path: root/taskflow/storage.py

# -*- coding: utf-8 -*-

#    Copyright (C) 2013 Yahoo! Inc. All Rights Reserved.
#
#    Licensed under the Apache License, Version 2.0 (the "License"); you may
#    not use this file except in compliance with the License. You may obtain
#    a copy of the License at
#
#         http://www.apache.org/licenses/LICENSE-2.0
#
#    Unless required by applicable law or agreed to in writing, software
#    distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
#    WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the
#    License for the specific language governing permissions and limitations
#    under the License.

import contextlib
import functools

import fasteners
from oslo_utils import reflection
from oslo_utils import uuidutils
import six
import tenacity

from taskflow import exceptions
from taskflow import logging
from taskflow.persistence.backends import impl_memory
from taskflow.persistence import models
from taskflow import retry
from taskflow import states
from taskflow import task
from taskflow.utils import misc

LOG = logging.getLogger(__name__)

RETRY_ATTEMPTS = 3
RETRY_WAIT_TIMEOUT = 5

_EXECUTE_STATES_WITH_RESULTS = (
    # The atom ``execute`` worked out :)
    states.SUCCESS,
    # The atom ``execute`` didn't work out :(
    states.FAILURE,
    # In this state we will still have access to prior SUCCESS (or FAILURE)
    # results, so make sure extraction is still allowed in this state...
    states.REVERTING,
)

_REVERT_STATES_WITH_RESULTS = (
    # The atom ``revert`` worked out :)
    states.REVERTED,
    # The atom ``revert`` didn't work out :(
    states.REVERT_FAILURE,
    # In this state we will still have access to prior SUCCESS (or FAILURE)
    # results, so make sure extraction is still allowed in this state...
    states.REVERTING,
)

# Atom states that may have results...
STATES_WITH_RESULTS = set()
STATES_WITH_RESULTS.update(_REVERT_STATES_WITH_RESULTS)
STATES_WITH_RESULTS.update(_EXECUTE_STATES_WITH_RESULTS)
STATES_WITH_RESULTS = tuple(sorted(STATES_WITH_RESULTS))

# TODO(harlowja): do this better (via a singleton or something else...)
_TRANSIENT_PROVIDER = object()

# Only for these intentions will we cache any failures that happened...
_SAVE_FAILURE_INTENTIONS = (states.EXECUTE, states.REVERT)

# NOTE(harlowja): Perhaps the container is a dictionary-like object and that
# key does not exist (key error), or the container is a tuple/list and a
# non-numeric key is being requested (index error), or there was no container
# and an attempt to index into none/other unsubscriptable type is being
# requested (type error).
#
# Overall this (along with the item_from* functions) try to handle the vast
# majority of wrong indexing operations on the wrong/invalid types so that we
# can fail extraction during lookup or emit warning on result reception...
_EXTRACTION_EXCEPTIONS = (IndexError, KeyError, ValueError, TypeError)

# Atom detail metadata key used to inject atom non-transient injected args.
META_INJECTED = 'injected'

# Atom detail metadata key(s) used to set atom progress (with any details).
META_PROGRESS = 'progress'
META_PROGRESS_DETAILS = 'progress_details'


class _ProviderLocator(object):
    """Helper to start to better decouple the finding logic from storage.

    WIP: part of the larger effort to cleanup/refactor the finding of named
         arguments so that the code can be more unified and easy to
         follow...
    """

    def __init__(self, transient_results,
                 providers_fetcher, result_fetcher):
        self.result_fetcher = result_fetcher
        self.providers_fetcher = providers_fetcher
        self.transient_results = transient_results

    def _try_get_results(self, looking_for, provider,
                         look_into_results=True, find_potentials=False):
        if provider.name is _TRANSIENT_PROVIDER:
            # TODO(harlowja): This 'is' check still sucks, do this
            # better in the future...
            results = self.transient_results
        else:
            try:
                results = self.result_fetcher(provider.name)
            except (exceptions.NotFound, exceptions.DisallowedAccess):
                if not find_potentials:
                    raise
                else:
                    # Ok, likely hasn't produced a result yet, but
                    # at a future point it hopefully will, so stub
                    # out the *expected* result.
                    results = {}
        if look_into_results:
            _item_from_single(provider, results, looking_for)
        return results

    def _find(self, looking_for, scope_walker=None,
              short_circuit=True, find_potentials=False):
        if scope_walker is None:
            scope_walker = []
        default_providers, atom_providers = self.providers_fetcher(looking_for)
        searched_providers = set()
        providers_and_results = []
        if default_providers:
            for p in default_providers:
                searched_providers.add(p)
                try:
                    provider_results = self._try_get_results(
                        looking_for, p, find_potentials=find_potentials,
                        # For default providers always look into there
                        # results as default providers are statically setup
                        # and therefore looking into there provided results
                        # should fail early.
                        look_into_results=True)
                except exceptions.NotFound:
                    if not find_potentials:
                        raise
                else:
                    providers_and_results.append((p, provider_results))
            if short_circuit:
                return (searched_providers, providers_and_results)
        if not atom_providers:
            return (searched_providers, providers_and_results)
        atom_providers_by_name = dict((p.name, p) for p in atom_providers)
        for accessible_atom_names in iter(scope_walker):
            # *Always* retain the scope ordering (if any matches
            # happen); instead of retaining the possible provider match
            # order (which isn't that important and may be different from
            # the scope requested ordering).
            maybe_atom_providers = [atom_providers_by_name[atom_name]
                                    for atom_name in accessible_atom_names
                                    if atom_name in atom_providers_by_name]
            tmp_providers_and_results = []
            if find_potentials:
                for p in maybe_atom_providers:
                    searched_providers.add(p)
                    tmp_providers_and_results.append((p, {}))
            else:
                for p in maybe_atom_providers:
                    searched_providers.add(p)
                    try:
                        # Don't at this point look into the provider results
                        # as calling code will grab all providers, and then
                        # get the result from the *first* provider that
                        # actually provided it (or die).
                        provider_results = self._try_get_results(
                            looking_for, p, find_potentials=find_potentials,
                            look_into_results=False)
                    except exceptions.DisallowedAccess as e:
                        if e.state != states.IGNORE:
                            exceptions.raise_with_cause(
                                exceptions.NotFound,
                                "Expected to be able to find output %r"
                                " produced by %s but was unable to get at"
                                " that providers results" % (looking_for, p))
                        else:
                            LOG.trace("Avoiding using the results of"
                                      " %r (from %s) for name %r because"
                                      " it was ignored", p.name, p,
                                      looking_for)
                    else:
                        tmp_providers_and_results.append((p, provider_results))
            if tmp_providers_and_results and short_circuit:
                return (searched_providers, tmp_providers_and_results)
            else:
                providers_and_results.extend(tmp_providers_and_results)
        return (searched_providers, providers_and_results)

    def find_potentials(self, looking_for, scope_walker=None):
        """Returns the accessible **potential** providers."""
        _searched_providers, providers_and_results = self._find(
            looking_for, scope_walker=scope_walker,
            short_circuit=False, find_potentials=True)
        return set(p for (p, _provider_results) in providers_and_results)

    def find(self, looking_for, scope_walker=None, short_circuit=True):
        """Returns the accessible providers."""
        return self._find(looking_for, scope_walker=scope_walker,
                          short_circuit=short_circuit,
                          find_potentials=False)


class _Provider(object):
    """A named symbol provider that produces a output at the given index."""

    def __init__(self, name, index):
        self.name = name
        self.index = index

    def __repr__(self):
        # TODO(harlowja): clean this up...
        if self.name is _TRANSIENT_PROVIDER:
            base = "<TransientProvider"
        else:
            base = "<Provider '%s'" % (self.name)
        if self.index is None:
            base += ">"
        else:
            base += " @ index %r>" % (self.index)
        return base

    def __hash__(self):
        return hash((self.name, self.index))

    def __eq__(self, other):
        return (self.name, self.index) == (other.name, other.index)

    def __ne__(self, other):
        return not self.__eq__(other)


def _item_from(container, index):
    """Attempts to fetch a index/key from a given container."""
    if index is None:
        return container
    return container[index]


def _item_from_single(provider, container, looking_for):
    """Returns item from a *single* provider."""
    try:
        return _item_from(container, provider.index)
    except _EXTRACTION_EXCEPTIONS:
        exceptions.raise_with_cause(
            exceptions.NotFound,
            "Unable to find result %r, expected to be able to find it"
            " created by %s but was unable to perform successful"
            " extraction" % (looking_for, provider))


def _item_from_first_of(providers, looking_for):
    """Returns item from the *first* successful container extraction."""
    for (provider, container) in providers:
        try:
            return (provider, _item_from(container, provider.index))
        except _EXTRACTION_EXCEPTIONS:
            pass
    providers = [p[0] for p in providers]
    raise exceptions.NotFound(
        "Unable to find result %r, expected to be able to find it"
        " created by one of %s but was unable to perform successful"
        " extraction" % (looking_for, providers))


class Storage(object):
    """Interface between engines and logbook and its backend (if any).

    This class provides a simple interface to save atoms of a given flow and
    associated activity and results to persistence layer (logbook,
    atom_details, flow_details) for use by engines. This makes it easier to
    interact with the underlying storage & backend mechanism through this
    interface rather than accessing those objects directly.

    NOTE(harlowja): if no backend is provided then a in-memory backend will
    be automatically used and the provided flow detail object will be placed
    into it for the duration of this objects existence.
    """

    injector_name = '_TaskFlow_INJECTOR'
    """Injector task detail name.

    This task detail is a **special** detail that will be automatically
    created and saved to store **persistent** injected values (name conflicts
    with it must be avoided) that are *global* to the flow being executed.
    """

    def __init__(self, flow_detail, backend=None, scope_fetcher=None):
        self._result_mappings = {}
        self._reverse_mapping = {}
        if backend is None:
            # Err on the likely-hood that most people don't make there
            # objects able to be deepcopyable (resources, locks and such
            # can't be deepcopied)...
            backend = impl_memory.MemoryBackend({'deep_copy': False})
            with contextlib.closing(backend.get_connection()) as conn:
                conn.update_flow_details(flow_detail, ignore_missing=True)
        self._backend = backend
        self._flowdetail = flow_detail
        self._transients = {}
        self._injected_args = {}
        self._lock = fasteners.ReaderWriterLock()
        self._ensure_matchers = [
            ((task.Task,), (models.TaskDetail, 'Task')),
            ((retry.Retry,), (models.RetryDetail, 'Retry')),
        ]
        if scope_fetcher is None:
            scope_fetcher = lambda atom_name: None
        self._scope_fetcher = scope_fetcher

        # NOTE(imelnikov): failure serialization looses information,
        # so we cache failures here, in atom name -> failure mapping.
        self._failures = {}
        for ad in self._flowdetail:
            fail_cache = {}
            if ad.failure is not None:
                fail_cache[states.EXECUTE] = ad.failure
            if ad.revert_failure is not None:
                fail_cache[states.REVERT] = ad.revert_failure
            self._failures[ad.name] = fail_cache

        self._atom_name_to_uuid = dict((ad.name, ad.uuid)
                                       for ad in self._flowdetail)
        try:
            source, _clone = self._atomdetail_by_name(
                self.injector_name, expected_type=models.TaskDetail)
        except exceptions.NotFound:
            pass
        else:
            names_iter = six.iterkeys(source.results)
            self._set_result_mapping(source.name,
                                     dict((name, name) for name in names_iter))

    def _with_connection(self, functor, *args, **kwargs):
        # Run the given functor with a backend connection as its first
        # argument (providing the additional positional arguments and keyword
        # arguments as subsequent arguments).
        with contextlib.closing(self._backend.get_connection()) as conn:
            return functor(conn, *args, **kwargs)

    @staticmethod
    def _create_atom_detail(atom_name, atom_detail_cls,
                            atom_version=None, atom_state=states.PENDING):
        ad = atom_detail_cls(atom_name, uuidutils.generate_uuid())
        ad.state = atom_state
        if atom_version is not None:
            ad.version = atom_version
        return ad

    @fasteners.write_locked
    def ensure_atoms(self, atoms):
        """Ensure there is an atomdetail for **each** of the given atoms.

        Returns list of atomdetail uuids for each atom processed.
        """
        atom_ids = []
        missing_ads = []
        for i, atom in enumerate(atoms):
            match = misc.match_type(atom, self._ensure_matchers)
            if not match:
                raise TypeError("Unknown atom '%s' (%s) requested to ensure"
                                % (atom, type(atom)))
            atom_detail_cls, kind = match
            atom_name = atom.name
            if not atom_name:
                raise ValueError("%s name must be non-empty" % (kind))
            try:
                atom_id = self._atom_name_to_uuid[atom_name]
            except KeyError:
                missing_ads.append((i, atom, atom_detail_cls))
                # This will be later replaced with the uuid that is created...
                atom_ids.append(None)
            else:
                ad = self._flowdetail.find(atom_id)
                if not isinstance(ad, atom_detail_cls):
                    raise exceptions.Duplicate(
                        "Atom detail '%s' already exists in flow"
                        " detail '%s'" % (atom_name, self._flowdetail.name))
                else:
                    atom_ids.append(ad.uuid)
                    self._set_result_mapping(atom_name, atom.save_as)
        if missing_ads:
            needs_to_be_created_ads = []
            for (i, atom, atom_detail_cls) in missing_ads:
                ad = self._create_atom_detail(
                    atom.name, atom_detail_cls,
                    atom_version=misc.get_version_string(atom))
                needs_to_be_created_ads.append((i, atom, ad))
            # Add the atom detail(s) to a clone, which upon success will be
            # updated into the contained flow detail; if it does not get saved
            # then no update will happen.
            source, clone = self._fetch_flowdetail(clone=True)
            for (_i, _atom, ad) in needs_to_be_created_ads:
                clone.add(ad)
            self._with_connection(self._save_flow_detail, source, clone)
            # Insert the needed data, and get outta here...
            for (i, atom, ad) in needs_to_be_created_ads:
                atom_name = atom.name
                atom_ids[i] = ad.uuid
                self._atom_name_to_uuid[atom_name] = ad.uuid
                self._set_result_mapping(atom_name, atom.save_as)
                self._failures.setdefault(atom_name, {})
        return atom_ids

    @property
    def lock(self):
        """Reader/writer lock used to ensure multi-thread safety.

        This does **not** protect against the **same** storage objects being
        used by multiple engines/users across multiple processes (or
        different machines); certain backends handle that situation better
        than others (for example by using sequence identifiers) and it's a
        ongoing work in progress to make that better).
        """
        return self._lock

    def ensure_atom(self, atom):
        """Ensure there is an atomdetail for the **given** atom.

        Returns the uuid for the atomdetail that corresponds to the given atom.
        """
        return self.ensure_atoms([atom])[0]

    @property
    def flow_name(self):
        """The flow detail name this storage unit is associated with."""
        # This never changes (so no read locking needed).
        return self._flowdetail.name

    @property
    def flow_uuid(self):
        """The flow detail uuid this storage unit is associated with."""
        # This never changes (so no read locking needed).
        return self._flowdetail.uuid

    @property
    def flow_meta(self):
        """The flow detail metadata this storage unit is associated with."""
        return self._flowdetail.meta

    @property
    def backend(self):
        """The backend this storage unit is associated with."""
        # This never changes (so no read locking needed).
        return self._backend

    @tenacity.retry(retry=tenacity.retry_if_exception_type(
                    exception_types=exceptions.StorageFailure),
                    stop=tenacity.stop_after_attempt(RETRY_ATTEMPTS),
                    wait=tenacity.wait_fixed(RETRY_WAIT_TIMEOUT))
    def _save_flow_detail(self, conn, original_flow_detail, flow_detail):
        # NOTE(harlowja): we need to update our contained flow detail if
        # the result of the update actually added more (aka another process
        # added item to the flow detail).
        original_flow_detail.update(conn.update_flow_details(flow_detail))
        return original_flow_detail

    def _fetch_flowdetail(self, clone=False):
        source = self._flowdetail
        if clone:
            return (source, source.copy())
        else:
            return (source, source)

    def _atomdetail_by_name(self, atom_name, expected_type=None, clone=False):
        try:
            ad = self._flowdetail.find(self._atom_name_to_uuid[atom_name])
        except KeyError:
            exceptions.raise_with_cause(exceptions.NotFound,
                                        "Unknown atom name '%s'" % atom_name)
        else:
            # TODO(harlowja): we need to figure out how to get away from doing
            # these kinds of type checks in general (since they likely mean
            # we aren't doing something right).
            if expected_type and not isinstance(ad, expected_type):
                raise TypeError("Atom '%s' is not of the expected type: %s"
                                % (atom_name,
                                   reflection.get_class_name(expected_type)))
            if clone:
                return (ad, ad.copy())
            else:
                return (ad, ad)

    @tenacity.retry(retry=tenacity.retry_if_exception_type(
                    exception_types=exceptions.StorageFailure),
                    stop=tenacity.stop_after_attempt(RETRY_ATTEMPTS),
                    wait=tenacity.wait_fixed(RETRY_WAIT_TIMEOUT))
    def _save_atom_detail(self, conn, original_atom_detail, atom_detail):
        # NOTE(harlowja): we need to update our contained atom detail if
        # the result of the update actually added more (aka another process
        # is also modifying the task detail), since python is by reference
        # and the contained atom detail will reflect the old state if we don't
        # do this update.
        original_atom_detail.update(conn.update_atom_details(atom_detail))
        return original_atom_detail

    @fasteners.read_locked
    def get_atom_uuid(self, atom_name):
        """Gets an atoms uuid given a atoms name."""
        source, _clone = self._atomdetail_by_name(atom_name)
        return source.uuid

    @fasteners.write_locked
    def set_atom_state(self, atom_name, state):
        """Sets an atoms state."""
        source, clone = self._atomdetail_by_name(atom_name, clone=True)
        if source.state != state:
            clone.state = state
            self._with_connection(self._save_atom_detail, source, clone)

    @fasteners.read_locked
    def get_atom_state(self, atom_name):
        """Gets the state of an atom given an atoms name."""
        source, _clone = self._atomdetail_by_name(atom_name)
        return source.state

    @fasteners.write_locked
    def set_atom_intention(self, atom_name, intention):
        """Sets the intention of an atom given an atoms name."""
        source, clone = self._atomdetail_by_name(atom_name, clone=True)
        if source.intention != intention:
            clone.intention = intention
            self._with_connection(self._save_atom_detail, source, clone)

    @fasteners.read_locked
    def get_atom_intention(self, atom_name):
        """Gets the intention of an atom given an atoms name."""
        source, _clone = self._atomdetail_by_name(atom_name)
        return source.intention

    @fasteners.read_locked
    def get_atoms_states(self, atom_names):
        """Gets a dict of atom name => (state, intention) given atom names."""
        details = {}
        for name in set(atom_names):
            source, _clone = self._atomdetail_by_name(name)
            details[name] = (source.state, source.intention)
        return details

    @fasteners.write_locked
    def _update_atom_metadata(self, atom_name, update_with,
                              expected_type=None):
        source, clone = self._atomdetail_by_name(atom_name,
                                                 expected_type=expected_type,
                                                 clone=True)
        if update_with:
            clone.meta.update(update_with)
            self._with_connection(self._save_atom_detail, source, clone)

    def update_atom_metadata(self, atom_name, update_with):
        """Updates a atoms associated metadata.

        This update will take a provided dictionary or a list of (key, value)
        pairs to include in the updated metadata (newer keys will overwrite
        older keys) and after merging saves the updated data into the
        underlying persistence layer.
        """
        self._update_atom_metadata(atom_name, update_with)

    def set_task_progress(self, task_name, progress, details=None):
        """Set a tasks progress.

        :param task_name: task name
        :param progress: tasks progress (0.0 <-> 1.0)
        :param details: any task specific progress details
        """
        update_with = {
            META_PROGRESS: progress,
        }
        if details is not None:
            # NOTE(imelnikov): as we can update progress without
            # updating details (e.g. automatically from engine)
            # we save progress value with details, too.
            if details:
                update_with[META_PROGRESS_DETAILS] = {
                    'at_progress': progress,
                    'details': details,
                }
            else:
                update_with[META_PROGRESS_DETAILS] = None
        self._update_atom_metadata(task_name, update_with,
                                   expected_type=models.TaskDetail)

    @fasteners.read_locked
    def get_task_progress(self, task_name):
        """Get the progress of a task given a tasks name.

        :param task_name: tasks name
        :returns: current task progress value
        """
        source, _clone = self._atomdetail_by_name(
            task_name, expected_type=models.TaskDetail)
        try:
            return source.meta[META_PROGRESS]
        except KeyError:
            return 0.0

    @fasteners.read_locked
    def get_task_progress_details(self, task_name):
        """Get the progress details of a task given a tasks name.

        :param task_name: task name
        :returns: None if progress_details not defined, else progress_details
                 dict
        """
        source, _clone = self._atomdetail_by_name(
            task_name, expected_type=models.TaskDetail)
        try:
            return source.meta[META_PROGRESS_DETAILS]
        except KeyError:
            return None

    def _check_all_results_provided(self, atom_name, container):
        """Warn if an atom did not provide some of its expected results.

        This may happen if atom returns shorter tuple or list or dict
        without all needed keys. It may also happen if atom returns
        result of wrong type.
        """
        result_mapping = self._result_mappings.get(atom_name)
        if not result_mapping:
            return
        for name, index in six.iteritems(result_mapping):
            try:
                _item_from(container, index)
            except _EXTRACTION_EXCEPTIONS:
                LOG.warning("Atom '%s' did not supply result "
                            "with index %r (name '%s')", atom_name, index,
                            name)

    @fasteners.write_locked
    def save(self, atom_name, result, state=states.SUCCESS):
        """Put result for atom with provided name to storage."""
        source, clone = self._atomdetail_by_name(atom_name, clone=True)
        if clone.put(state, result):
            self._with_connection(self._save_atom_detail, source, clone)
        # We need to somehow place more of this responsibility on the atom
        # detail class itself, vs doing it here; since it ties those two
        # together (which is bad)...
        if state in (states.FAILURE, states.REVERT_FAILURE):
            # NOTE(imelnikov): failure serialization looses information,
            # so we cache failures here, in atom name -> failure mapping so
            # that we can later use the better version on fetch/get.
            if clone.intention in _SAVE_FAILURE_INTENTIONS:
                fail_cache = self._failures[clone.name]
                fail_cache[clone.intention] = result
        if state == states.SUCCESS and clone.intention == states.EXECUTE:
            self._check_all_results_provided(clone.name, result)

    @fasteners.write_locked
    def save_retry_failure(self, retry_name, failed_atom_name, failure):
        """Save subflow failure to retry controller history."""
        source, clone = self._atomdetail_by_name(
            retry_name, expected_type=models.RetryDetail, clone=True)
        try:
            failures = clone.last_failures
        except exceptions.NotFound:
            exceptions.raise_with_cause(exceptions.StorageFailure,
                                        "Unable to fetch most recent retry"
                                        " failures so new retry failure can"
                                        " be inserted")
        else:
            if failed_atom_name not in failures:
                failures[failed_atom_name] = failure
                self._with_connection(self._save_atom_detail, source, clone)

    @fasteners.write_locked
    def cleanup_retry_history(self, retry_name, state):
        """Cleanup history of retry atom with given name."""
        source, clone = self._atomdetail_by_name(
            retry_name, expected_type=models.RetryDetail, clone=True)
        clone.state = state
        clone.results = []
        self._with_connection(self._save_atom_detail, source, clone)

    @fasteners.read_locked
    def _get(self, atom_name,
             results_attr_name, fail_attr_name,
             allowed_states, fail_cache_key):
        source, _clone = self._atomdetail_by_name(atom_name)
        failure = getattr(source, fail_attr_name)
        if failure is not None:
            fail_cache = self._failures[atom_name]
            try:
                fail = fail_cache[fail_cache_key]
                if failure.matches(fail):
                    # Try to give the version back that should have the
                    # backtrace instead of one that has it
                    # stripped (since backtraces are not serializable).
                    failure = fail
            except KeyError:
                pass
            return failure
        else:
            if source.state not in allowed_states:
                raise exceptions.DisallowedAccess(
                    "Result for atom '%s' is not known/accessible"
                    " due to it being in %s state when result access"
                    " is restricted to %s states" % (atom_name,
                                                     source.state,
                                                     allowed_states),
                    state=source.state)
            return getattr(source, results_attr_name)

    def get_execute_result(self, atom_name):
        """Gets the ``execute`` results for an atom from storage."""
        try:
            results = self._get(atom_name, 'results', 'failure',
                                _EXECUTE_STATES_WITH_RESULTS, states.EXECUTE)
        except exceptions.DisallowedAccess as e:
            if e.state == states.IGNORE:
                exceptions.raise_with_cause(exceptions.NotFound,
                                            "Result for atom '%s' execution"
                                            " is not known (as it was"
                                            " ignored)" % atom_name)
            else:
                exceptions.raise_with_cause(exceptions.NotFound,
                                            "Result for atom '%s' execution"
                                            " is not known" % atom_name)
        else:
            return results

    @fasteners.read_locked
    def _get_failures(self, fail_cache_key):
        failures = {}
        for atom_name, fail_cache in six.iteritems(self._failures):
            try:
                failures[atom_name] = fail_cache[fail_cache_key]
            except KeyError:
                pass
        return failures

    def get_execute_failures(self):
        """Get all ``execute`` failures that happened with this flow."""
        return self._get_failures(states.EXECUTE)

    # TODO(harlowja): remove these in the future?
    get = get_execute_result
    get_failures = get_execute_failures

    def get_revert_result(self, atom_name):
        """Gets the ``revert`` results for an atom from storage."""
        try:
            results = self._get(atom_name, 'revert_results', 'revert_failure',
                                _REVERT_STATES_WITH_RESULTS, states.REVERT)
        except exceptions.DisallowedAccess as e:
            if e.state == states.IGNORE:
                exceptions.raise_with_cause(exceptions.NotFound,
                                            "Result for atom '%s' revert is"
                                            " not known (as it was"
                                            " ignored)" % atom_name)
            else:
                exceptions.raise_with_cause(exceptions.NotFound,
                                            "Result for atom '%s' revert is"
                                            " not known" % atom_name)
        else:
            return results

    def get_revert_failures(self):
        """Get all ``revert`` failures that happened with this flow."""
        return self._get_failures(states.REVERT)

    @fasteners.read_locked
    def has_failures(self):
        """Returns true if there are **any** failures in storage."""
        for fail_cache in six.itervalues(self._failures):
            if fail_cache:
                return True
        return False

    @fasteners.write_locked
    def reset(self, atom_name, state=states.PENDING):
        """Reset atom with given name (if the atom is not in a given state)."""
        if atom_name == self.injector_name:
            return
        source, clone = self._atomdetail_by_name(atom_name, clone=True)
        if source.state == state:
            return
        clone.reset(state)
        self._with_connection(self._save_atom_detail, source, clone)
        self._failures[clone.name].clear()

    def inject_atom_args(self, atom_name, pairs, transient=True):
        """Add values into storage for a specific atom only.

        :param transient: save the data in-memory only instead of persisting
                the data to backend storage (useful for resource-like objects
                or similar objects which can **not** be persisted)

        This method injects a dictionary/pairs of arguments for an atom so that
        when that atom is scheduled for execution it will have immediate access
        to these arguments.

        .. note::

            Injected atom arguments take precedence over arguments
            provided by predecessor atoms or arguments provided by injecting
            into the flow scope (using
            the :py:meth:`~taskflow.storage.Storage.inject` method).

        .. warning::

            It should be noted that injected atom arguments (that are scoped
            to the atom with the given name) *should* be serializable
            whenever possible. This is a **requirement** for the
            :doc:`worker based engine <workers>` which **must**
            serialize (typically using ``json``) all
            atom :py:meth:`~taskflow.atom.Atom.execute` and
            :py:meth:`~taskflow.atom.Atom.revert` arguments to
            be able to transmit those arguments to the target worker(s). If
            the use-case being applied/desired is to later use the worker
            based engine then it is highly recommended to ensure all injected
            atoms (even transient ones) are serializable to avoid issues
            that *may* appear later (when a object turned out to not actually
            be serializable).
        """
        if atom_name not in self._atom_name_to_uuid:
            raise exceptions.NotFound("Unknown atom name '%s'" % atom_name)

        def save_transient():
            self._injected_args.setdefault(atom_name, {})
            self._injected_args[atom_name].update(pairs)

        def save_persistent():
            source, clone = self._atomdetail_by_name(atom_name, clone=True)
            injected = source.meta.get(META_INJECTED)
            if not injected:
                injected = {}
            injected.update(pairs)
            clone.meta[META_INJECTED] = injected
            self._with_connection(self._save_atom_detail, source, clone)

        with self._lock.write_lock():
            if transient:
                save_transient()
            else:
                save_persistent()

    @fasteners.write_locked
    def inject(self, pairs, transient=False):
        """Add values into storage.

        This method should be used to put flow parameters (requirements that
        are not satisfied by any atom in the flow) into storage.

        :param transient: save the data in-memory only instead of persisting
                the data to backend storage (useful for resource-like objects
                or similar objects which can **not** be persisted)

        .. warning::

            It should be noted that injected flow arguments (that are scoped
            to all atoms in this flow) *should* be serializable whenever
            possible. This is a **requirement** for
            the :doc:`worker based engine <workers>` which **must**
            serialize (typically using ``json``) all
            atom :py:meth:`~taskflow.atom.Atom.execute` and
            :py:meth:`~taskflow.atom.Atom.revert` arguments to
            be able to transmit those arguments to the target worker(s). If
            the use-case being applied/desired is to later use the worker
            based engine then it is highly recommended to ensure all injected
            atoms (even transient ones) are serializable to avoid issues
            that *may* appear later (when a object turned out to not actually
            be serializable).
        """

        def save_persistent():
            try:
                source, clone = self._atomdetail_by_name(
                    self.injector_name,
                    expected_type=models.TaskDetail,
                    clone=True)
            except exceptions.NotFound:
                # Ensure we have our special task detail...
                #
                # TODO(harlowja): get this removed when
                # https://review.openstack.org/#/c/165645/ merges.
                source = self._create_atom_detail(self.injector_name,
                                                  models.TaskDetail,
                                                  atom_state=None)
                fd_source, fd_clone = self._fetch_flowdetail(clone=True)
                fd_clone.add(source)
                self._with_connection(self._save_flow_detail, fd_source,
                                      fd_clone)
                self._atom_name_to_uuid[source.name] = source.uuid
                clone = source
                clone.results = dict(pairs)
                clone.state = states.SUCCESS
            else:
                clone.results.update(pairs)
            result = self._with_connection(self._save_atom_detail,
                                           source, clone)
            return (self.injector_name, six.iterkeys(result.results))

        def save_transient():
            self._transients.update(pairs)
            return (_TRANSIENT_PROVIDER, six.iterkeys(self._transients))

        if transient:
            provider_name, names = save_transient()
        else:
            provider_name, names = save_persistent()

        self._set_result_mapping(provider_name,
                                 dict((name, name) for name in names))

    def _fetch_providers(self, looking_for, providers=None):
        """Return pair of (default providers, atom providers)."""
        if providers is None:
            providers = self._reverse_mapping.get(looking_for, [])
        default_providers = []
        atom_providers = []
        for p in providers:
            if p.name in (_TRANSIENT_PROVIDER, self.injector_name):
                default_providers.append(p)
            else:
                atom_providers.append(p)
        return default_providers, atom_providers

    def _set_result_mapping(self, provider_name, mapping):
        """Sets the result mapping for a given producer.

        The result saved with given name would be accessible by names
        defined in mapping. Mapping is a dict name => index. If index
        is None, the whole result will have this name; else, only
        part of it, result[index].
        """
        provider_mapping = self._result_mappings.setdefault(provider_name, {})
        if mapping:
            provider_mapping.update(mapping)
            # Ensure the reverse mapping/index is updated (for faster lookups).
            for name, index in six.iteritems(provider_mapping):
                entries = self._reverse_mapping.setdefault(name, [])
                provider = _Provider(provider_name, index)
                if provider not in entries:
                    entries.append(provider)

    @fasteners.read_locked
    def fetch(self, name, many_handler=None):
        """Fetch a named ``execute`` result."""
        def _many_handler(values):
            # By default we just return the first of many (unless provided
            # a different callback that can translate many results into
            # something more meaningful).
            return values[0]
        if many_handler is None:
            many_handler = _many_handler
        try:
            maybe_providers = self._reverse_mapping[name]
        except KeyError:
            raise exceptions.NotFound("Name %r is not mapped as a produced"
                                      " output by any providers" % name)
        locator = _ProviderLocator(
            self._transients,
            functools.partial(self._fetch_providers,
                              providers=maybe_providers),
            lambda atom_name:
                self._get(atom_name, 'last_results', 'failure',
                          _EXECUTE_STATES_WITH_RESULTS, states.EXECUTE))
        values = []
        searched_providers, providers = locator.find(
            name, short_circuit=False,
            # NOTE(harlowja): There are no scopes used here (as of now), so
            # we just return all known providers as if it was one large
            # scope.
            scope_walker=[[p.name for p in maybe_providers]])
        for provider, results in providers:
            values.append(_item_from_single(provider, results, name))
        if not values:
            raise exceptions.NotFound(
                "Unable to find result %r, searched %s providers"
                % (name, len(searched_providers)))
        else:
            return many_handler(values)

    @fasteners.read_locked
    def fetch_unsatisfied_args(self, atom_name, args_mapping,
                               scope_walker=None, optional_args=None):
        """Fetch unsatisfied ``execute`` arguments using an atoms args mapping.

        NOTE(harlowja): this takes into account the provided scope walker
        atoms who should produce the required value at runtime, as well as
        the transient/persistent flow and atom specific injected arguments.
        It does **not** check if the providers actually have produced the
        needed values; it just checks that they are registered to produce
        it in the future.
        """
        source, _clone = self._atomdetail_by_name(atom_name)
        if scope_walker is None:
            scope_walker = self._scope_fetcher(atom_name)
        if optional_args is None:
            optional_args = []
        injected_sources = [
            self._injected_args.get(atom_name, {}),
            source.meta.get(META_INJECTED, {}),
        ]
        missing = set(six.iterkeys(args_mapping))
        locator = _ProviderLocator(
            self._transients, self._fetch_providers,
            lambda atom_name:
                self._get(atom_name, 'last_results', 'failure',
                          _EXECUTE_STATES_WITH_RESULTS, states.EXECUTE))
        for (bound_name, name) in six.iteritems(args_mapping):
            if LOG.isEnabledFor(logging.TRACE):
                LOG.trace("Looking for %r <= %r for atom '%s'",
                          bound_name, name, atom_name)
            if bound_name in optional_args:
                LOG.trace("Argument %r is optional, skipping", bound_name)
                missing.discard(bound_name)
                continue
            maybe_providers = 0
            for source in injected_sources:
                if not source:
                    continue
                if name in source:
                    maybe_providers += 1
            maybe_providers += len(
                locator.find_potentials(name, scope_walker=scope_walker))
            if maybe_providers:
                LOG.trace("Atom '%s' will have %s potential providers"
                          " of %r <= %r", atom_name, maybe_providers,
                          bound_name, name)
                missing.discard(bound_name)
        return missing

    @fasteners.read_locked
    def fetch_all(self, many_handler=None):
        """Fetch all named ``execute`` results known so far."""
        def _many_handler(values):
            if len(values) > 1:
                return values
            return values[0]
        if many_handler is None:
            many_handler = _many_handler
        results = {}
        for name in six.iterkeys(self._reverse_mapping):
            try:
                results[name] = self.fetch(name, many_handler=many_handler)
            except exceptions.NotFound:
                pass
        return results

    @fasteners.read_locked
    def fetch_mapped_args(self, args_mapping,
                          atom_name=None, scope_walker=None,
                          optional_args=None):
        """Fetch ``execute`` arguments for an atom using its args mapping."""
        def _extract_first_from(name, sources):
            """Extracts/returns first occurrence of key in list of dicts."""
            for i, source in enumerate(sources):
                if not source:
                    continue
                if name in source:
                    return (i, source[name])
            raise KeyError(name)
        if optional_args is None:
            optional_args = []
        if atom_name:
            source, _clone = self._atomdetail_by_name(atom_name)
            injected_sources = [
                self._injected_args.get(atom_name, {}),
                source.meta.get(META_INJECTED, {}),
            ]
            if scope_walker is None:
                scope_walker = self._scope_fetcher(atom_name)
        else:
            injected_sources = []
        if not args_mapping:
            return {}
        get_results = lambda atom_name: \
            self._get(atom_name, 'last_results', 'failure',
                      _EXECUTE_STATES_WITH_RESULTS, states.EXECUTE)
        mapped_args = {}
        for (bound_name, name) in six.iteritems(args_mapping):
            if LOG.isEnabledFor(logging.TRACE):
                if atom_name:
                    LOG.trace("Looking for %r <= %r for atom '%s'",
                              bound_name, name, atom_name)
                else:
                    LOG.trace("Looking for %r <= %r", bound_name, name)
            try:
                source_index, value = _extract_first_from(
                    name, injected_sources)
                mapped_args[bound_name] = value
                if LOG.isEnabledFor(logging.TRACE):
                    if source_index == 0:
                        LOG.trace("Matched %r <= %r to %r (from injected"
                                  " atom-specific transient"
                                  " values)", bound_name, name, value)
                    else:
                        LOG.trace("Matched %r <= %r to %r (from injected"
                                  " atom-specific persistent"
                                  " values)", bound_name, name, value)
            except KeyError:
                try:
                    maybe_providers = self._reverse_mapping[name]
                except KeyError:
                    if bound_name in optional_args:
                        LOG.trace("Argument %r is optional, skipping",
                                  bound_name)
                        continue
                    raise exceptions.NotFound("Name %r is not mapped as a"
                                              " produced output by any"
                                              " providers" % name)
                locator = _ProviderLocator(
                    self._transients,
                    functools.partial(self._fetch_providers,
                                      providers=maybe_providers), get_results)
                searched_providers, providers = locator.find(
                    name, scope_walker=scope_walker)
                if not providers:
                    raise exceptions.NotFound(
                        "Mapped argument %r <= %r was not produced"
                        " by any accessible provider (%s possible"
                        " providers were scanned)"
                        % (bound_name, name, len(searched_providers)))
                provider, value = _item_from_first_of(providers, name)
                mapped_args[bound_name] = value
                LOG.trace("Matched %r <= %r to %r (from %s)",
                          bound_name, name, value, provider)
        return mapped_args

    @fasteners.write_locked
    def set_flow_state(self, state):
        """Set flow details state and save it."""
        source, clone = self._fetch_flowdetail(clone=True)
        clone.state = state
        self._with_connection(self._save_flow_detail, source, clone)

    @fasteners.write_locked
    def update_flow_metadata(self, update_with):
        """Update flowdetails metadata and save it."""
        if update_with:
            source, clone = self._fetch_flowdetail(clone=True)
            clone.meta.update(update_with)
            self._with_connection(self._save_flow_detail, source, clone)

    @fasteners.write_locked
    def change_flow_state(self, state):
        """Transition flow from old state to new state.

        Returns ``(True, old_state)`` if transition was performed,
        or ``(False, old_state)`` if it was ignored, or raises a
        :py:class:`~taskflow.exceptions.InvalidState` exception if transition
        is invalid.
        """
        old_state = self.get_flow_state()
        if not states.check_flow_transition(old_state, state):
            return (False, old_state)
        self.set_flow_state(state)
        return (True, old_state)

    @fasteners.read_locked
    def get_flow_state(self):
        """Get state from flow details."""
        source = self._flowdetail
        state = source.state
        if state is None:
            state = states.PENDING
        return state

    def _translate_into_history(self, ad):
        failure = None
        if ad.failure is not None:
            # NOTE(harlowja): Try to use our local cache to get a more
            # complete failure object that has a traceback (instead of the
            # one that is saved which will *typically* not have one)...
            failure = ad.failure
            fail_cache = self._failures[ad.name]
            try:
                fail = fail_cache[states.EXECUTE]
                if failure.matches(fail):
                    failure = fail
            except KeyError:
                pass
        return retry.History(ad.results, failure=failure)

    @fasteners.read_locked
    def get_retry_history(self, retry_name):
        """Fetch a single retrys history."""
        source, _clone = self._atomdetail_by_name(
            retry_name, expected_type=models.RetryDetail)
        return self._translate_into_history(source)

    @fasteners.read_locked
    def get_retry_histories(self):
        """Fetch all retrys histories."""
        histories = []
        for ad in self._flowdetail:
            if isinstance(ad, models.RetryDetail):
                histories.append((ad.name,
                                  self._translate_into_history(ad)))
        return histories