executor = ConcurrentExecutor(
+            *self._prepped_sinks, executor=self._executor_factory()
+        )
+
+    @not_disposed
+    @override
+    def __enter__(self) -> Self:
+        """Return ``self`` upon entering the runtime context.
+
+        .. admonition:: Don't use after dispose
+            :class: error
+
+            Invoking this method on an instance that is disposed(i.e. the
+            :attr:`is_disposed` property on the instance is ``True``) will
+            result in a :exc:`ResourceDisposedError` being raised.
+
+        :return: This instance.
+
+        :raise ResourceDisposedError: If this sink has already been disposed.
+        """
+        return super(Sink, self).__enter__()
+
+    @property
+    @override
+    def is_disposed(self) -> bool:
+        return self._is_disposed
+
+    @not_disposed
+    @override
+    def drain(self, processed_data: Sequence[_PDT]) -> None:
+        """Split the supplied processed data and consume each data part.
+
+        This method decomposes the provided aggregated processed data into its
+        constituent data parts and maps each data part to an embedded sink;
+        before executing all the embedded sinks concurrently. It then applies
+        the result-gatherer function assigned to this instance (at creation) to
+        the :class:`Future` objects resulting from the concurrent execution.
+        Each of these ``Future`` objects wraps the result of draining each data
+        part to an embedded sink contained in this ``SplitSink``, and they
+        preserve the same order.
+
+        .. admonition:: Don't use after dispose
+            :class: error
+
+            Invoking this method on an instance that is disposed(i.e. the
+            :attr:`is_disposed` property on the instance is ``True``) will
+            result in a :exc:`ResourceDisposedError` being raised.
+
+        :param processed_data: The aggregated processed data to consume. This
+            *MUST* be a ``Sequence`` of processed data values whose size *MUST
+            EQUAL* the number of embedded sinks contained by this
+            ``SplitSink``.
+
+        :return: None.
+
+        :raise ResourceDisposedError: If this sink has already been disposed.
+        :raise TypeError: If ``processed_data`` *IS NOT* a ``Sequence``.
+        :raise ValueError: If the size of ``processed_data`` *DOES NOT EQUAL*
+            the number of embedded sinks in this ``SplitSink``.
+        """
+        ensure_instance_of(
+            value=processed_data,
+            klass=Sequence,
+            message=(
+                "'processed_data' MUST be a collections.abc.Sequence object."
+            ),
+        )
+        ensure_predicate(
+            test=len(processed_data) == len(self._sinks),
+            message=(
+                f"Expected 'processed_data' to be of size {len(self._sinks)} "
+                f"but got size {len(processed_data)} instead."
+            ),
+            exc_factory=ValueError,
+        )
+        self._logger.info(
+            "Splitting aggregated processed data and draining each data part "
+            "to all available sinks."
+        )
+
+        with self._executor as executor:
+            futures = executor.execute(processed_data)
+
+        self._result_gatherer(futures)
+
+    @override
+    def dispose(self) -> None:
+        """Release any underlying resources contained by this sink.
+
+        All embedded sinks are also disposed. After this method returns
+        successfully, the :attr:`is_disposed` property should return ``True``.
+
+        .. note::
+            Unless otherwise specified, trying to use methods of a
+            ``Disposable`` instance decorated with the
+            :func:`~sghi.disposable.not_disposed` decorator after this method
+            returns should generally be considered a programming error and
+            should result in a :exc:`~sghi.disposable.ResourceDisposedError`
+            being raised.
+
+            This method should be idempotent allowing it to be called more
+            than once; only the first call, however, should have an effect.
+
+        :return: None.
+        """
+        self._is_disposed = True
+        self._exit_stack.close()
+        self._executor.dispose()
+        self._logger.info("Disposal complete.")
+
+    def _sink_to_task(
+        self,
+        s: Sink[_PDT],
+        i: int,
+    ) -> Task[Sequence[_PDT], None]:
+        @task
+        def do_drain(processed_data: Sequence[_PDT]) -> None:
+            with s as _s:
+                drain = self._retry_policy_factory().retry(_s.drain)
+                return drain(processed_data[i])
+
+        return do_drain
+
+
 @final
 class _SinkOfCallable(Sink[_PDT], Generic[_PDT]):
     __slots__ = ("_delegate_to", "_is_disposed", "_logger")
@@ -227,5 +473,6 @@ def dispose(self) -> None:
 
 __all__ = [
     "NullSink",
+    "SplitSink",
     "sink",
 ]