def setUp(self):
super().setUp()
self.addCleanup(patch.stopall)
self._patched_items = {}
self._blacklist_methods_not_allowed_in_unit_tests()
# Stub out a few library dependencies that launch subprocesses.
self.patch(
'app.util.autoversioning.get_version').return_value = '0.0.0'
self.patch('app.util.conf.base_config_loader.platform.node'
).return_value = self._fake_hostname
if self._do_network_mocks:
# requests.Session() also makes some subprocess calls on instantiation.
self.patch('app.util.network.requests.Session')
# Stub out Network.are_hosts_same() call with a simple string comparison.
self.patch('app.util.network.Network.are_hosts_same',
new=lambda host_a, host_b: host_a == host_b)
# Reset singletons so that they get recreated for every test that uses them.
Configuration.reset_singleton()
UnhandledExceptionHandler.reset_singleton()
SlaveRegistry.reset_singleton()
# Explicitly initialize UnhandledExceptionHandler singleton here (on the main thread) since it sets up signal
# handlers that must execute on the main thread.
UnhandledExceptionHandler.singleton()
MasterConfigLoader().configure_defaults(Configuration.singleton())
MasterConfigLoader().configure_postload(Configuration.singleton())
self.patch(
'app.util.conf.master_config_loader.MasterConfigLoader.load_from_config_file'
)
# Reset counters
Slave._slave_id_counter = Counter()
Build._build_id_counter = Counter()
analytics._event_id_generator = Counter()
# Configure logging to go to stdout. This makes debugging easier by allowing us to see logs for failed tests.
log.configure_logging('DEBUG')
# Then stub out configure_logging so we don't end up logging to real files during testing.
self.patch('app.util.log.configure_logging')
# Set up TestHandler. This allows asserting on log messages in tests.
self.log_handler = logbook.TestHandler(bubble=True)
self.log_handler.push_application()
self._base_setup_called = True
def __init__(self, filename=None):
"""
:param filename: The name of the logfile
:type filename: str | None
"""
self.filename = filename
self.logging_disabled = filename is None
self._analytics_logger = None
self._event_id_generator = Counter()
self._log_cache = collections.deque()
self._logger = log.get_logger(__name__)
def _create_mock_slave(self, num_executors=5):
"""
:type num_executors: int
:rtype: Slave | MagicMock
"""
slave_spec = Slave('', 0) # constructor values don't matter since this is just a spec object
mock_slave = MagicMock(spec_set=slave_spec, url=self._FAKE_SLAVE_URL, num_executors=num_executors)
counter = Counter()
mock_slave.claim_executor.side_effect = counter.increment
mock_slave.free_executor.side_effect = counter.decrement
return mock_slave
def __init__(self, slave_url, num_executors):
"""
:type slave_url: str
:type num_executors: int
"""
self.url = slave_url
self.num_executors = num_executors
self.id = self._slave_id_counter.increment()
self._num_executors_in_use = Counter()
self._network = Network(min_connection_poolsize=num_executors)
self.current_build_id = None
self._is_alive = True
self._slave_api = UrlBuilder(slave_url, self.API_VERSION)
self._logger = log.get_logger(__name__)
def __init__(self, slave_url, num_executors, slave_session_id=None):
"""
:type slave_url: str
:type num_executors: int
:type slave_session_id: str
"""
self.url = slave_url
self.num_executors = num_executors
self.id = self._slave_id_counter.increment()
self._num_executors_in_use = Counter()
self._network = Network(min_connection_poolsize=num_executors)
self.current_build_id = None
self._last_heartbeat_time = datetime.now()
self._is_alive = True
self._is_in_shutdown_mode = False
self._slave_api = UrlBuilder(slave_url, self.API_VERSION)
self._session_id = slave_session_id
self._logger = log.get_logger(__name__)
class Slave(object):
API_VERSION = 'v1'
_slave_id_counter = Counter()
def __init__(self, slave_url, num_executors):
"""
:type slave_url: str
:type num_executors: int
"""
self.url = slave_url
self.num_executors = num_executors
self.id = self._slave_id_counter.increment()
self._num_executors_in_use = Counter()
self._network = Network(min_connection_poolsize=num_executors)
self.current_build_id = None
self._is_alive = True
self._is_in_shutdown_mode = False
self._slave_api = UrlBuilder(slave_url, self.API_VERSION)
self._logger = log.get_logger(__name__)
def api_representation(self):
return {
'url': self.url,
'id': self.id,
'num_executors': self.num_executors,
'num_executors_in_use': self.num_executors_in_use(),
'current_build_id': self.current_build_id,
'is_alive': self.is_alive(),
'is_in_shutdown_mode': self._is_in_shutdown_mode,
}
def mark_as_idle(self):
"""
Do bookkeeping when this slave becomes idle. Error if the slave cannot be idle.
If the slave is in shutdown mode, clear the build_id, kill the slave, and raise an error.
"""
if self._num_executors_in_use.value() != 0:
raise Exception(
'Trying to mark slave idle while {} executors still in use.',
self._num_executors_in_use.value())
self.current_build_id = None
if self._is_in_shutdown_mode:
self.kill()
raise SlaveMarkedForShutdownError
def setup(self, build, executor_start_index):
"""
Execute a setup command on the slave for the specified build. The setup process executes asynchronously on the
slave and the slave will alert the master when setup is complete and it is ready to start working on subjobs.
:param build: The build to set up this slave to work on
:type build: Build
:param executor_start_index: The index the slave should number its executors from for this build
:type executor_start_index: int
"""
slave_project_type_params = build.build_request.build_parameters(
).copy()
slave_project_type_params.update(
build.project_type.slave_param_overrides())
setup_url = self._slave_api.url('build', build.build_id(), 'setup')
post_data = {
'project_type_params': slave_project_type_params,
'build_executor_start_index': executor_start_index,
}
self.current_build_id = build.build_id()
self._network.post_with_digest(setup_url, post_data, Secret.get())
def teardown(self):
"""
Tell the slave to run the build teardown
"""
if self.is_alive():
teardown_url = self._slave_api.url('build', self.current_build_id,
'teardown')
self._network.post(teardown_url)
else:
self._logger.notice(
'Teardown request to slave {} was not sent since slave is disconnected.',
self.url)
def start_subjob(self, subjob):
"""
:type subjob: Subjob
"""
if not self.is_alive():
raise DeadSlaveError(
'Tried to start a subjob on a dead slave! ({}, id: {})'.format(
self.url, self.id))
if self._is_in_shutdown_mode:
raise SlaveMarkedForShutdownError(
'Tried to start a subjob on a slave in shutdown mode. ({}, id: {})'
.format(self.url, self.id))
SafeThread(target=self._async_start_subjob, args=(subjob, )).start()
def _async_start_subjob(self, subjob):
"""
:type subjob: Subjob
"""
execution_url = self._slave_api.url('build', subjob.build_id(),
'subjob', subjob.subjob_id())
post_data = {'atomic_commands': subjob.atomic_commands()}
response = self._network.post_with_digest(execution_url,
post_data,
Secret.get(),
error_on_failure=True)
subjob_executor_id = response.json().get('executor_id')
analytics.record_event(analytics.MASTER_TRIGGERED_SUBJOB,
executor_id=subjob_executor_id,
build_id=subjob.build_id(),
subjob_id=subjob.subjob_id(),
slave_id=self.id)
def num_executors_in_use(self):
return self._num_executors_in_use.value()
def claim_executor(self):
new_count = self._num_executors_in_use.increment()
if new_count > self.num_executors:
raise Exception(
'Cannot claim executor on slave {}. No executors left.'.format(
self.url))
return new_count
def free_executor(self):
new_count = self._num_executors_in_use.decrement()
if new_count < 0:
raise Exception(
'Cannot free executor on slave {}. All are free.'.format(
self.url))
return new_count
def is_alive(self, use_cached=True):
"""
Is the slave API responsive?
:param use_cached: Should we use the last returned value of the network check to the slave? If True,
will return cached value. If False, this method will perform an actual network call to the slave.
:type use_cached: bool
:rtype: bool
"""
if use_cached:
return self._is_alive
try:
response = self._network.get(self._slave_api.url())
if not response.ok:
self._is_alive = False
else:
response_data = response.json()
if 'slave' not in response_data or 'is_alive' not in response_data[
'slave']:
self._logger.warning(
'{}\'s API is missing key slave[\'is_alive\'].',
self.url)
self._is_alive = False
elif not isinstance(response_data['slave']['is_alive'], bool):
self._logger.warning(
'{}\'s API key \'is_alive\' is not a boolean.',
self.url)
self._is_alive = False
else:
self._is_alive = response_data['slave']['is_alive']
except requests.exceptions.ConnectionError:
self._logger.warning('Slave with url {} is offline.', self.url)
self._is_alive = False
return self._is_alive
def set_is_alive(self, value):
"""
Setter for the self._is_alive attribute.
:type value: bool
"""
self._is_alive = value
def set_shutdown_mode(self):
"""
Mark this slave as being in shutdown mode. Slaves in shutdown mode will not get new subjobs and will be
killed when they finish teardown, or killed immediately if they are not processing a build.
"""
self._is_in_shutdown_mode = True
if self.current_build_id is None:
self.kill()
def is_shutdown(self):
"""
Whether the slave is in shutdown mode.
"""
return self._is_in_shutdown_mode
def kill(self):
"""
Instructs the slave process to kill itself.
"""
kill_url = self._slave_api.url('kill')
self._network.post_with_digest(kill_url, {}, Secret.get())
self.mark_dead()
def mark_dead(self):
"""
Marks the slave dead.
"""
self.set_is_alive(False)
self.current_build_id = None
class Build(object):
"""
A build is a single execution of any configured job. This class:
- exposes the overall status of the build
- keeps track of the build's subjobs and their completion state
- manages slaves that have been assigned to accept this build's subjobs
"""
_build_id_counter = Counter(
) # class-level counter for assigning build ids
def __init__(self, build_request):
"""
:type build_request: BuildRequest
"""
self._logger = get_logger(__name__)
self._build_id = self._build_id_counter.increment()
self.build_request = build_request
self._artifacts_archive_file = None
self._build_artifact = None
""" :type : BuildArtifact"""
self._error_message = None
self.is_prepared = False
self._preparation_coin = SingleUseCoin(
) # protects against separate threads calling prepare() more than once
self._is_canceled = False
self._project_type = None
self._build_completion_lock = Lock(
) # protects against more than one thread detecting the build's finish
self._slaves_allocated = []
self._num_executors_allocated = 0
self._num_executors_in_use = 0
self._max_executors = float('inf')
self._max_executors_per_slave = float('inf')
self._all_subjobs_by_id = {}
self._unstarted_subjobs = None
self._finished_subjobs = None
self._postbuild_tasks_are_finished = False
self._teardowns_finished = False
self._timing_file_path = None
def api_representation(self):
return {
'id': self._build_id,
'status': self._status(),
'artifacts': self.
_artifacts_archive_file, # todo: this should probably be a url, not a file path
'details': self._detail_message,
'error_message': self._error_message,
'num_atoms': self._num_atoms,
'num_subjobs': len(self._all_subjobs_by_id),
'failed_atoms': self._failed_atoms(
), # todo: print the file contents instead of paths
'result': self._result(),
}
def prepare(self, subjobs, project_type, job_config):
"""
:type subjobs: list[Subjob]
:type project_type: project_type.project_type.ProjectType
:type job_config: JobConfig
"""
if not self._preparation_coin.spend():
raise RuntimeError(
'prepare() was called more than once on build {}.'.format(
self._build_id))
self._project_type = project_type
self._unstarted_subjobs = Queue(maxsize=len(subjobs))
self._finished_subjobs = Queue(maxsize=len(subjobs))
for subjob in subjobs:
self._all_subjobs_by_id[subjob.subjob_id()] = subjob
self._unstarted_subjobs.put(subjob)
self._max_executors = job_config.max_executors
self._max_executors_per_slave = job_config.max_executors_per_slave
self._timing_file_path = project_type.timing_file_path(job_config.name)
self.is_prepared = True
def finish(self):
"""
Called when all slaves are done with this build (and any teardown is complete)
"""
if self._subjobs_are_finished:
self._teardowns_finished = True
else:
raise RuntimeError(
'Tried to finish build {} but not all subjobs are complete'.
format(self._build_id))
def build_id(self):
"""
:rtype: int
"""
return self._build_id
def needs_more_slaves(self):
"""
Determine whether or not this build should have more slaves allocated to it.
:rtype: bool
"""
return self._num_executors_allocated < self._max_executors and not self._unstarted_subjobs.empty(
)
def allocate_slave(self, slave):
"""
Allocate a slave to this build. This tells the slave to execute setup commands for this build.
:type slave: Slave
"""
self._slaves_allocated.append(slave)
self._num_executors_allocated += min(slave.num_executors,
self._max_executors_per_slave)
slave.setup(self.build_id(),
project_type_params=self.build_request.build_parameters())
def all_subjobs(self):
"""
Returns a list of subjobs for this build
:rtype: list[Subjob]
"""
return [subjob for subjob in self._all_subjobs_by_id.values()]
def subjob(self, subjob_id):
"""
Returns a single subjob
:type subjob_id: int
:rtype: Subjob
"""
subjob = self._all_subjobs_by_id.get(subjob_id)
if subjob is None:
raise ItemNotFoundError('Invalid subjob id.')
return subjob
def begin_subjob_executions_on_slave(self, slave):
"""
Begin subjob executions on a slave. This should be called once after the specified slave has already run
build_setup commands for this build.
:type slave: Slave
"""
for slave_executor_count in range(slave.num_executors):
if (self._num_executors_in_use >= self._max_executors
or slave_executor_count >= self._max_executors_per_slave):
break
slave.claim_executor()
self._num_executors_in_use += 1
self.execute_next_subjob_on_slave(slave)
def execute_next_subjob_on_slave(self, slave):
"""
Grabs an unstarted subjob off the queue and sends it to the specified slave to be executed. If the unstarted
subjob queue is empty, we mark the slave as idle.
:type slave: Slave
"""
try:
subjob = self._unstarted_subjobs.get(block=False)
self._logger.debug('Sending subjob {} (build {}) to slave {}.',
subjob.subjob_id(), subjob.build_id(),
slave.url)
slave.start_subjob(subjob)
except Empty:
num_executors_in_use = slave.free_executor()
if num_executors_in_use == 0:
try:
self._slaves_allocated.remove(slave)
except ValueError:
pass # We have already deallocated this slave, no need to teardown
else:
slave.teardown()
def handle_subjob_payload(self, subjob_id, payload=None):
if not payload:
self._logger.warning('No payload for subjob {}.', subjob_id)
return
# Assertion: all payloads received from subjobs are uniquely named.
result_file_path = os.path.join(self._build_results_dir(),
payload['filename'])
try:
app.util.fs.write_file(payload['body'], result_file_path)
app.util.fs.extract_tar(result_file_path, delete=True)
self._logger.debug('Payload for subjob {} written.', subjob_id)
except:
self._logger.warning('Writing payload for subjob {} FAILED.',
subjob_id)
raise
def _read_subjob_timings_from_results(self):
"""
Collect timing data from all subjobs
:rtype: dict [str, float]
"""
timings = {}
for _, subjob in self._all_subjobs_by_id.items():
timings.update(subjob.read_timings())
return timings
def mark_subjob_complete(self, subjob_id):
"""
:type subjob_id: int
"""
subjob = self._all_subjobs_by_id[int(subjob_id)]
with self._build_completion_lock:
self._finished_subjobs.put(subjob, block=False)
subjobs_are_finished = self._subjobs_are_finished
# We use a local variable here which was set inside the _build_completion_lock to prevent a race condition
if subjobs_are_finished:
self._logger.info("All results received for build {}!",
self._build_id)
SafeThread(target=self._perform_async_postbuild_tasks,
name='PostBuild{}'.format(self._build_id)).start()
def mark_failed(self, failure_reason):
"""
Mark a build as failed and set a failure reason. The failure reason should be something we can present to the
end user of ClusterRunner, so try not to include detailed references to internal implementation.
:type failure_reason: str
"""
self._logger.error('Build {} failed: {}', self.build_id(),
failure_reason)
self._error_message = failure_reason
def cancel(self):
"""
Cancel a running build
"""
# Early exit if build is not running
if self._status() in [
BuildStatus.FINISHED, BuildStatus.ERROR, BuildStatus.CANCELED
]:
return
self._is_canceled = True
# Deplete the unstarted subjob queue.
# TODO: Handle situation where cancel() is called while subjobs are being added to _unstarted_subjobs
while self._unstarted_subjobs is not None and not self._unstarted_subjobs.empty(
):
try:
# A subjob may be asynchronously pulled from this queue, so we need to avoid blocking when empty.
self._unstarted_subjobs.get(block=False)
except Empty:
break
def validate_update_params(self, update_params):
"""
Determine if a dict of update params are valid, and generate an error if not
:param update_params: Params passed into a PUT for this build
:type update_params: dict [str, str]
:return: Whether the params are valid and a response containing an error message if not
:rtype: tuple [bool, dict [str, str]]
"""
keys_and_values_allowed = {'status': ['canceled']}
message = None
for key, value in update_params.items():
if key not in keys_and_values_allowed.keys():
message = 'Key ({}) is not in list of allowed keys ({})'.\
format(key, ",".join(keys_and_values_allowed.keys()))
elif value not in keys_and_values_allowed[key]:
message = 'Value ({}) is not in list of allowed values ({}) for {}'.\
format(value, keys_and_values_allowed[key], key)
if message is not None:
return False, {'error': message}
return True, {}
def update_state(self, update_params):
"""
Make updates to the state of this build given a set of update params
:param update_params: The keys and values to update on this build
:type update_params: dict [str, str]
"""
success = False
for key, value in update_params.items():
if key == 'status':
if value == 'canceled':
self.cancel()
success = True
return success
@property
def artifacts_archive_file(self):
return self._artifacts_archive_file
@property
def _num_subjobs_total(self):
return len(self._all_subjobs_by_id)
@property
def _num_subjobs_finished(self):
return 0 if not self._finished_subjobs else self._finished_subjobs.qsize(
)
@property
def _num_atoms(self):
if self._status() not in [BuildStatus.BUILDING, BuildStatus.FINISHED]:
return None
return sum([
len(subjob.atomic_commands())
for subjob in self._all_subjobs_by_id.values()
])
@property
def _subjobs_are_finished(self):
return self._is_canceled or (self.is_prepared
and self._finished_subjobs.full())
@property
def is_finished(self):
# TODO: Clean up this logic or move everything into a state machine
build_fully_completed = self._postbuild_tasks_are_finished and self._teardowns_finished
return self._is_canceled or build_fully_completed
@property
def is_unstarted(self):
return self.is_prepared and self._num_executors_allocated == 0 and self._unstarted_subjobs.full(
)
@property
def has_error(self):
return self._error_message is not None
@property
def _detail_message(self):
if self._num_subjobs_total > 0:
return '{} of {} subjobs are complete ({:.1f}%).'.format(
self._num_subjobs_finished, self._num_subjobs_total,
100 * self._num_subjobs_finished / self._num_subjobs_total)
return None
def _status(self):
"""
:rtype: BuildStatus
"""
if self.has_error:
return BuildStatus.ERROR
elif self._is_canceled:
return BuildStatus.CANCELED
elif not self.is_prepared or self.is_unstarted:
return BuildStatus.QUEUED
elif self.is_finished:
return BuildStatus.FINISHED
else:
return BuildStatus.BUILDING
def _failed_atoms(self):
"""
The commands which failed
:rtype: list [str] | None
"""
if self._is_canceled:
return []
if self.is_finished:
# dict.values() returns a view object in python 3, so wrapping values() in a list
return list(self._build_artifact.get_failed_commands().values())
return None
def _result(self):
"""
:rtype: str | None
"""
if self._is_canceled:
return BuildResult.FAILURE
if self.is_finished:
if len(self._build_artifact.get_failed_commands()) == 0:
return BuildResult.NO_FAILURES
return BuildResult.FAILURE
return None
def _perform_async_postbuild_tasks(self):
"""
Once a build is complete, certain tasks can be performed asynchronously.
"""
# @TODO There is a race condition here where the build is marked finished before the results archive
# is prepared. If the user requests the build status before archival finishes, the 'artifacts'
# value in the post body will be None. self.is_finished should be conditional on whether archival
# is finished.
self._create_build_artifact()
self._logger.debug('Postbuild tasks completed for build {}',
self.build_id())
self._postbuild_tasks_are_finished = True
def _create_build_artifact(self):
self._build_artifact = BuildArtifact(self._build_results_dir())
self._build_artifact.generate_failures_file()
self._build_artifact.write_timing_data(
self._timing_file_path, self._read_subjob_timings_from_results())
self._artifacts_archive_file = app.util.fs.compress_directory(
self._build_results_dir(), 'results.tar.gz')
def _build_results_dir(self):
return os.path.join(
Configuration['results_directory'],
str(self.build_id()),
)
import time
from app.util import fs
from app.util.conf.configuration import Configuration
from app.util.counter import Counter
BUILD_REQUEST_QUEUED = 'BUILD_REQUEST_QUEUED'
MASTER_RECEIVED_RESULT = 'MASTER_RECEIVED_RESULT'
MASTER_TRIGGERED_SUBJOB = 'MASTER_TRIGGERED_SUBJOB'
SERVICE_STARTED = 'SERVICE_STARTED'
SUBJOB_EXECUTION_FINISH = 'SUBJOB_EXECUTION_FINISH'
SUBJOB_EXECUTION_START = 'SUBJOB_EXECUTION_START'
_analytics_logger = None
_eventlog_file = None
_event_id_generator = Counter()
def initialize(eventlog_file=None):
"""
Initialize the analytics output. This will cause analytics events to be output to either a file or stdout.
If this function is not called, analytics events will not be output. If it is called with a filename, the events
will be output to that file. If it is called with 'STDOUT' or None, the events will be output to stdout.
:param eventlog_file: The filename to output events to, 'STDOUT' to output to stdout, None to disable event logging
:type eventlog_file: str | None
"""
global _analytics_logger, _eventlog_file
_eventlog_file = eventlog_file
class Slave(object):
API_VERSION = 'v1'
_slave_id_counter = Counter()
def __init__(self, slave_url, num_executors, slave_session_id=None):
"""
:type slave_url: str
:type num_executors: int
:type slave_session_id: str
"""
self.url = slave_url
self.num_executors = num_executors
self.id = self._slave_id_counter.increment()
self._num_executors_in_use = Counter()
self._network = Network(min_connection_poolsize=num_executors)
self.current_build_id = None
self._is_alive = True
self._is_in_shutdown_mode = False
self._slave_api = UrlBuilder(slave_url, self.API_VERSION)
self._session_id = slave_session_id
self._logger = log.get_logger(__name__)
def __str__(self):
return '<slave #{} - {}>'.format(self.id, self.url)
def api_representation(self):
return {
'url': self.url,
'id': self.id,
'session_id': self._session_id,
'num_executors': self.num_executors,
'num_executors_in_use': self.num_executors_in_use(),
'current_build_id': self.current_build_id,
'is_alive': self.is_alive(),
'is_in_shutdown_mode': self._is_in_shutdown_mode,
}
def mark_as_idle(self):
"""
Do bookkeeping when this slave becomes idle. Error if the slave cannot be idle.
If the slave is in shutdown mode, clear the build_id, kill the slave, and raise an error.
"""
if self._num_executors_in_use.value() != 0:
raise Exception(
'Trying to mark slave idle while {} executors still in use.',
self._num_executors_in_use.value())
self.current_build_id = None
if self._is_in_shutdown_mode:
self.kill()
raise SlaveMarkedForShutdownError
def setup(self, build: Build, executor_start_index: int) -> bool:
"""
Execute a setup command on the slave for the specified build. The setup process executes asynchronously on the
slave and the slave will alert the master when setup is complete and it is ready to start working on subjobs.
:param build: The build to set up this slave to work on
:param executor_start_index: The index the slave should number its executors from for this build
:return: Whether or not the call to start setup on the slave was successful
"""
slave_project_type_params = build.build_request.build_parameters(
).copy()
slave_project_type_params.update(
build.project_type.slave_param_overrides())
setup_url = self._slave_api.url('build', build.build_id(), 'setup')
post_data = {
'project_type_params': slave_project_type_params,
'build_executor_start_index': executor_start_index,
}
self.current_build_id = build.build_id()
try:
self._network.post_with_digest(setup_url, post_data, Secret.get())
except (requests.ConnectionError, requests.Timeout) as ex:
self._logger.warning('Setup call to {} failed with {}: {}.', self,
ex.__class__.__name__, str(ex))
self.mark_dead()
return False
return True
def teardown(self):
"""
Tell the slave to run the build teardown
"""
if not self.is_alive():
self._logger.notice(
'Teardown request to slave {} was not sent since slave is disconnected.',
self.url)
return
teardown_url = self._slave_api.url('build', self.current_build_id,
'teardown')
try:
self._network.post(teardown_url)
except (requests.ConnectionError, requests.Timeout):
self._logger.warning(
'Teardown request to slave failed because slave is unresponsive.'
)
self.mark_dead()
def start_subjob(self, subjob):
"""
:type subjob: Subjob
"""
if not self.is_alive():
raise DeadSlaveError(
'Tried to start a subjob on a dead slave! ({}, id: {})'.format(
self.url, self.id))
if self._is_in_shutdown_mode:
raise SlaveMarkedForShutdownError(
'Tried to start a subjob on a slave in shutdown mode. ({}, id: {})'
.format(self.url, self.id))
# todo: This should not be a SafeThread. https://github.com/box/ClusterRunner/issues/337
SafeThread(target=self._async_start_subjob, args=(subjob, )).start()
def _async_start_subjob(self, subjob):
"""
:type subjob: Subjob
"""
execution_url = self._slave_api.url('build', subjob.build_id(),
'subjob', subjob.subjob_id())
post_data = {'atomic_commands': subjob.atomic_commands()}
response = self._network.post_with_digest(execution_url,
post_data,
Secret.get(),
error_on_failure=True)
subjob_executor_id = response.json().get('executor_id')
analytics.record_event(analytics.MASTER_TRIGGERED_SUBJOB,
executor_id=subjob_executor_id,
build_id=subjob.build_id(),
subjob_id=subjob.subjob_id(),
slave_id=self.id)
def num_executors_in_use(self):
return self._num_executors_in_use.value()
def claim_executor(self):
new_count = self._num_executors_in_use.increment()
if new_count > self.num_executors:
raise Exception(
'Cannot claim executor on slave {}. No executors left.'.format(
self.url))
return new_count
def free_executor(self):
new_count = self._num_executors_in_use.decrement()
if new_count < 0:
raise Exception(
'Cannot free executor on slave {}. All are free.'.format(
self.url))
return new_count
def is_alive(self, use_cached: bool = True) -> bool:
"""
Is the slave API responsive?
Note that if the slave API responds but its session id does not match the one we've stored in this
instance, then this method will still return false.
:param use_cached: Should we use the last returned value of the network check to the slave? If True,
will return cached value. If False, this method will perform an actual network call to the slave.
:return: Whether or not the slave is alive
"""
if use_cached:
return self._is_alive
try:
response = self._network.get(
self._slave_api.url(), headers=self._expected_session_header())
if not response.ok:
self.mark_dead()
else:
response_data = response.json()
if 'slave' not in response_data or 'is_alive' not in response_data[
'slave']:
self._logger.warning(
'{}\'s API is missing key slave[\'is_alive\'].',
self.url)
self.mark_dead()
elif not isinstance(response_data['slave']['is_alive'], bool):
self._logger.warning(
'{}\'s API key \'is_alive\' is not a boolean.',
self.url)
self.mark_dead()
else:
self._is_alive = response_data['slave']['is_alive']
except (requests.ConnectionError, requests.Timeout):
self.mark_dead()
return self._is_alive
def set_is_alive(self, value):
"""
Setter for the self._is_alive attribute.
:type value: bool
"""
self._is_alive = value
def set_shutdown_mode(self):
"""
Mark this slave as being in shutdown mode. Slaves in shutdown mode will not get new subjobs and will be
killed when they finish teardown, or killed immediately if they are not processing a build.
"""
self._is_in_shutdown_mode = True
if self.current_build_id is None:
self.kill()
def is_shutdown(self):
"""
Whether the slave is in shutdown mode.
"""
return self._is_in_shutdown_mode
def kill(self):
"""
Instruct the slave process to kill itself.
"""
self._logger.notice('Killing {}', self)
kill_url = self._slave_api.url('kill')
try:
self._network.post_with_digest(kill_url, {}, Secret.get())
except (requests.ConnectionError, requests.Timeout):
pass
self.mark_dead()
def mark_dead(self):
"""
Mark the slave dead.
"""
self._logger.warning('{} has gone offline. Last build: {}', self,
self.current_build_id)
self._is_alive = False
self.current_build_id = None
self._network.reset_session(
) # Close any pooled connections for this slave.
def _expected_session_header(self):
"""
Return headers that should be sent with slave requests to verify that the master is still talking to
the same slave service that it originally connected to.
Note that adding these headers to existing requests may add new failure cases (e.g., slave API would
start returning a 412) so we should make sure all potential 412 errors are handled appropriately when
adding these headers to existing requests.
:rtype: dict
"""
headers = {}
if self._session_id:
headers[SessionId.EXPECTED_SESSION_HEADER_KEY] = self._session_id
return headers
class Build(object):
"""
A build is a single execution of any configured job. This class:
- exposes the overall status of the build
- keeps track of the build's subjobs and their completion state
- manages slaves that have been assigned to accept this build's subjobs
:type _build_id: int
:type _build_request: BuildRequest
:type _build_artifact: None | BuildArtifact
:type _error_message: None | str
:type _project_type: None | ProjectType
:type _timing_file_path: None | str
"""
_build_id_counter = Counter() # class-level counter for assigning build ids
def __init__(self, build_request):
"""
:type build_request: BuildRequest
"""
self._logger = get_logger(__name__)
self._build_id = self._build_id_counter.increment()
self._build_request = build_request
self._artifacts_tar_file = None # DEPRECATED - Use zip file instead
self._artifacts_zip_file = None
self._build_artifact = None
self._error_message = None
self._preparation_coin = SingleUseCoin() # protects against separate threads calling prepare() more than once
self._project_type = None
self._build_completion_lock = Lock() # protects against more than one thread detecting the build's finish
self._all_subjobs_by_id = OrderedDict()
self._unstarted_subjobs = None # WIP(joey): Move subjob queues to BuildScheduler class.
self._finished_subjobs = None
self._failed_atoms = None
self._postbuild_tasks_are_finished = False # WIP(joey): Remove and use build state.
self._timing_file_path = None
leave_state_callbacks = {build_state: self._on_leave_state
for build_state in BuildState}
self._state_machine = BuildFsm(
build_id=self._build_id,
enter_state_callbacks={
BuildState.ERROR: self._on_enter_error_state,
BuildState.CANCELED: self._on_enter_canceled_state,
BuildState.PREPARING: self._on_enter_preparing_state,
},
leave_state_callbacks=leave_state_callbacks
)
# Number of times build_setup has failed on this build. If
# setup_failures increases beyond MAX_SETUP_FAILURES, the build is
# cancelled
self.setup_failures = 0
def api_representation(self):
failed_atoms_api_representation = None
if self._get_failed_atoms() is not None:
failed_atoms_api_representation = [failed_atom.api_representation()
for failed_atom in self._get_failed_atoms()]
build_state = self._status()
# todo: PREPARING/PREPARED are new states -- make sure clients can handle them before exposing.
if build_state in (BuildState.PREPARING, BuildState.PREPARED):
build_state = BuildState.QUEUED
return {
'id': self._build_id,
'status': build_state,
'details': self._detail_message,
'error_message': self._error_message,
'num_atoms': self._num_atoms,
'num_subjobs': len(self._all_subjobs_by_id),
'failed_atoms': failed_atoms_api_representation,
'result': self._result(),
'request_params': self.build_request.build_parameters(),
# Convert self._state_timestamps to OrderedDict to make raw API response more readable. Sort the entries
# by numerically increasing dict value, with None values sorting highest.
'state_timestamps': OrderedDict(sorted(
[(state.lower(), timestamp) for state, timestamp in self._state_machine.transition_timestamps.items()],
key=lambda item: item[1] or float('inf'))),
}
def generate_project_type(self):
"""
Instantiate the project type for this build, populating the self._project_type instance variable.
As a side effect, this method also updates the build request's build_parameters dictionary
with the unique workspace directory path for this build.
:raises BuildProjectError when failed to instantiate project type
"""
# Generate a unique project build directory name that will be symlinked to the actual project directory
# later on when the project gets fetched.
build_specific_project_directory = self._generate_unique_symlink_path_for_build_repo()
# Because build_specific_project_directory is entirely internal and generated by ClusterRunner (it is a
# build-unique generated symlink), we must manually add it to the project_type_params
project_type_params = self.build_request.build_parameters()
project_type_params.update({'build_project_directory': build_specific_project_directory})
self._project_type = util.create_project_type(project_type_params)
if self._project_type is None:
raise BuildProjectError('Build failed due to an invalid project type.')
def prepare(self):
if not isinstance(self.build_request, BuildRequest):
raise RuntimeError('Build {} has no associated request object.'.format(self._build_id))
if not isinstance(self.project_type, ProjectType):
raise RuntimeError('Build {} has no project set.'.format(self._build_id))
if not self._preparation_coin.spend():
raise RuntimeError('prepare() was called more than once on build {}.'.format(self._build_id))
self._state_machine.trigger(BuildEvent.START_PREPARE)
def build_id(self):
"""
:rtype: int
"""
return self._build_id
@property
def build_request(self):
"""
:rtype: BuildRequest
"""
return self._build_request
def get_subjobs(self, offset: int=None, limit: int=None) -> List['Subjob']:
"""
Returns a list of subjobs for this build
:param offset: The starting index of the requested build
:param limit: The number of builds requested
"""
num_subjobs = len(self._all_subjobs_by_id)
start, end = get_paginated_indices(offset, limit, num_subjobs)
requested_subjobs = islice(self._all_subjobs_by_id, start, end)
return [self._all_subjobs_by_id[key] for key in requested_subjobs]
def subjob(self, subjob_id: int) -> Subjob:
"""Return the subjob for this build with the specified id."""
subjob = self._all_subjobs_by_id.get(subjob_id)
if subjob is None:
raise ItemNotFoundError('Invalid subjob id.')
return subjob
def complete_subjob(self, subjob_id, payload=None):
"""
Handle the subjob payload and mark the given subjob id for this build as complete.
:type subjob_id: int
:type payload: dict
"""
try:
self._handle_subjob_payload(subjob_id, payload)
self._mark_subjob_complete(subjob_id)
except Exception:
self._logger.exception('Error while completing subjob; marking build as failed.')
self.mark_failed('Error occurred while completing subjob {}.'.format(subjob_id))
raise
def _parse_payload_for_atom_exit_code(self, subjob_id):
subjob = self.subjob(subjob_id)
for atom_id in range(len(subjob.atoms)):
artifact_dir = BuildArtifact.atom_artifact_directory(
self.build_id(),
subjob.subjob_id(),
atom_id,
result_root=Configuration['results_directory']
)
atom_exit_code_file_sys_path = os.path.join(artifact_dir, BuildArtifact.EXIT_CODE_FILE)
with open(atom_exit_code_file_sys_path, 'r') as atom_exit_code_file:
subjob.atoms[atom_id].exit_code = int(atom_exit_code_file.read())
def _handle_subjob_payload(self, subjob_id, payload):
if not payload:
self._logger.warning('No payload for subjob {} of build {}.', subjob_id, self._build_id)
return
# Assertion: all payloads received from subjobs are uniquely named.
result_file_path = os.path.join(self._build_results_dir(), payload['filename'])
try:
app.util.fs.write_file(payload['body'], result_file_path)
app.util.fs.extract_tar(result_file_path, delete=True)
self._parse_payload_for_atom_exit_code(subjob_id)
except:
internal_errors.labels(ErrorType.SubjobWriteFailure).inc() # pylint: disable=no-member
self._logger.warning('Writing payload for subjob {} of build {} FAILED.', subjob_id, self._build_id)
raise
def _read_subjob_timings_from_results(self):
"""
Collect timing data from all subjobs
:rtype: dict [str, float]
"""
timings = {}
for _, subjob in self._all_subjobs_by_id.items():
timings.update(subjob.read_timings())
return timings
def _mark_subjob_complete(self, subjob_id):
"""
:type subjob_id: int
"""
subjob = self.subjob(subjob_id)
subjob.mark_completed()
with self._build_completion_lock:
self._finished_subjobs.put(subjob, block=False)
should_trigger_postbuild_tasks = self._all_subjobs_are_finished() and not self.is_stopped
# We use a local variable here which was set inside the _build_completion_lock to prevent a race condition
if should_trigger_postbuild_tasks:
self._logger.info("All results received for build {}!", self._build_id)
self.finish()
def mark_started(self):
"""
Mark the build as started.
"""
self._state_machine.trigger(BuildEvent.START_BUILDING)
def finish(self):
"""
Perform postbuild task and mark this build as finished.
"""
Thread(
target=self._perform_async_postbuild_tasks,
name='PostBuild{}'.format(self._build_id),
).start()
def mark_failed(self, failure_reason):
"""
Mark a build as failed and set a failure reason. The failure reason should be something we can present to the
end user of ClusterRunner, so try not to include detailed references to internal implementation.
:type failure_reason: str
"""
self._state_machine.trigger(BuildEvent.FAIL, error_msg=failure_reason)
def _on_enter_error_state(self, event):
"""
Store an error message for the build and log the failure. This method is triggered by
a state machine transition to the ERROR state.
:param event: The Fysom event object
"""
# WIP(joey): Should this be a reenter_state callback also? Should it check for previous error message?
default_error_msg = 'An unspecified error occurred.'
self._error_message = getattr(event, 'error_msg', default_error_msg)
self._logger.warning('Build {} failed: {}', self.build_id(), self._error_message)
def _on_enter_preparing_state(self, event):
"""
Prepare the build by atomization and subjobs creation.
This method is triggered by a state machine transition to the PREPARING state.
:param event: The Fysom event object
:type event: BuildEvent
"""
self._logger.info('Fetching project for build {}.', self._build_id)
self.project_type.fetch_project()
self._logger.info('Successfully fetched project for build {}.', self._build_id)
job_config = self.project_type.job_config()
if job_config is None:
raise RuntimeError('Build failed while trying to parse clusterrunner.yaml.')
subjobs = compute_subjobs_for_build(self._build_id, job_config, self.project_type)
self._unstarted_subjobs = Queue(maxsize=len(subjobs)) # WIP(joey): Move this into BuildScheduler?
self._finished_subjobs = Queue(maxsize=len(subjobs)) # WIP(joey): Remove this and just record finished count.
for subjob in subjobs:
self._all_subjobs_by_id[subjob.subjob_id()] = subjob
self._unstarted_subjobs.put(subjob)
self._timing_file_path = self._project_type.timing_file_path(job_config.name)
app.util.fs.create_dir(self._build_results_dir())
self._state_machine.trigger(BuildEvent.FINISH_PREPARE)
def _on_leave_state(self, event):
start_time = self._state_machine.transition_timestamps.get(event.src)
if start_time is not None:
elapsed = time.time() - start_time
build_state_duration_seconds.labels(event.src.value).observe(elapsed) # pylint: disable=no-member
else:
self._logger.warn('Build {} transitioned from state {} to state {} but never marked started timestamp for {}',
self._build_id, event.src, event.dst, event.src)
def cancel(self):
"""
Cancel a running build.
"""
self._state_machine.trigger(BuildEvent.CANCEL)
def _on_enter_canceled_state(self, event):
"""
:param event: The Fysom event object
:type event: BuildEvent
"""
self._logger.notice('Canceling build {}.', self._build_id)
# Set the kill_event to kill the subprocesses for the build
self.project_type.kill_subprocesses()
# Deplete the unstarted subjob queue.
# WIP(joey): Just remove this completely and adjust behavior of other methods based on self._is_canceled().
# TODO: Handle situation where cancel() is called while subjobs are being added to _unstarted_subjobs
while self._unstarted_subjobs is not None and not self._unstarted_subjobs.empty():
try:
# A subjob may be asynchronously pulled from this queue, so we need to avoid blocking when empty.
self._unstarted_subjobs.get(block=False)
except Empty:
break
def validate_update_params(self, update_params):
"""
Determine if a dict of update params are valid, and generate an error if not
:param update_params: Params passed into a PUT for this build
:type update_params: dict [str, str]
:return: Whether the params are valid and a response containing an error message if not
:rtype: tuple [bool, dict [str, str]]
"""
keys_and_values_allowed = {'status': ['canceled']}
message = None
for key, value in update_params.items():
if key not in keys_and_values_allowed.keys():
message = 'Key ({}) is not in list of allowed keys ({})'.\
format(key, ",".join(keys_and_values_allowed.keys()))
elif value not in keys_and_values_allowed[key]:
message = 'Value ({}) is not in list of allowed values ({}) for {}'.\
format(value, keys_and_values_allowed[key], key)
if message is not None:
return False, {'error': message}
return True, {}
def update_state(self, update_params):
"""
Make updates to the state of this build given a set of update params
:param update_params: The keys and values to update on this build
:type update_params: dict [str, str]
"""
success = False
for key, value in update_params.items():
if key == 'status':
if value == 'canceled':
self.cancel()
success = True
return success
@property
def project_type(self):
"""
:rtype: ProjectType
"""
return self._project_type
@property
def artifacts_zip_file(self):
"""Return the local path to the artifacts zip archive."""
return self._artifacts_zip_file
@property
def artifacts_tar_file(self):
"""
DEPRECATED: We are transitioning to zip files from tar.gz files for artifacts.
Return the local path to the artifacts tar.gz archive.
"""
self._logger.warning('The tar format for build artifact files is deprecated. File: {}',
self._artifacts_tar_file)
return self._artifacts_tar_file
# WIP(joey): Change some of these private @properties to methods.
@property
def _num_subjobs_total(self):
return len(self._all_subjobs_by_id)
@property
def _num_subjobs_finished(self):
return 0 if not self._finished_subjobs else self._finished_subjobs.qsize()
@property
def _num_atoms(self):
# todo: blacklist states instead of whitelist, or just check _all_subjobs_by_id directly
if self._status() not in [BuildState.BUILDING, BuildState.FINISHED]:
return None
return sum([len(subjob.atomic_commands()) for subjob in self._all_subjobs_by_id.values()])
def _all_subjobs_are_finished(self):
return self._finished_subjobs and self._finished_subjobs.full()
@property
def is_finished(self):
# WIP(joey): Calling logic should check _is_canceled if it needs to instead of including the check here.
return self.is_canceled or self._postbuild_tasks_are_finished
@property
def _detail_message(self):
if self._num_subjobs_total > 0:
return '{} of {} subjobs are complete ({:.1f}%).'.format(
self._num_subjobs_finished,
self._num_subjobs_total,
100 * self._num_subjobs_finished / self._num_subjobs_total
)
return None
def _status(self): # WIP(joey): Rename to _state.
"""
:rtype: BuildState
"""
return self._state_machine.state
@property
def has_error(self):
return self._status() is BuildState.ERROR
@property
def is_canceled(self):
return self._status() is BuildState.CANCELED
@property
def is_stopped(self):
return self._status() in (BuildState.ERROR, BuildState.CANCELED)
def _get_failed_atoms(self):
"""
The atoms that failed. Returns None if the build hasn't completed yet. Returns empty set if
build has completed and no atoms have failed.
:rtype: list[Atom] | None
"""
if self._failed_atoms is None and self.is_finished:
if self.is_canceled:
return []
self._failed_atoms = []
for subjob_id, atom_id in self._build_artifact.get_failed_subjob_and_atom_ids():
subjob = self.subjob(subjob_id)
atom = subjob.atoms[atom_id]
self._failed_atoms.append(atom)
return self._failed_atoms
def _result(self):
"""
Can return three states:
None:
FAILURE:
NO_FAILURES:
:rtype: BuildResult | None
"""
if self.is_canceled:
return BuildResult.FAILURE
if self.is_finished:
if len(self._build_artifact.get_failed_subjob_and_atom_ids()) == 0:
return BuildResult.NO_FAILURES
return BuildResult.FAILURE
return None
def _perform_async_postbuild_tasks(self):
"""
Once a build is complete, execute certain tasks like archiving the artifacts and writing timing
data. This method also transitions the FSM to finished after the postbuild tasks are complete.
"""
try:
timing_data = self._read_subjob_timings_from_results()
self._create_build_artifact(timing_data)
serialized_build_time_seconds.observe(sum(timing_data.values()))
self._delete_temporary_build_artifact_files()
self._postbuild_tasks_are_finished = True
self._state_machine.trigger(BuildEvent.POSTBUILD_TASKS_COMPLETE)
self._logger.notice('Completed build (id: {}), saving to database.'.format(self._build_id))
self.save()
except Exception as ex: # pylint: disable=broad-except
internal_errors.labels(ErrorType.PostBuildFailure).inc() # pylint: disable=no-member
self._logger.exception('Postbuild tasks failed for build {}.'.format(self._build_id))
self.mark_failed('Postbuild tasks failed due to an internal error: "{}"'.format(ex))
def _create_build_artifact(self, timing_data: Dict[str, float]): # pylint: disable=unsubscriptable-object
self._build_artifact = BuildArtifact(self._build_results_dir())
self._build_artifact.generate_failures_file()
self._build_artifact.write_timing_data(self._timing_file_path, timing_data)
self._artifacts_tar_file = app.util.fs.tar_directory(self._build_results_dir(),
BuildArtifact.ARTIFACT_TARFILE_NAME)
temp_tar_path = None
try:
# Temporarily move aside tar file so we can create a zip file, then move it back.
# This juggling can be removed once we're no longer creating tar artifacts.
temp_tar_path = shutil.move(self._artifacts_tar_file, tempfile.mktemp())
self._artifacts_zip_file = app.util.fs.zip_directory(self._build_results_dir(),
BuildArtifact.ARTIFACT_ZIPFILE_NAME)
except Exception: # pylint: disable=broad-except
internal_errors.labels(ErrorType.ZipFileCreationFailure).inc() # pylint: disable=no-member
# Due to issue #339 we are ignoring exceptions in the zip file creation for now.
self._logger.exception('Zipping of artifacts failed. This error will be ignored.')
finally:
if temp_tar_path:
shutil.move(temp_tar_path, self._artifacts_tar_file)
def _delete_temporary_build_artifact_files(self):
"""
Delete the temporary build result files that are no longer needed, due to the creation of the
build artifact tarball.
ONLY call this method after _create_build_artifact() has completed. Otherwise we have lost the build results.
"""
build_result_dir = self._build_results_dir()
start_time = time.time()
for path in os.listdir(build_result_dir):
# The build result archive is also stored in this same directory, so we must not delete it.
if path in (BuildArtifact.ARTIFACT_TARFILE_NAME, BuildArtifact.ARTIFACT_ZIPFILE_NAME):
continue
full_path = os.path.join(build_result_dir, path)
# Do NOT use app.util.fs.async_delete() here. That call will generate a temp directory for every
# atom, which can be in the thousands per build, and can lead to running up against the ulimit -Hn.
if os.path.isdir:
shutil.rmtree(full_path, ignore_errors=True)
else:
os.remove(full_path)
end_time = time.time() - start_time
self._logger.info('Completed deleting artifact files for {}, took {:.1f} seconds.', self._build_id, end_time)
def _build_results_dir(self):
return BuildArtifact.build_artifact_directory(self.build_id(), result_root=Configuration['results_directory'])
def _generate_unique_symlink_path_for_build_repo(self):
"""
Generate a unique symlink path for a build-specific repo. This method does NOT generate the symlink itself.
:rtype: str
"""
return os.path.join(Configuration['build_symlink_directory'], str(uuid.uuid4()))
# pylint: disable=protected-access
def save(self):
"""Serialize the Build object and update all of the parts to the database."""
with Connection.get() as session:
build_schema = session.query(BuildSchema).filter(BuildSchema.build_id == self._build_id).first()
failed_artifact_directories_schema = session.query(FailedArtifactDirectoriesSchema) \
.filter(FailedArtifactDirectoriesSchema.build_id == self._build_id) \
.all()
failed_subjob_atom_pairs_schema = session.query(FailedSubjobAtomPairsSchema) \
.filter(FailedSubjobAtomPairsSchema.build_id == self._build_id) \
.all()
atoms_schema = session.query(AtomsSchema).filter(AtomsSchema.build_id == self._build_id).all()
subjobs_schema = session.query(SubjobsSchema).filter(SubjobsSchema.build_id == self._build_id).all()
# If this wasn't found, it's safe to assume that the build doesn't exist within the database
if build_schema is None:
raise ItemNotFoundError('Unable to find build (id: {}) in database.'.format(self._build_id))
build_schema.artifacts_tar_file = self._artifacts_tar_file
build_schema.artifacts_zip_file = self._artifacts_zip_file
build_schema.error_message = self._error_message
build_schema.postbuild_tasks_are_finished = self._postbuild_tasks_are_finished
build_schema.setup_failures = self.setup_failures
build_schema.timing_file_path = self._timing_file_path
build_artifact_dir = None
if self._build_artifact is not None:
build_artifact_dir = self._build_artifact.build_artifact_dir
build_schema.build_artifact_dir = build_artifact_dir
if self._build_artifact is not None:
# Clear all old directories
session.query(FailedArtifactDirectoriesSchema) \
.filter(FailedArtifactDirectoriesSchema.build_id == self._build_id) \
.delete()
# Commit changes so we don't delete the newly added rows later
session.commit()
# Add all the updated versions of the directories
for directory in self._build_artifact._get_failed_artifact_directories():
failed_artifact_directory = FailedArtifactDirectoriesSchema(
build_id=self._build_id,
failed_artifact_directory=directory
)
session.add(failed_artifact_directory)
if self._build_artifact is not None:
# Clear all old directories
session.query(FailedSubjobAtomPairsSchema) \
.filter(FailedSubjobAtomPairsSchema.build_id == self._build_id) \
.delete()
# Commit changes so we don't delete the newly added rows later
session.commit()
# Add all the updated versions of the data
for subjob_id, atom_id in self._build_artifact.get_failed_subjob_and_atom_ids():
failed_subjob_and_atom_ids = FailedSubjobAtomPairsSchema(
build_id=self._build_id,
subjob_id=subjob_id,
atom_id=atom_id
)
session.add(failed_subjob_and_atom_ids)
build_schema.build_parameters = json.dumps(self._build_request.build_parameters())
fsm_timestamps = {state.lower(): timestamp for state, timestamp in self._state_machine.transition_timestamps.items()}
build_schema.state = self._status()
build_schema.queued_ts = fsm_timestamps['queued']
build_schema.finished_ts = fsm_timestamps['finished']
build_schema.prepared_ts = fsm_timestamps['prepared']
build_schema.preparing_ts = fsm_timestamps['preparing']
build_schema.error_ts = fsm_timestamps['error']
build_schema.canceled_ts = fsm_timestamps['canceled']
build_schema.building_ts = fsm_timestamps['building']
# Subjobs
# Clear all old Subjobs and Atoms
session.query(SubjobsSchema) \
.filter(SubjobsSchema.build_id == self._build_id) \
.delete()
session.query(AtomsSchema) \
.filter(AtomsSchema.build_id == self._build_id) \
.delete()
# Commit changes so we don't delete the newly added rows later
session.commit()
# Add all the updated versions of Subjobs and Atoms
subjobs = self._all_subjobs_by_id
for subjob_id in subjobs:
subjob = self._all_subjobs_by_id[subjob_id]
subjob_schema = SubjobsSchema(
subjob_id=subjob_id,
build_id=self._build_id,
completed=subjob.completed
)
session.add(subjob_schema)
# Atoms
for atom in subjob._atoms:
atom_schema = AtomsSchema(
atom_id=atom.id,
build_id=self._build_id,
subjob_id=subjob_id,
command_string=atom.command_string,
expected_time=atom.expected_time,
actual_time=atom.actual_time,
exit_code=atom.exit_code,
state=atom.state
)
session.add(atom_schema)
@classmethod
def load_from_db(cls, build_id):
"""
Given a build_id, fetch all the stored information from the database to reconstruct
a Build object to represent that build.
:param build_id: The id of the build to recreate.
"""
with Connection.get() as session:
build_schema = session.query(BuildSchema).filter(BuildSchema.build_id == build_id).first()
failed_artifact_directories_schema = session.query(FailedArtifactDirectoriesSchema) \
.filter(FailedArtifactDirectoriesSchema.build_id == build_id) \
.all()
failed_subjob_atom_pairs_schema = session.query(FailedSubjobAtomPairsSchema) \
.filter(FailedSubjobAtomPairsSchema.build_id == build_id) \
.all()
atoms_schema = session.query(AtomsSchema).filter(AtomsSchema.build_id == build_id).all()
subjobs_schema = session.query(SubjobsSchema).filter(SubjobsSchema.build_id == build_id).all()
# If a query returns None, then we know the build wasn't found in the database
if not build_schema:
return None
build_parameters = json.loads(build_schema.build_parameters)
# Genereate a BuildRequest object with our query response
build_request = BuildRequest(build_parameters)
# Create initial Build object, we will be altering the state of this as we get more data
build = Build(build_request)
build._build_id = build_id
# Manually generate ProjectType object for build and create a `job_config` since this is usually done in `prepare()`
build.generate_project_type()
job_config = build.project_type.job_config()
# Manually update build data
build._artifacts_tar_file = build_schema.artifacts_tar_file
build._artifacts_zip_file = build_schema.artifacts_zip_file
build._error_message = build_schema.error_message
build._postbuild_tasks_are_finished = bool(int(build_schema.postbuild_tasks_are_finished))
build.setup_failures = build_schema.setup_failures
build._timing_file_path = build_schema.timing_file_path
# Manually set the state machine timestamps
build._state_machine._transition_timestamps = {
BuildState.QUEUED: build_schema.queued_ts,
BuildState.FINISHED: build_schema.finished_ts,
BuildState.PREPARED: build_schema.prepared_ts,
BuildState.PREPARING: build_schema.preparing_ts,
BuildState.ERROR: build_schema.error_ts,
BuildState.CANCELED: build_schema.canceled_ts,
BuildState.BUILDING: build_schema.building_ts
}
build._state_machine._fsm.current = BuildState[build_schema.state]
build_artifact = BuildArtifact(build_schema.build_artifact_dir)
directories = []
for directory in failed_artifact_directories_schema:
directories.append(directory.failed_artifact_directory)
build_artifact._failed_artifact_directories = directories
pairs = []
for pair in failed_subjob_atom_pairs_schema:
pairs.append((pair.subjob_id, pair.atom_id))
build_artifact._q_failed_subjob_atom_pairs = pairs
build._build_artifact = build_artifact
atoms_by_subjob_id = {}
for atom in atoms_schema:
atoms_by_subjob_id.setdefault(atom.subjob_id, [])
atoms_by_subjob_id[atom.subjob_id].append(Atom(
atom.command_string,
atom.expected_time,
atom.actual_time,
atom.exit_code,
atom.state,
atom.atom_id,
atom.subjob_id
))
subjobs = OrderedDict()
for subjob in subjobs_schema:
atoms = atoms_by_subjob_id[subjob.subjob_id]
# Add atoms after subjob is created so we don't alter their state on initialization
subjob_to_add = Subjob(build_id, subjob.subjob_id, build.project_type, job_config, [])
subjob_to_add._atoms = atoms
subjob_to_add.completed = subjob.completed
subjobs[subjob.subjob_id] = subjob_to_add
build._all_subjobs_by_id = subjobs
# Place subjobs into correct queues within the build
build._unstarted_subjobs = Queue(maxsize=len(subjobs))
build._finished_subjobs = Queue(maxsize=len(subjobs))
for _, subjob in subjobs.items():
if subjob.completed:
build._finished_subjobs.put(subjob)
else:
build._unstarted_subjobs.put(subjob)
return build
class Build(object):
"""
A build is a single execution of any configured job. This class:
- exposes the overall status of the build
- keeps track of the build's subjobs and their completion state
- manages slaves that have been assigned to accept this build's subjobs
:type _build_id: int
:type _build_request: BuildRequest
:type _build_artifact: None | BuildArtifact
:type _error_message: None | str
:type _project_type: None | ProjectType
:type _timing_file_path: None | str
"""
_build_id_counter = Counter(
) # class-level counter for assigning build ids
def __init__(self, build_request):
"""
:type build_request: BuildRequest
"""
self._logger = get_logger(__name__)
self._build_id = self._build_id_counter.increment()
self._build_request = build_request
self._artifacts_archive_file = None
self._build_artifact = None
self._error_message = None
self._preparation_coin = SingleUseCoin(
) # protects against separate threads calling prepare() more than once
self._project_type = None
self._build_completion_lock = Lock(
) # protects against more than one thread detecting the build's finish
self._all_subjobs_by_id = {}
self._unstarted_subjobs = None # WIP(joey): Move subjob queues to BuildScheduler class.
self._finished_subjobs = None
self._failed_atoms = None
self._postbuild_tasks_are_finished = False # WIP(joey): Remove and use build state.
self._timing_file_path = None
self._state_machine = BuildFsm(build_id=self._build_id,
enter_state_callbacks={
BuildState.ERROR:
self._on_enter_error_state,
BuildState.CANCELED:
self._on_enter_canceled_state,
})
def api_representation(self):
failed_atoms_api_representation = None
if self._get_failed_atoms() is not None:
failed_atoms_api_representation = [
failed_atom.api_representation()
for failed_atom in self._get_failed_atoms()
]
build_state = self._status()
# todo: PREPARING/PREPARED are new states -- make sure clients can handle them before exposing.
if build_state in (BuildState.PREPARING, BuildState.PREPARED):
build_state = BuildState.QUEUED
return {
'id':
self._build_id,
'status':
build_state,
'artifacts':
self.
_artifacts_archive_file, # todo: this should probably be a url, not a file path
'details':
self._detail_message,
'error_message':
self._error_message,
'num_atoms':
self._num_atoms,
'num_subjobs':
len(self._all_subjobs_by_id),
'failed_atoms':
failed_atoms_api_representation,
'result':
self._result(),
'request_params':
self.build_request.build_parameters(),
# Convert self._state_timestamps to OrderedDict to make raw API response more readable. Sort the entries
# by numerically increasing dict value, with None values sorting highest.
'state_timestamps':
OrderedDict(
sorted([(state.lower(), timestamp) for state, timestamp in
self._state_machine.transition_timestamps.items()],
key=lambda item: item[1] or float('inf'))),
}
def generate_project_type(self):
"""
Instantiate the project type for this build, populating the self._project_type instance variable.
As a side effect, this method also updates the build request's build_parameters dictionary
with the unique workspace directory path for this build.
:raises BuildProjectError when failed to instantiate project type
"""
# Generate a unique project build directory name that will be symlinked to the actual project directory
# later on when the project gets fetched.
build_specific_project_directory = self._generate_unique_symlink_path_for_build_repo(
)
# Because build_specific_project_directory is entirely internal and generated by ClusterRunner (it is a
# build-unique generated symlink), we must manually add it to the project_type_params
project_type_params = self.build_request.build_parameters()
project_type_params.update(
{'build_project_directory': build_specific_project_directory})
self._project_type = util.create_project_type(project_type_params)
if self._project_type is None:
raise BuildProjectError(
'Build failed due to an invalid project type.')
def prepare(self, subjob_calculator):
"""
:param subjob_calculator: Used after project fetch to atomize and group subjobs for this build
:type subjob_calculator: SubjobCalculator
"""
if not isinstance(self.build_request, BuildRequest):
raise RuntimeError(
'Build {} has no associated request object.'.format(
self._build_id))
if not isinstance(self.project_type, ProjectType):
raise RuntimeError('Build {} has no project set.'.format(
self._build_id))
if not self._preparation_coin.spend():
raise RuntimeError(
'prepare() was called more than once on build {}.'.format(
self._build_id))
self._state_machine.trigger(BuildEvent.START_PREPARE)
# WIP(joey): Move the following code into a PREPARING state callback
# (so that it won't execute if the build has already been canceled.)
self._logger.info('Fetching project for build {}.', self._build_id)
self.project_type.fetch_project()
self._logger.info('Successfully fetched project for build {}.',
self._build_id)
job_config = self.project_type.job_config()
if job_config is None:
raise RuntimeError(
'Build failed while trying to parse clusterrunner.yaml.')
subjobs = subjob_calculator.compute_subjobs_for_build(
self._build_id, job_config, self.project_type)
self._unstarted_subjobs = Queue(
maxsize=len(subjobs)) # WIP(joey): Move this into BuildScheduler?
self._finished_subjobs = Queue(maxsize=len(
subjobs)) # WIP(joey): Remove this and just record finished count.
for subjob in subjobs:
self._all_subjobs_by_id[subjob.subjob_id()] = subjob
self._unstarted_subjobs.put(subjob)
self._timing_file_path = self._project_type.timing_file_path(
job_config.name)
app.util.fs.create_dir(self._build_results_dir())
self._state_machine.trigger(BuildEvent.FINISH_PREPARE)
def build_id(self):
"""
:rtype: int
"""
return self._build_id
@property
def build_request(self):
"""
:rtype: BuildRequest
"""
return self._build_request
def all_subjobs(self):
"""
Returns a list of subjobs for this build
:rtype: list[Subjob]
"""
return [subjob for subjob in self._all_subjobs_by_id.values()]
def subjob(self, subjob_id):
"""
Returns a single subjob
:type subjob_id: int
:rtype: Subjob
"""
subjob = self._all_subjobs_by_id.get(subjob_id)
if subjob is None:
raise ItemNotFoundError('Invalid subjob id.')
return subjob
def complete_subjob(self, subjob_id, payload=None):
"""
Handle the subjob payload and mark the given subjob id for this build as complete.
:type subjob_id: int
:type payload: dict
"""
try:
self._handle_subjob_payload(subjob_id, payload)
self._mark_subjob_complete(subjob_id)
except Exception:
self._logger.exception(
'Error while completing subjob; marking build as failed.')
self.mark_failed(
'Error occurred while completing subjob {}.'.format(subjob_id))
raise
def _parse_payload_for_atom_exit_code(self, subjob_id):
subjob = self.subjob(subjob_id)
for atom_id in range(len(subjob.atoms)):
artifact_dir = BuildArtifact.atom_artifact_directory(
self.build_id(),
subjob.subjob_id(),
atom_id,
result_root=Configuration['results_directory'])
atom_exit_code_file_sys_path = os.path.join(
artifact_dir, BuildArtifact.EXIT_CODE_FILE)
with open(atom_exit_code_file_sys_path,
'r') as atom_exit_code_file:
subjob.atoms[atom_id].exit_code = int(
atom_exit_code_file.read())
def _handle_subjob_payload(self, subjob_id, payload):
if not payload:
self._logger.warning('No payload for subjob {} of build {}.',
subjob_id, self._build_id)
return
# Assertion: all payloads received from subjobs are uniquely named.
result_file_path = os.path.join(self._build_results_dir(),
payload['filename'])
try:
app.util.fs.write_file(payload['body'], result_file_path)
app.util.fs.extract_tar(result_file_path, delete=True)
self._parse_payload_for_atom_exit_code(subjob_id)
except:
self._logger.warning(
'Writing payload for subjob {} of build {} FAILED.', subjob_id,
self._build_id)
raise
def _read_subjob_timings_from_results(self):
"""
Collect timing data from all subjobs
:rtype: dict [str, float]
"""
timings = {}
for _, subjob in self._all_subjobs_by_id.items():
timings.update(subjob.read_timings())
return timings
def _mark_subjob_complete(self, subjob_id):
"""
:type subjob_id: int
"""
subjob = self.subjob(subjob_id)
subjob.mark_completed()
with self._build_completion_lock:
self._finished_subjobs.put(subjob, block=False)
should_trigger_postbuild_tasks = self._all_subjobs_are_finished(
) and not self._is_stopped()
# We use a local variable here which was set inside the _build_completion_lock to prevent a race condition
if should_trigger_postbuild_tasks:
self._logger.info("All results received for build {}!",
self._build_id)
SafeThread(target=self._perform_async_postbuild_tasks,
name='PostBuild{}'.format(self._build_id)).start()
def mark_started(self):
"""
Mark the build as started.
"""
self._state_machine.trigger(BuildEvent.START_BUILDING)
def finish(self):
"""
Perform postbuild task and mark this build as finished.
"""
# This method also transitions the FSM to finished after the postbuild tasks are complete.
self._perform_async_postbuild_tasks()
def mark_failed(self, failure_reason):
"""
Mark a build as failed and set a failure reason. The failure reason should be something we can present to the
end user of ClusterRunner, so try not to include detailed references to internal implementation.
:type failure_reason: str
"""
self._state_machine.trigger(BuildEvent.FAIL, error_msg=failure_reason)
def _on_enter_error_state(self, event):
"""
Store an error message for the build and log the failure. This method is triggered by
a state machine transition to the ERROR state.
:param event: The Fysom event object
"""
# WIP(joey): Should this be a reenter_state callback also? Should it check for previous error message?
default_error_msg = 'An unspecified error occurred.'
self._error_message = getattr(event, 'error_msg', default_error_msg)
self._logger.warning('Build {} failed: {}', self.build_id(),
self._error_message)
def cancel(self):
"""
Cancel a running build.
"""
self._logger.notice('Request received to cancel build {}.',
self._build_id)
self._state_machine.trigger(BuildEvent.CANCEL)
def _on_enter_canceled_state(self, event):
# Deplete the unstarted subjob queue.
# WIP(joey): Just remove this completely and adjust behavior of other methods based on self._is_canceled().
# TODO: Handle situation where cancel() is called while subjobs are being added to _unstarted_subjobs
while self._unstarted_subjobs is not None and not self._unstarted_subjobs.empty(
):
try:
# A subjob may be asynchronously pulled from this queue, so we need to avoid blocking when empty.
self._unstarted_subjobs.get(block=False)
except Empty:
break
def validate_update_params(self, update_params):
"""
Determine if a dict of update params are valid, and generate an error if not
:param update_params: Params passed into a PUT for this build
:type update_params: dict [str, str]
:return: Whether the params are valid and a response containing an error message if not
:rtype: tuple [bool, dict [str, str]]
"""
keys_and_values_allowed = {'status': ['canceled']}
message = None
for key, value in update_params.items():
if key not in keys_and_values_allowed.keys():
message = 'Key ({}) is not in list of allowed keys ({})'.\
format(key, ",".join(keys_and_values_allowed.keys()))
elif value not in keys_and_values_allowed[key]:
message = 'Value ({}) is not in list of allowed values ({}) for {}'.\
format(value, keys_and_values_allowed[key], key)
if message is not None:
return False, {'error': message}
return True, {}
def update_state(self, update_params):
"""
Make updates to the state of this build given a set of update params
:param update_params: The keys and values to update on this build
:type update_params: dict [str, str]
"""
success = False
for key, value in update_params.items():
if key == 'status':
if value == 'canceled':
self.cancel()
success = True
return success
@property
def project_type(self):
"""
:rtype: ProjectType
"""
return self._project_type
@property
def artifacts_archive_file(self):
return self._artifacts_archive_file
# WIP(joey): Change some of these private @properties to methods.
@property
def _num_subjobs_total(self):
return len(self._all_subjobs_by_id)
@property
def _num_subjobs_finished(self):
return 0 if not self._finished_subjobs else self._finished_subjobs.qsize(
)
@property
def _num_atoms(self):
# todo: blacklist states instead of whitelist, or just check _all_subjobs_by_id directly
if self._status() not in [BuildState.BUILDING, BuildState.FINISHED]:
return None
return sum([
len(subjob.atomic_commands())
for subjob in self._all_subjobs_by_id.values()
])
def _all_subjobs_are_finished(self):
return self._finished_subjobs and self._finished_subjobs.full()
@property
def is_finished(self):
# WIP(joey): Calling logic should check _is_canceled if it needs to instead of including the check here.
return self._is_canceled() or self._postbuild_tasks_are_finished
@property
def _detail_message(self):
if self._num_subjobs_total > 0:
return '{} of {} subjobs are complete ({:.1f}%).'.format(
self._num_subjobs_finished, self._num_subjobs_total,
100 * self._num_subjobs_finished / self._num_subjobs_total)
return None
def _status(self): # WIP(joey): Rename to _state.
"""
:rtype: BuildState
"""
return self._state_machine.state
@property
def has_error(self):
return self._status() is BuildState.ERROR
def _is_canceled(self):
return self._status() is BuildState.CANCELED
def _is_stopped(self):
return self._status() in (BuildState.ERROR, BuildState.CANCELED)
def _get_failed_atoms(self):
"""
The atoms that failed. Returns None if the build hasn't completed yet. Returns empty set if
build has completed and no atoms have failed.
:rtype: list[Atom] | None
"""
if self._failed_atoms is None and self.is_finished:
if self._is_canceled():
return []
self._failed_atoms = []
for subjob_id, atom_id in self._build_artifact.get_failed_subjob_and_atom_ids(
):
subjob = self.subjob(subjob_id)
atom = subjob.atoms[atom_id]
self._failed_atoms.append(atom)
return self._failed_atoms
def _result(self):
"""
Can return three states:
None:
FAILURE:
NO_FAILURES:
:rtype: BuildResult | None
"""
if self._is_canceled():
return BuildResult.FAILURE
if self.is_finished:
if len(self._build_artifact.get_failed_subjob_and_atom_ids()) == 0:
return BuildResult.NO_FAILURES
return BuildResult.FAILURE
return None
def _perform_async_postbuild_tasks(self):
"""
Once a build is complete, certain tasks can be performed asynchronously.
"""
self._create_build_artifact()
self._delete_temporary_build_artifact_files()
self._postbuild_tasks_are_finished = True
self._state_machine.trigger(BuildEvent.POSTBUILD_TASKS_COMPLETE)
def _create_build_artifact(self):
self._build_artifact = BuildArtifact(self._build_results_dir())
self._build_artifact.generate_failures_file()
self._build_artifact.write_timing_data(
self._timing_file_path, self._read_subjob_timings_from_results())
self._artifacts_archive_file = app.util.fs.compress_directory(
self._build_results_dir(), BuildArtifact.ARTIFACT_FILE_NAME)
def _delete_temporary_build_artifact_files(self):
"""
Delete the temporary build result files that are no longer needed, due to the creation of the
build artifact tarball.
ONLY call this method after _create_build_artifact() has completed. Otherwise we have lost the build results.
"""
build_result_dir = self._build_results_dir()
start_time = time.time()
for path in os.listdir(build_result_dir):
# The build result tar-ball is also stored in this same directory, so we must not delete it.
if path == BuildArtifact.ARTIFACT_FILE_NAME:
continue
full_path = os.path.join(build_result_dir, path)
# Do NOT use app.util.fs.async_delete() here. That call will generate a temp directory for every
# atom, which can be in the thousands per build, and can lead to running up against the ulimit -Hn.
if os.path.isdir:
shutil.rmtree(full_path, ignore_errors=True)
else:
os.remove(full_path)
end_time = time.time() - start_time
self._logger.info(
'Completed deleting artifact files for {}, took {:.1f} seconds.',
self._build_id, end_time)
def _build_results_dir(self):
return BuildArtifact.build_artifact_directory(
self.build_id(), result_root=Configuration['results_directory'])
def _generate_unique_symlink_path_for_build_repo(self):
"""
Generate a unique symlink path for a build-specific repo. This method does NOT generate the symlink itself.
:rtype: str
"""
return os.path.join(Configuration['build_symlink_directory'],
str(uuid.uuid4()))
class Slave(object):
API_VERSION = 'v1'
_slave_id_counter = Counter()
def __init__(self, slave_url, num_executors):
"""
:type slave_url: str
:type num_executors: int
"""
self.url = slave_url
self.num_executors = num_executors
self.id = self._slave_id_counter.increment()
self._num_executors_in_use = Counter()
self._network = Network(min_connection_poolsize=num_executors)
self.current_build_id = None
self._is_alive = True
self._slave_api = UrlBuilder(slave_url, self.API_VERSION)
self._logger = log.get_logger(__name__)
def api_representation(self):
return {
'url': self.url,
'id': self.id,
'num_executors': self.num_executors,
'num_executors_in_use': self.num_executors_in_use(),
'current_build_id': self.current_build_id,
}
def mark_as_idle(self):
"""
Do bookkeeping when this slave becomes idle. Error if the slave cannot be idle.
"""
if self._num_executors_in_use.value() != 0:
raise Exception(
'Trying to mark slave idle while {} executors still in use.',
self._num_executors_in_use.value())
self.current_build_id = None
def setup(self, build_id, project_type_params):
"""
Execute a setup command on the slave for the specified build. The command is executed asynchronously from the
perspective of this method, but any subjobs will block until the slave finishes executing the setup command.
:param build_id: The build id that this setup command is for.
:type build_id: int
:param project_type_params: The parameters that define the project type this build will execute in
:type project_type_params: dict
"""
setup_url = self._slave_api.url('build', build_id, 'setup')
slave_project_type_params = util.project_type_params_for_slave(
project_type_params)
post_data = {
'project_type_params': slave_project_type_params,
}
self._network.post_with_digest(setup_url, post_data, Secret.get())
self.current_build_id = build_id
def teardown(self):
"""
Tell the slave to run the build teardown
"""
if self.is_alive():
teardown_url = self._slave_api.url('build', self.current_build_id,
'teardown')
self._network.post(teardown_url)
else:
self._logger.notice(
'Teardown request to slave {} was not sent since slave is disconnected.',
self.url)
def start_subjob(self, subjob):
"""
:type subjob: Subjob
"""
if not self.is_alive():
raise RuntimeError(
'Tried to start a subjob on a dead slave! ({}, id: {})'.format(
self.url, self.id))
SafeThread(target=self._async_start_subjob, args=(subjob, )).start()
def _async_start_subjob(self, subjob):
"""
:type subjob: Subjob
"""
execution_url = self._slave_api.url('build', subjob.build_id(),
'subjob', subjob.subjob_id())
post_data = {
'subjob_artifact_dir': subjob.artifact_dir(),
'atomic_commands': subjob.atomic_commands(),
}
response = self._network.post_with_digest(execution_url,
post_data,
Secret.get(),
error_on_failure=True)
subjob_executor_id = response.json().get('executor_id')
analytics.record_event(analytics.MASTER_TRIGGERED_SUBJOB,
executor_id=subjob_executor_id,
build_id=subjob.build_id(),
subjob_id=subjob.subjob_id(),
slave_id=self.id)
def num_executors_in_use(self):
return self._num_executors_in_use.value()
def claim_executor(self):
new_count = self._num_executors_in_use.increment()
if new_count > self.num_executors:
raise Exception(
'Cannot claim executor on slave {}. No executors left.'.format(
self.url))
return new_count
def free_executor(self):
new_count = self._num_executors_in_use.decrement()
if new_count < 0:
raise Exception(
'Cannot free executor on slave {}. All are free.'.format(
self.url))
return new_count
def is_alive(self, use_cached=True):
"""
Is the slave API responsive?
:param use_cached: Should we use the last returned value of the network check to the slave? If True,
will return cached value. If False, this method will perform an actual network call to the slave.
:type use_cached: bool
:rtype: bool
"""
if use_cached:
return self._is_alive
try:
response = self._network.get(self._slave_api.url())
if not response.ok:
self._is_alive = False
else:
response_data = response.json()
if 'slave' not in response_data or 'is_alive' not in response_data[
'slave']:
self._logger.warning(
'{}\'s API is missing key slave[\'is_alive\'].',
self.url)
self._is_alive = False
elif not isinstance(response_data['slave']['is_alive'], bool):
self._logger.warning(
'{}\'s API key \'is_alive\' is not a boolean.',
self.url)
self._is_alive = False
else:
self._is_alive = response_data['slave']['is_alive']
except ConnectionError:
self._logger.warning('Slave with url {} is offline.', self.url)
self._is_alive = False
return self._is_alive
def set_is_alive(self, value):
"""
Setter for the self._is_alive attribute.
:type value: bool
"""
self._is_alive = value
class Build(object):
"""
A build is a single execution of any configured job. This class:
- exposes the overall status of the build
- keeps track of the build's subjobs and their completion state
- manages slaves that have been assigned to accept this build's subjobs
"""
_build_id_counter = Counter(
) # class-level counter for assigning build ids
def __init__(self, build_request):
"""
:type build_request: BuildRequest
"""
self._logger = get_logger(__name__)
self._build_id = self._build_id_counter.increment()
self.build_request = build_request
self._artifacts_archive_file = None
self._build_artifact = None
""" :type : BuildArtifact"""
self._error_message = None
self.is_prepared = False
self._setup_is_started = False
self._preparation_coin = SingleUseCoin(
) # protects against separate threads calling prepare() more than once
self._is_canceled = False
self._project_type = None
self._build_completion_lock = Lock(
) # protects against more than one thread detecting the build's finish
self._all_subjobs_by_id = {}
self._unstarted_subjobs = None # WIP: Move subjob queues to BuildScheduler class.
self._finished_subjobs = None
self._failed_atoms = None
self._postbuild_tasks_are_finished = False
self._timing_file_path = None
self._state_timestamps = {status: None
for status in BuildStatus
} # initialize all timestamps to None
self._record_state_timestamp(BuildStatus.QUEUED)
def api_representation(self):
failed_atoms_api_representation = None
if self._get_failed_atoms() is not None:
failed_atoms_api_representation = [
failed_atom.api_representation()
for failed_atom in self._get_failed_atoms()
]
return {
'id':
self._build_id,
'status':
self._status(),
'artifacts':
self.
_artifacts_archive_file, # todo: this should probably be a url, not a file path
'details':
self._detail_message,
'error_message':
self._error_message,
'num_atoms':
self._num_atoms,
'num_subjobs':
len(self._all_subjobs_by_id),
'failed_atoms':
failed_atoms_api_representation,
'result':
self._result(),
'request_params':
self.build_request.build_parameters(),
# Convert self._state_timestamps to OrderedDict to make raw API response more readable. Sort the entries
# by numerically increasing dict value, with None values sorting highest.
'state_timestamps':
OrderedDict(
sorted(
[(state.lower(), timestamp)
for state, timestamp in self._state_timestamps.items()],
key=lambda item: item[1] or float('inf'))),
}
def generate_project_type(self):
"""
Instantiate the project type for this build, populating the self._project_type instance variable.
As a side effect, this method also updates the build request's build_parameters dictionary
with the unique workspace directory path for this build.
:raises BuildProjectError when failed to instantiate project type
"""
# Generate a unique project build directory name that will be symlinked to the actual project directory
# later on when the project gets fetched.
build_specific_project_directory = self._generate_unique_symlink_path_for_build_repo(
)
# Because build_specific_project_directory is entirely internal and generated by ClusterRunner (it is a
# build-unique generated symlink), we must manually add it to the project_type_params
project_type_params = self.build_request.build_parameters()
project_type_params.update(
{'build_project_directory': build_specific_project_directory})
self._project_type = util.create_project_type(project_type_params)
if self._project_type is None:
raise BuildProjectError(
'Build failed due to an invalid project type.')
def prepare(self, subjob_calculator):
"""
:param subjob_calculator: Used after project fetch to atomize and group subjobs for this build
:type subjob_calculator: SubjobCalculator
"""
if not isinstance(self.build_request, BuildRequest):
raise RuntimeError(
'Build {} has no associated request object.'.format(
self._build_id))
if not isinstance(self.project_type, ProjectType):
raise RuntimeError('Build {} has no project set.'.format(
self._build_id))
if not self._preparation_coin.spend():
raise RuntimeError(
'prepare() was called more than once on build {}.'.format(
self._build_id))
self._logger.info('Fetching project for build {}.', self._build_id)
self.project_type.fetch_project()
self._logger.info('Successfully fetched project for build {}.',
self._build_id)
job_config = self.project_type.job_config()
if job_config is None:
raise RuntimeError(
'Build failed while trying to parse clusterrunner.yaml.')
subjobs = subjob_calculator.compute_subjobs_for_build(
self._build_id, job_config, self.project_type)
self._unstarted_subjobs = Queue(maxsize=len(subjobs))
self._finished_subjobs = Queue(maxsize=len(subjobs))
for subjob in subjobs:
self._all_subjobs_by_id[subjob.subjob_id()] = subjob
self._unstarted_subjobs.put(subjob)
self._timing_file_path = self._project_type.timing_file_path(
job_config.name)
self.is_prepared = True
self._record_state_timestamp(BuildStatus.PREPARED)
def build_id(self):
"""
:rtype: int
"""
return self._build_id
def all_subjobs(self):
"""
Returns a list of subjobs for this build
:rtype: list[Subjob]
"""
return [subjob for subjob in self._all_subjobs_by_id.values()]
def subjob(self, subjob_id):
"""
Returns a single subjob
:type subjob_id: int
:rtype: Subjob
"""
subjob = self._all_subjobs_by_id.get(subjob_id)
if subjob is None:
raise ItemNotFoundError('Invalid subjob id.')
return subjob
def complete_subjob(self, subjob_id, payload=None):
"""
Handle the subjob payload and mark the given subjob id for this build as complete.
:type subjob_id: int
:type payload: dict
"""
try:
self._handle_subjob_payload(subjob_id, payload)
self._mark_subjob_complete(subjob_id)
except Exception:
self._logger.exception(
'Error while completing subjob; marking build as failed.')
self.mark_failed(
'Error occurred while completing subjob {}.'.format(subjob_id))
raise
def _parse_payload_for_atom_exit_code(self, subjob_id):
subjob = self.subjob(subjob_id)
for atom_id in range(len(subjob.atoms)):
artifact_dir = BuildArtifact.atom_artifact_directory(
self.build_id(),
subjob.subjob_id(),
atom_id,
result_root=Configuration['results_directory'])
atom_exit_code_file_sys_path = os.path.join(
artifact_dir, BuildArtifact.EXIT_CODE_FILE)
with open(atom_exit_code_file_sys_path,
'r') as atom_exit_code_file:
subjob.atoms[atom_id].exit_code = int(
atom_exit_code_file.read())
def _handle_subjob_payload(self, subjob_id, payload):
if not payload:
self._logger.warning('No payload for subjob {} of build {}.',
subjob_id, self._build_id)
return
# Assertion: all payloads received from subjobs are uniquely named.
result_file_path = os.path.join(self._build_results_dir(),
payload['filename'])
try:
app.util.fs.write_file(payload['body'], result_file_path)
app.util.fs.extract_tar(result_file_path, delete=True)
self._parse_payload_for_atom_exit_code(subjob_id)
except:
self._logger.warning(
'Writing payload for subjob {} of build {} FAILED.', subjob_id,
self._build_id)
raise
def _read_subjob_timings_from_results(self):
"""
Collect timing data from all subjobs
:rtype: dict [str, float]
"""
timings = {}
for _, subjob in self._all_subjobs_by_id.items():
timings.update(subjob.read_timings())
return timings
def _mark_subjob_complete(self, subjob_id):
"""
:type subjob_id: int
"""
subjob = self.subjob(subjob_id)
subjob.mark_completed()
with self._build_completion_lock:
self._finished_subjobs.put(subjob, block=False)
subjobs_are_finished = self._subjobs_are_finished
# We use a local variable here which was set inside the _build_completion_lock to prevent a race condition
if subjobs_are_finished:
self._logger.info("All results received for build {}!",
self._build_id)
SafeThread(target=self._perform_async_postbuild_tasks,
name='PostBuild{}'.format(self._build_id)).start()
def mark_started(self):
self._setup_is_started = True
self._record_state_timestamp(BuildStatus.BUILDING)
def mark_failed(self, failure_reason):
"""
Mark a build as failed and set a failure reason. The failure reason should be something we can present to the
end user of ClusterRunner, so try not to include detailed references to internal implementation.
:type failure_reason: str
"""
self._logger.error('Build {} failed: {}', self.build_id(),
failure_reason)
self._error_message = failure_reason
self._record_state_timestamp(BuildStatus.ERROR)
def cancel(self):
"""
Cancel a running build
"""
# Early exit if build is not running
if self._status() in [
BuildStatus.FINISHED, BuildStatus.ERROR, BuildStatus.CANCELED
]:
self._logger.notice(
'Ignoring cancel request for build {}. Build is already in state {}.',
self._build_id, self._status())
return
self._logger.notice('Canceling build {}.', self._build_id)
self._is_canceled = True
self._record_state_timestamp(BuildStatus.CANCELED)
# Deplete the unstarted subjob queue.
# TODO: Handle situation where cancel() is called while subjobs are being added to _unstarted_subjobs
while self._unstarted_subjobs is not None and not self._unstarted_subjobs.empty(
):
try:
# A subjob may be asynchronously pulled from this queue, so we need to avoid blocking when empty.
self._unstarted_subjobs.get(block=False)
except Empty:
break
def validate_update_params(self, update_params):
"""
Determine if a dict of update params are valid, and generate an error if not
:param update_params: Params passed into a PUT for this build
:type update_params: dict [str, str]
:return: Whether the params are valid and a response containing an error message if not
:rtype: tuple [bool, dict [str, str]]
"""
keys_and_values_allowed = {'status': ['canceled']}
message = None
for key, value in update_params.items():
if key not in keys_and_values_allowed.keys():
message = 'Key ({}) is not in list of allowed keys ({})'.\
format(key, ",".join(keys_and_values_allowed.keys()))
elif value not in keys_and_values_allowed[key]:
message = 'Value ({}) is not in list of allowed values ({}) for {}'.\
format(value, keys_and_values_allowed[key], key)
if message is not None:
return False, {'error': message}
return True, {}
def update_state(self, update_params):
"""
Make updates to the state of this build given a set of update params
:param update_params: The keys and values to update on this build
:type update_params: dict [str, str]
"""
success = False
for key, value in update_params.items():
if key == 'status':
if value == 'canceled':
self.cancel()
success = True
return success
@property
def project_type(self):
"""
:rtype: ProjectType
"""
return self._project_type
@property
def artifacts_archive_file(self):
return self._artifacts_archive_file
@property
def _num_subjobs_total(self):
return len(self._all_subjobs_by_id)
@property
def _num_subjobs_finished(self):
return 0 if not self._finished_subjobs else self._finished_subjobs.qsize(
)
@property
def _num_atoms(self):
if self._status() not in [BuildStatus.BUILDING, BuildStatus.FINISHED]:
return None
return sum([
len(subjob.atomic_commands())
for subjob in self._all_subjobs_by_id.values()
])
@property
def _subjobs_are_finished(self):
return self._is_canceled or (self.is_prepared
and self._finished_subjobs.full())
@property
def is_finished(self):
# TODO: Clean up this logic or move everything into a state machine
return self._is_canceled or self._postbuild_tasks_are_finished
@property
def is_unstarted(self):
return self.is_prepared and not self._setup_is_started and self._unstarted_subjobs.full(
)
@property
def has_error(self):
return self._error_message is not None
@property
def _detail_message(self):
if self._num_subjobs_total > 0:
return '{} of {} subjobs are complete ({:.1f}%).'.format(
self._num_subjobs_finished, self._num_subjobs_total,
100 * self._num_subjobs_finished / self._num_subjobs_total)
return None
def _status(self):
"""
:rtype: BuildStatus
"""
if self.has_error:
return BuildStatus.ERROR
elif self._is_canceled:
return BuildStatus.CANCELED
elif not self.is_prepared or self.is_unstarted:
return BuildStatus.QUEUED
elif self.is_finished:
return BuildStatus.FINISHED
else:
return BuildStatus.BUILDING
def _get_failed_atoms(self):
"""
The atoms that failed. Returns None if the build hasn't completed yet. Returns empty set if
build has completed and no atoms have failed.
:rtype: list[Atom] | None
"""
if self._failed_atoms is None and self.is_finished:
if self._is_canceled:
return []
self._failed_atoms = []
for subjob_id, atom_id in self._build_artifact.get_failed_subjob_and_atom_ids(
):
subjob = self.subjob(subjob_id)
atom = subjob.atoms[atom_id]
self._failed_atoms.append(atom)
return self._failed_atoms
def _result(self):
"""
:rtype: str | None
"""
if self._is_canceled:
return BuildResult.FAILURE
if self.is_finished:
if len(self._build_artifact.get_failed_subjob_and_atom_ids()) == 0:
return BuildResult.NO_FAILURES
return BuildResult.FAILURE
return None
def _perform_async_postbuild_tasks(self):
"""
Once a build is complete, certain tasks can be performed asynchronously.
"""
self._create_build_artifact()
self._logger.debug('Postbuild tasks completed for build {}',
self.build_id())
self._postbuild_tasks_are_finished = True
self._record_state_timestamp(BuildStatus.FINISHED)
def _create_build_artifact(self):
self._build_artifact = BuildArtifact(self._build_results_dir())
self._build_artifact.generate_failures_file()
self._build_artifact.write_timing_data(
self._timing_file_path, self._read_subjob_timings_from_results())
self._artifacts_archive_file = app.util.fs.compress_directory(
self._build_results_dir(), 'results.tar.gz')
def _build_results_dir(self):
return BuildArtifact.build_artifact_directory(
self.build_id(), result_root=Configuration['results_directory'])
def _generate_unique_symlink_path_for_build_repo(self):
"""
Generate a unique symlink path for a build-specific repo. This method does NOT generate the symlink itself.
:rtype: str
"""
return os.path.join(Configuration['build_symlink_directory'],
str(uuid.uuid4()))
def get_state_timestamp(self, build_status):
"""
Get the recorded timestamp for a given build status. This may be None if the build has not yet reached
the specified state.
:param build_status: The build status for which to retrieve the corresponding timestamp
:type build_status: BuildStatus
:return: The timestamp for the specified status
:rtype: float | None
"""
return self._state_timestamps.get(build_status)
def _record_state_timestamp(self, build_status):
"""
Record a timestamp for a given build status. This is used to record the timing of the various build phases and
is exposed via the Build object's API representation.
:param build_status: The build status for which to record a timestamp
:type build_status: BuildStatus
"""
if self._state_timestamps.get(build_status) is not None:
self._logger.warning(
'Overwriting timestamp for build {}, status {}'.format(
self.build_id(), build_status))
self._state_timestamps[build_status] = time.time()