# -*- 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 collections import contextlib import functools import threading from concurrent import futures from kazoo import exceptions as k_exceptions from kazoo.protocol import paths as k_paths from kazoo.recipe import watchers from oslo_serialization import jsonutils from oslo_utils import excutils from oslo_utils import uuidutils import six from taskflow import exceptions as excp from taskflow.jobs import base from taskflow import logging from taskflow import states from taskflow.types import timing as tt from taskflow.utils import kazoo_utils from taskflow.utils import lock_utils from taskflow.utils import misc LOG = logging.getLogger(__name__) UNCLAIMED_JOB_STATES = ( states.UNCLAIMED, ) ALL_JOB_STATES = ( states.UNCLAIMED, states.COMPLETE, states.CLAIMED, ) # Transaction support was added in 3.4.0 MIN_ZK_VERSION = (3, 4, 0) LOCK_POSTFIX = ".lock" TRASH_FOLDER = ".trash" JOB_PREFIX = 'job' def _check_who(who): if not isinstance(who, six.string_types): raise TypeError("Job applicant must be a string type") if len(who) == 0: raise ValueError("Job applicant must be non-empty") class ZookeeperJob(base.Job): """A zookeeper job.""" def __init__(self, name, board, client, backend, path, uuid=None, details=None, book=None, book_data=None, created_on=None): super(ZookeeperJob, self).__init__(name, uuid=uuid, details=details) self._board = board self._book = book if not book_data: book_data = {} self._book_data = book_data self._client = client self._backend = backend if all((self._book, self._book_data)): raise ValueError("Only one of 'book_data' or 'book'" " can be provided") self._path = k_paths.normpath(path) self._lock_path = self._path + LOCK_POSTFIX self._created_on = created_on self._node_not_found = False basename = k_paths.basename(self._path) self._root = self._path[0:-len(basename)] self._sequence = int(basename[len(JOB_PREFIX):]) @property def lock_path(self): return self._lock_path @property def path(self): return self._path @property def sequence(self): """Sequence number of the current job.""" return self._sequence @property def root(self): """The parent path of the job in zookeeper.""" return self._root def _get_node_attr(self, path, attr_name, trans_func=None): try: _data, node_stat = self._client.get(path) attr = getattr(node_stat, attr_name) if trans_func is not None: return trans_func(attr) else: return attr except k_exceptions.NoNodeError as e: raise excp.NotFound("Can not fetch the %r attribute" " of job %s (%s), path %s not found" % (attr_name, self.uuid, self.path, path), e) except self._client.handler.timeout_exception as e: raise excp.JobFailure("Can not fetch the %r attribute" " of job %s (%s), operation timed out" % (attr_name, self.uuid, self.path), e) except k_exceptions.SessionExpiredError as e: raise excp.JobFailure("Can not fetch the %r attribute" " of job %s (%s), session expired" % (attr_name, self.uuid, self.path), e) except (AttributeError, k_exceptions.KazooException) as e: raise excp.JobFailure("Can not fetch the %r attribute" " of job %s (%s), internal error" % (attr_name, self.uuid, self.path), e) @property def last_modified(self): modified_on = None try: if not self._node_not_found: modified_on = self._get_node_attr( self.path, 'mtime', trans_func=misc.millis_to_datetime) except excp.NotFound: self._node_not_found = True return modified_on @property def created_on(self): # This one we can cache (since it won't change after creation). if self._node_not_found: return None if self._created_on is None: try: self._created_on = self._get_node_attr( self.path, 'ctime', trans_func=misc.millis_to_datetime) except excp.NotFound: self._node_not_found = True return self._created_on @property def board(self): return self._board def _load_book(self): book_uuid = self.book_uuid if self._backend is not None and book_uuid is not None: # TODO(harlowja): we are currently limited by assuming that the # job posted has the same backend as this loader (to start this # seems to be a ok assumption, and can be adjusted in the future # if we determine there is a use-case for multi-backend loaders, # aka a registry of loaders). with contextlib.closing(self._backend.get_connection()) as conn: return conn.get_logbook(book_uuid) # No backend to fetch from or no uuid specified return None @property def state(self): owner = self.board.find_owner(self) job_data = {} try: raw_data, _data_stat = self._client.get(self.path) job_data = misc.decode_json(raw_data) except k_exceptions.NoNodeError: pass except k_exceptions.SessionExpiredError as e: raise excp.JobFailure("Can not fetch the state of %s," " session expired" % (self.uuid), e) except self._client.handler.timeout_exception as e: raise excp.JobFailure("Can not fetch the state of %s," " operation timed out" % (self.uuid), e) except k_exceptions.KazooException as e: raise excp.JobFailure("Can not fetch the state of %s, internal" " error" % (self.uuid), e) if not job_data: # No data this job has been completed (the owner that we might have # fetched will not be able to be fetched again, since the job node # is a parent node of the owner/lock node). return states.COMPLETE if not owner: # No owner, but data, still work to be done. return states.UNCLAIMED return states.CLAIMED def __lt__(self, other): if self.root == other.root: return self.sequence < other.sequence else: return self.root < other.root def __hash__(self): return hash(self.path) @property def book(self): if self._book is None: self._book = self._load_book() return self._book @property def book_uuid(self): if self._book: return self._book.uuid if self._book_data: return self._book_data.get('uuid') else: return None @property def book_name(self): if self._book: return self._book.name if self._book_data: return self._book_data.get('name') else: return None class ZookeeperJobBoardIterator(six.Iterator): """Iterator over a zookeeper jobboard that iterates over potential jobs. It supports the following attributes/constructor arguments: * ``ensure_fresh``: boolean that requests that during every fetch of a new set of jobs this will cause the iterator to force the backend to refresh (ensuring that the jobboard has the most recent job listings). * ``only_unclaimed``: boolean that indicates whether to only iterate over unclaimed jobs. """ def __init__(self, board, only_unclaimed=False, ensure_fresh=False): self._board = board self._jobs = collections.deque() self._fetched = False self.ensure_fresh = ensure_fresh self.only_unclaimed = only_unclaimed @property def board(self): return self._board def __iter__(self): return self def _next_job(self): if self.only_unclaimed: allowed_states = UNCLAIMED_JOB_STATES else: allowed_states = ALL_JOB_STATES job = None while self._jobs and job is None: maybe_job = self._jobs.popleft() try: if maybe_job.state in allowed_states: job = maybe_job except excp.JobFailure: LOG.warn("Failed determining the state of job: %s (%s)", maybe_job.uuid, maybe_job.path, exc_info=True) except excp.NotFound: self._board._remove_job(maybe_job.path) return job def __next__(self): if not self._jobs: if not self._fetched: jobs = self._board._fetch_jobs(ensure_fresh=self.ensure_fresh) self._jobs.extend(jobs) self._fetched = True job = self._next_job() if job is None: raise StopIteration else: return job class ZookeeperJobBoard(base.NotifyingJobBoard): """A jobboard backend by zookeeper. Powered by the `kazoo `_ library. This jobboard creates *sequenced* persistent znodes in a directory in zookeeper (that directory defaults ``/taskflow/jobs``) and uses zookeeper watches to notify other jobboards that the job which was posted using the :meth:`.post` method (this creates a znode with contents/details in json) The users of those jobboard(s) (potentially on disjoint sets of machines) can then iterate over the available jobs and decide if they want to attempt to claim one of the jobs they have iterated over. If so they will then attempt to contact zookeeper and will attempt to create a ephemeral znode using the name of the persistent znode + ".lock" as a postfix. If the entity trying to use the jobboard to :meth:`.claim` the job is able to create a ephemeral znode with that name then it will be allowed (and expected) to perform whatever *work* the contents of that job that it locked described. Once finished the ephemeral znode and persistent znode may be deleted (if successfully completed) in a single transcation or if not successfull (or the entity that claimed the znode dies) the ephemeral znode will be released (either manually by using :meth:`.abandon` or automatically by zookeeper the ephemeral is deemed to be lost). """ def __init__(self, name, conf, client=None, persistence=None, emit_notifications=True): super(ZookeeperJobBoard, self).__init__(name, conf) if client is not None: self._client = client self._owned = False else: self._client = kazoo_utils.make_client(self._conf) self._owned = True path = str(conf.get("path", "/taskflow/jobs")) if not path: raise ValueError("Empty zookeeper path is disallowed") if not k_paths.isabs(path): raise ValueError("Zookeeper path must be absolute") self._path = path self._trash_path = self._path.replace(k_paths.basename(self._path), TRASH_FOLDER) # The backend to load the full logbooks from, since whats sent over # the zookeeper data connection is only the logbook uuid and name, and # not currently the full logbook (later when a zookeeper backend # appears we can likely optimize for that backend usage by directly # reading from the path where the data is stored, if we want). self._persistence = persistence # Misc. internal details self._known_jobs = {} self._job_cond = threading.Condition() self._open_close_lock = threading.RLock() self._client.add_listener(self._state_change_listener) self._bad_paths = frozenset([path]) self._job_watcher = None # Since we use sequenced ids this will be the path that the sequences # are prefixed with, for example, job0000000001, job0000000002, ... self._job_base = k_paths.join(path, JOB_PREFIX) self._worker = None self._emit_notifications = bool(emit_notifications) self._connected = False def _emit(self, state, details): # Submit the work to the executor to avoid blocking the kazoo queue. try: self._worker.submit(self.notifier.notify, state, details) except (AttributeError, RuntimeError): # Notification thread is shutdown or non-existent, either case we # just want to skip submitting a notification... pass @property def path(self): return self._path @property def trash_path(self): return self._trash_path @property def job_count(self): return len(self._known_jobs) def _fetch_jobs(self, ensure_fresh=False): if ensure_fresh: self._force_refresh() with self._job_cond: return sorted(six.itervalues(self._known_jobs)) def _force_refresh(self): try: children = self._client.get_children(self.path) except self._client.handler.timeout_exception as e: raise excp.JobFailure("Refreshing failure, operation timed out", e) except k_exceptions.SessionExpiredError as e: raise excp.JobFailure("Refreshing failure, session expired", e) except k_exceptions.NoNodeError: pass except k_exceptions.KazooException as e: raise excp.JobFailure("Refreshing failure, internal error", e) else: self._on_job_posting(children, delayed=False) def iterjobs(self, only_unclaimed=False, ensure_fresh=False): return ZookeeperJobBoardIterator(self, only_unclaimed=only_unclaimed, ensure_fresh=ensure_fresh) def _remove_job(self, path): if path not in self._known_jobs: return with self._job_cond: job = self._known_jobs.pop(path, None) if job is not None: LOG.debug("Removed job that was at path '%s'", path) self._emit(base.REMOVAL, details={'job': job}) def _process_child(self, path, request): """Receives the result of a child data fetch request.""" job = None try: raw_data, node_stat = request.get() job_data = misc.decode_json(raw_data) created_on = misc.millis_to_datetime(node_stat.ctime) except (ValueError, TypeError, KeyError): LOG.warn("Incorrectly formatted job data found at path: %s", path, exc_info=True) except self._client.handler.timeout_exception: LOG.warn("Operation timed out fetching job data from path: %s", path, exc_info=True) except k_exceptions.SessionExpiredError: LOG.warn("Session expired fetching job data from path: %s", path, exc_info=True) except k_exceptions.NoNodeError: LOG.debug("No job node found at path: %s, it must have" " disappeared or was removed", path) except k_exceptions.KazooException: LOG.warn("Internal error fetching job data from path: %s", path, exc_info=True) else: with self._job_cond: # Now we can offically check if someone already placed this # jobs information into the known job set (if it's already # existing then just leave it alone). if path not in self._known_jobs: job = ZookeeperJob(job_data['name'], self, self._client, self._persistence, path, uuid=job_data['uuid'], book_data=job_data.get("book"), details=job_data.get("details", {}), created_on=created_on) self._known_jobs[path] = job self._job_cond.notify_all() if job is not None: self._emit(base.POSTED, details={'job': job}) def _on_job_posting(self, children, delayed=True): LOG.debug("Got children %s under path %s", children, self.path) child_paths = [] for c in children: if c.endswith(LOCK_POSTFIX) or not c.startswith(JOB_PREFIX): # Skip lock paths or non-job-paths (these are not valid jobs) continue child_paths.append(k_paths.join(self.path, c)) # Figure out what we really should be investigating and what we # shouldn't (remove jobs that exist in our local version, but don't # exist in the children anymore) and accumulate all paths that we # need to trigger population of (without holding the job lock). investigate_paths = [] pending_removals = [] with self._job_cond: for path in six.iterkeys(self._known_jobs): if path not in child_paths: pending_removals.append(path) for path in child_paths: if path in self._bad_paths: continue # This pre-check will *not* guarantee that we will not already # have the job (if it's being populated elsewhere) but it will # reduce the amount of duplicated requests in general; later when # the job information has been populated we will ensure that we # are not adding duplicates into the currently known jobs... if path in self._known_jobs: continue if path not in investigate_paths: investigate_paths.append(path) if pending_removals: with self._job_cond: for path in pending_removals: self._remove_job(path) for path in investigate_paths: # Fire off the request to populate this job. # # This method is *usually* called from a asynchronous handler so # it's better to exit from this quickly to allow other asynchronous # handlers to be executed. request = self._client.get_async(path) if delayed: request.rawlink(functools.partial(self._process_child, path)) else: self._process_child(path, request) def post(self, name, book=None, details=None): def format_posting(job_uuid): posting = { 'uuid': job_uuid, 'name': name, } if details: posting['details'] = details else: posting['details'] = {} if book is not None: posting['book'] = { 'name': book.name, 'uuid': book.uuid, } return posting # NOTE(harlowja): Jobs are not ephemeral, they will persist until they # are consumed (this may change later, but seems safer to do this until # further notice). job_uuid = uuidutils.generate_uuid() with self._wrap(job_uuid, None, "Posting failure: %s", ensure_known=False): job_posting = format_posting(job_uuid) job_posting = misc.binary_encode(jsonutils.dumps(job_posting)) job_path = self._client.create(self._job_base, value=job_posting, sequence=True, ephemeral=False) job = ZookeeperJob(name, self, self._client, self._persistence, job_path, book=book, details=details, uuid=job_uuid) with self._job_cond: self._known_jobs[job_path] = job self._job_cond.notify_all() self._emit(base.POSTED, details={'job': job}) return job def claim(self, job, who): def _unclaimable_try_find_owner(cause): try: owner = self.find_owner(job) except Exception: owner = None if owner: msg = "Job %s already claimed by '%s'" % (job.uuid, owner) else: msg = "Job %s already claimed" % (job.uuid) return excp.UnclaimableJob(msg, cause) _check_who(who) with self._wrap(job.uuid, job.path, "Claiming failure: %s"): # NOTE(harlowja): post as json which will allow for future changes # more easily than a raw string/text. value = jsonutils.dumps({ 'owner': who, }) # Ensure the target job is still existent (at the right version). job_data, job_stat = self._client.get(job.path) txn = self._client.transaction() # This will abort (and not create the lock) if the job has been # removed (somehow...) or updated by someone else to a different # version... txn.check(job.path, version=job_stat.version) txn.create(job.lock_path, value=misc.binary_encode(value), ephemeral=True) try: kazoo_utils.checked_commit(txn) except k_exceptions.NodeExistsError as e: raise _unclaimable_try_find_owner(e) except kazoo_utils.KazooTransactionException as e: if len(e.failures) < 2: raise else: if isinstance(e.failures[0], k_exceptions.NoNodeError): raise excp.NotFound( "Job %s not found to be claimed" % job.uuid, e.failures[0]) if isinstance(e.failures[1], k_exceptions.NodeExistsError): raise _unclaimable_try_find_owner(e.failures[1]) else: raise excp.UnclaimableJob( "Job %s claim failed due to transaction" " not succeeding" % (job.uuid), e) @contextlib.contextmanager def _wrap(self, job_uuid, job_path, fail_msg_tpl="Failure: %s", ensure_known=True): if job_path: fail_msg_tpl += " (%s)" % (job_path) if ensure_known: if not job_path: raise ValueError("Unable to check if %r is a known path" % (job_path)) if job_path not in self._known_jobs: fail_msg_tpl += ", unknown job" raise excp.NotFound(fail_msg_tpl % (job_uuid)) try: yield except self._client.handler.timeout_exception as e: fail_msg_tpl += ", operation timed out" raise excp.JobFailure(fail_msg_tpl % (job_uuid), e) except k_exceptions.SessionExpiredError as e: fail_msg_tpl += ", session expired" raise excp.JobFailure(fail_msg_tpl % (job_uuid), e) except k_exceptions.NoNodeError: fail_msg_tpl += ", unknown job" raise excp.NotFound(fail_msg_tpl % (job_uuid)) except k_exceptions.KazooException as e: fail_msg_tpl += ", internal error" raise excp.JobFailure(fail_msg_tpl % (job_uuid), e) def find_owner(self, job): with self._wrap(job.uuid, job.path, "Owner query failure: %s"): try: self._client.sync(job.lock_path) raw_data, _lock_stat = self._client.get(job.lock_path) data = misc.decode_json(raw_data) owner = data.get("owner") except k_exceptions.NoNodeError: owner = None return owner def _get_owner_and_data(self, job): lock_data, lock_stat = self._client.get(job.lock_path) job_data, job_stat = self._client.get(job.path) return (misc.decode_json(lock_data), lock_stat, misc.decode_json(job_data), job_stat) def consume(self, job, who): _check_who(who) with self._wrap(job.uuid, job.path, "Consumption failure: %s"): try: owner_data = self._get_owner_and_data(job) lock_data, lock_stat, data, data_stat = owner_data except k_exceptions.NoNodeError: raise excp.JobFailure("Can not consume a job %s" " which we can not determine" " the owner of" % (job.uuid)) if lock_data.get("owner") != who: raise excp.JobFailure("Can not consume a job %s" " which is not owned by %s" % (job.uuid, who)) txn = self._client.transaction() txn.delete(job.lock_path, version=lock_stat.version) txn.delete(job.path, version=data_stat.version) kazoo_utils.checked_commit(txn) self._remove_job(job.path) def abandon(self, job, who): _check_who(who) with self._wrap(job.uuid, job.path, "Abandonment failure: %s"): try: owner_data = self._get_owner_and_data(job) lock_data, lock_stat, data, data_stat = owner_data except k_exceptions.NoNodeError: raise excp.JobFailure("Can not abandon a job %s" " which we can not determine" " the owner of" % (job.uuid)) if lock_data.get("owner") != who: raise excp.JobFailure("Can not abandon a job %s" " which is not owned by %s" % (job.uuid, who)) txn = self._client.transaction() txn.delete(job.lock_path, version=lock_stat.version) kazoo_utils.checked_commit(txn) def trash(self, job, who): _check_who(who) with self._wrap(job.uuid, job.path, "Trash failure: %s"): try: owner_data = self._get_owner_and_data(job) lock_data, lock_stat, data, data_stat = owner_data except k_exceptions.NoNodeError: raise excp.JobFailure("Can not trash a job %s" " which we can not determine" " the owner of" % (job.uuid)) if lock_data.get("owner") != who: raise excp.JobFailure("Can not trash a job %s" " which is not owned by %s" % (job.uuid, who)) trash_path = job.path.replace(self.path, self.trash_path) value = misc.binary_encode(jsonutils.dumps(data)) txn = self._client.transaction() txn.create(trash_path, value=value) txn.delete(job.lock_path, version=lock_stat.version) txn.delete(job.path, version=data_stat.version) kazoo_utils.checked_commit(txn) def _state_change_listener(self, state): LOG.debug("Kazoo client has changed to state: %s", state) def wait(self, timeout=None): # Wait until timeout expires (or forever) for jobs to appear. watch = tt.StopWatch(duration=timeout) watch.start() with self._job_cond: while True: if not self._known_jobs: if watch.expired(): raise excp.NotFound("Expired waiting for jobs to" " arrive; waited %s seconds" % watch.elapsed()) # This is done since the given timeout can not be provided # to the condition variable, since we can not ensure that # when we acquire the condition that there will actually # be jobs (especially if we are spuriously awaken), so we # must recalculate the amount of time we really have left. self._job_cond.wait(watch.leftover(return_none=True)) else: it = ZookeeperJobBoardIterator(self) it._jobs.extend(self._fetch_jobs()) it._fetched = True return it @property def connected(self): return self._connected and self._client.connected @lock_utils.locked(lock='_open_close_lock') def close(self): if self._owned: LOG.debug("Stopping client") kazoo_utils.finalize_client(self._client) if self._worker is not None: LOG.debug("Shutting down the notifier") self._worker.shutdown() self._worker = None with self._job_cond: self._known_jobs.clear() LOG.debug("Stopped & cleared local state") self._connected = False @lock_utils.locked(lock='_open_close_lock') def connect(self, timeout=10.0): def try_clean(): # Attempt to do the needed cleanup if post-connection setup does # not succeed (maybe the connection is lost right after it is # obtained). try: self.close() except k_exceptions.KazooException: LOG.exception("Failed cleaning-up after post-connection" " initialization failed") try: if timeout is not None: timeout = float(timeout) self._client.start(timeout=timeout) except (self._client.handler.timeout_exception, k_exceptions.KazooException) as e: raise excp.JobFailure("Failed to connect to zookeeper", e) try: if self._conf.get('check_compatible', True): kazoo_utils.check_compatible(self._client, MIN_ZK_VERSION) if self._worker is None and self._emit_notifications: self._worker = futures.ThreadPoolExecutor(max_workers=1) self._client.ensure_path(self.path) self._client.ensure_path(self.trash_path) if self._job_watcher is None: self._job_watcher = watchers.ChildrenWatch( self._client, self.path, func=self._on_job_posting, allow_session_lost=True) self._connected = True except excp.IncompatibleVersion: with excutils.save_and_reraise_exception(): try_clean() except (self._client.handler.timeout_exception, k_exceptions.KazooException) as e: try_clean() raise excp.JobFailure("Failed to do post-connection" " initialization", e)