def __init__(self, control_address, worker_count, credentials=None):
self._worker_count = worker_count
self._worker_index = 0
if credentials is None:
logging.info('Creating insecure control channel.')
self._control_channel = grpc.insecure_channel(control_address)
else:
logging.info('Creating secure control channel.')
self._control_channel = grpc.secure_channel(
control_address, credentials)
grpc.channel_ready_future(self._control_channel).result(timeout=60)
logging.info('Control channel established.')
self._control_channel = grpc.intercept_channel(self._control_channel,
WorkerIdInterceptor())
self._data_channel_factory = data_plane.GrpcClientDataChannelFactory(
credentials)
self._state_handler_factory = GrpcStateHandlerFactory()
self.workers = queue.Queue()
# one thread is enough for getting the progress report.
# Assumption:
# Progress report generation should not do IO or wait on other resources.
# Without wait, having multiple threads will not improve performance and
# will only add complexity.
self._progress_thread_pool = futures.ThreadPoolExecutor(max_workers=1)
self._process_thread_pool = futures.ThreadPoolExecutor(
max_workers=self._worker_count)
self._instruction_id_vs_worker = {}
self._fns = {}
self._responses = queue.Queue()
self._process_bundle_queue = queue.Queue()
self._unscheduled_process_bundle = set()
logging.info('Initializing SDKHarness with %s workers.',
self._worker_count)
def __init__(self, control_address):
self._control_channel = grpc.insecure_channel(control_address)
self._data_channel_factory = data_plane.GrpcClientDataChannelFactory()
# TODO: Ensure thread safety to run with more than 1 thread.
self._default_work_thread_pool = futures.ThreadPoolExecutor(
max_workers=1)
self._progress_thread_pool = futures.ThreadPoolExecutor(max_workers=1)
def test_source_split(self):
source = RangeSource(0, 100)
expected_splits = list(source.split(30))
worker = sdk_harness.SdkWorker(
None, data_plane.GrpcClientDataChannelFactory())
worker.register(
beam_fn_api_pb2.RegisterRequest(process_bundle_descriptor=[
beam_fn_api_pb2.ProcessBundleDescriptor(primitive_transform=[
beam_fn_api_pb2.PrimitiveTransform(
function_spec=sdk_harness.serialize_and_pack_py_fn(
SourceBundle(1.0, source, None, None),
sdk_harness.PYTHON_SOURCE_URN,
id="src"))
])
]))
split_response = worker.initial_source_split(
beam_fn_api_pb2.InitialSourceSplitRequest(
desired_bundle_size_bytes=30, source_reference="src"))
self.assertEqual(expected_splits, [
sdk_harness.unpack_and_deserialize_py_fn(s.source)
for s in split_response.splits
])
self.assertEqual([s.weight for s in expected_splits],
[s.relative_size for s in split_response.splits])
def __init__(
self,
control_address,
worker_count,
credentials=None,
worker_id=None,
# Caching is disabled by default
state_cache_size=0,
profiler_factory=None):
self._alive = True
self._worker_count = worker_count
self._worker_index = 0
self._worker_id = worker_id
self._state_cache = StateCache(state_cache_size)
if credentials is None:
logging.info('Creating insecure control channel for %s.',
control_address)
self._control_channel = GRPCChannelFactory.insecure_channel(
control_address)
else:
logging.info('Creating secure control channel for %s.',
control_address)
self._control_channel = GRPCChannelFactory.secure_channel(
control_address, credentials)
grpc.channel_ready_future(self._control_channel).result(timeout=60)
logging.info('Control channel established.')
self._control_channel = grpc.intercept_channel(
self._control_channel, WorkerIdInterceptor(self._worker_id))
self._data_channel_factory = data_plane.GrpcClientDataChannelFactory(
credentials, self._worker_id)
self._state_handler_factory = GrpcStateHandlerFactory(
self._state_cache, credentials)
self._profiler_factory = profiler_factory
self._fns = {}
# BundleProcessor cache across all workers.
self._bundle_processor_cache = BundleProcessorCache(
state_handler_factory=self._state_handler_factory,
data_channel_factory=self._data_channel_factory,
fns=self._fns)
# workers for process/finalize bundle.
self.workers = queue.Queue()
# one worker for progress/split request.
self.progress_worker = SdkWorker(
self._bundle_processor_cache,
profiler_factory=self._profiler_factory)
# one thread is enough for getting the progress report.
# Assumption:
# Progress report generation should not do IO or wait on other resources.
# Without wait, having multiple threads will not improve performance and
# will only add complexity.
self._progress_thread_pool = futures.ThreadPoolExecutor(max_workers=1)
# finalize and process share one thread pool.
self._process_thread_pool = futures.ThreadPoolExecutor(
max_workers=self._worker_count)
self._responses = queue.Queue()
self._process_bundle_queue = queue.Queue()
self._unscheduled_process_bundle = {}
logging.info('Initializing SDKHarness with %s workers.',
self._worker_count)
def __init__(self,
control_address, # type: str
credentials=None,
worker_id=None, # type: Optional[str]
# Caching is disabled by default
state_cache_size=0,
# time-based data buffering is disabled by default
data_buffer_time_limit_ms=0,
profiler_factory=None, # type: Optional[Callable[..., Profile]]
status_address=None, # type: Optional[str, unicode]
):
self._alive = True
self._worker_index = 0
self._worker_id = worker_id
self._state_cache = StateCache(state_cache_size)
if credentials is None:
_LOGGER.info('Creating insecure control channel for %s.', control_address)
self._control_channel = GRPCChannelFactory.insecure_channel(
control_address)
else:
_LOGGER.info('Creating secure control channel for %s.', control_address)
self._control_channel = GRPCChannelFactory.secure_channel(
control_address, credentials)
grpc.channel_ready_future(self._control_channel).result(timeout=60)
_LOGGER.info('Control channel established.')
self._control_channel = grpc.intercept_channel(
self._control_channel, WorkerIdInterceptor(self._worker_id))
self._data_channel_factory = data_plane.GrpcClientDataChannelFactory(
credentials, self._worker_id, data_buffer_time_limit_ms)
self._state_handler_factory = GrpcStateHandlerFactory(
self._state_cache, credentials)
self._profiler_factory = profiler_factory
self._fns = {} # type: Dict[str, beam_fn_api_pb2.ProcessBundleDescriptor]
# BundleProcessor cache across all workers.
self._bundle_processor_cache = BundleProcessorCache(
state_handler_factory=self._state_handler_factory,
data_channel_factory=self._data_channel_factory,
fns=self._fns)
if status_address:
try:
self._status_handler = FnApiWorkerStatusHandler(
status_address, self._bundle_processor_cache)
except Exception:
traceback_string = traceback.format_exc()
_LOGGER.warning(
'Error creating worker status request handler, '
'skipping status report. Trace back: %s' % traceback_string)
else:
self._status_handler = None
# TODO(BEAM-8998) use common UnboundedThreadPoolExecutor to process bundle
# progress once dataflow runner's excessive progress polling is removed.
self._report_progress_executor = futures.ThreadPoolExecutor(max_workers=1)
self._worker_thread_pool = UnboundedThreadPoolExecutor()
self._responses = queue.Queue(
) # type: queue.Queue[beam_fn_api_pb2.InstructionResponse]
_LOGGER.info('Initializing SDKHarness with unbounded number of workers.')
def __init__(
self,
control_address,
credentials=None,
worker_id=None,
# Caching is disabled by default
state_cache_size=0,
profiler_factory=None):
self._alive = True
self._worker_index = 0
self._worker_id = worker_id
self._state_cache = StateCache(state_cache_size)
if credentials is None:
_LOGGER.info('Creating insecure control channel for %s.',
control_address)
self._control_channel = GRPCChannelFactory.insecure_channel(
control_address)
else:
_LOGGER.info('Creating secure control channel for %s.',
control_address)
self._control_channel = GRPCChannelFactory.secure_channel(
control_address, credentials)
grpc.channel_ready_future(self._control_channel).result(timeout=60)
_LOGGER.info('Control channel established.')
self._control_channel = grpc.intercept_channel(
self._control_channel, WorkerIdInterceptor(self._worker_id))
self._data_channel_factory = data_plane.GrpcClientDataChannelFactory(
credentials, self._worker_id)
self._state_handler_factory = GrpcStateHandlerFactory(
self._state_cache, credentials)
self._profiler_factory = profiler_factory
self._fns = {}
# BundleProcessor cache across all workers.
self._bundle_processor_cache = BundleProcessorCache(
state_handler_factory=self._state_handler_factory,
data_channel_factory=self._data_channel_factory,
fns=self._fns)
# TODO(BEAM-8998) use common UnboundedThreadPoolExecutor to process bundle
# progress once dataflow runner's excessive progress polling is removed.
self._report_progress_executor = futures.ThreadPoolExecutor(
max_workers=1)
self._worker_thread_pool = UnboundedThreadPoolExecutor()
self._responses = queue.Queue()
_LOGGER.info(
'Initializing SDKHarness with unbounded number of workers.')
def __init__(self, control_address, worker_count):
self._worker_count = worker_count
self._worker_index = 0
self._control_channel = grpc.insecure_channel(control_address)
self._data_channel_factory = data_plane.GrpcClientDataChannelFactory()
self.workers = queue.Queue()
# one thread is enough for getting the progress report.
# Assumption:
# Progress report generation should not do IO or wait on other resources.
# Without wait, having multiple threads will not improve performance and
# will only add complexity.
self._progress_thread_pool = futures.ThreadPoolExecutor(max_workers=1)
self._process_thread_pool = futures.ThreadPoolExecutor(
max_workers=self._worker_count)
self._instruction_id_vs_worker = {}
self._fns = {}
self._responses = queue.Queue()
self._process_bundle_queue = queue.Queue()
logging.info('Initializing SDKHarness with %s workers.', self._worker_count)
def __init__(self,
control_address, # type: str
credentials=None,
worker_id=None, # type: Optional[str]
# Caching is disabled by default
state_cache_size=0,
profiler_factory=None # type: Optional[Callable[..., Profile]]
):
self._alive = True
self._worker_index = 0
self._worker_id = worker_id
self._state_cache = StateCache(state_cache_size)
if credentials is None:
_LOGGER.info('Creating insecure control channel for %s.', control_address)
self._control_channel = GRPCChannelFactory.insecure_channel(
control_address)
else:
_LOGGER.info('Creating secure control channel for %s.', control_address)
self._control_channel = GRPCChannelFactory.secure_channel(
control_address, credentials)
grpc.channel_ready_future(self._control_channel).result(timeout=60)
_LOGGER.info('Control channel established.')
self._control_channel = grpc.intercept_channel(
self._control_channel, WorkerIdInterceptor(self._worker_id))
self._data_channel_factory = data_plane.GrpcClientDataChannelFactory(
credentials, self._worker_id)
self._state_handler_factory = GrpcStateHandlerFactory(self._state_cache,
credentials)
self._profiler_factory = profiler_factory
self._fns = {} # type: Dict[str, beam_fn_api_pb2.ProcessBundleDescriptor]
# BundleProcessor cache across all workers.
self._bundle_processor_cache = BundleProcessorCache(
state_handler_factory=self._state_handler_factory,
data_channel_factory=self._data_channel_factory,
fns=self._fns)
self._worker_thread_pool = UnboundedThreadPoolExecutor()
self._responses = queue.Queue() # type: queue.Queue[beam_fn_api_pb2.InstructionResponse]
_LOGGER.info('Initializing SDKHarness with unbounded number of workers.')
def __init__(
self,
control_address, # type: str
credentials=None, # type: Optional[grpc.ChannelCredentials]
worker_id=None, # type: Optional[str]
# Caching is disabled by default
state_cache_size=0, # type: int
# time-based data buffering is disabled by default
data_buffer_time_limit_ms=0, # type: int
profiler_factory=None, # type: Optional[Callable[..., Profile]]
status_address=None, # type: Optional[str]
# Heap dump through status api is disabled by default
enable_heap_dump=False, # type: bool
):
# type: (...) -> None
self._alive = True
self._worker_index = 0
self._worker_id = worker_id
self._state_cache = StateCache(state_cache_size)
options = [('grpc.max_receive_message_length', -1),
('grpc.max_send_message_length', -1)]
if credentials is None:
_LOGGER.info('Creating insecure control channel for %s.', control_address)
self._control_channel = GRPCChannelFactory.insecure_channel(
control_address, options=options)
else:
_LOGGER.info('Creating secure control channel for %s.', control_address)
self._control_channel = GRPCChannelFactory.secure_channel(
control_address, credentials, options=options)
grpc.channel_ready_future(self._control_channel).result(timeout=60)
_LOGGER.info('Control channel established.')
self._control_channel = grpc.intercept_channel(
self._control_channel, WorkerIdInterceptor(self._worker_id))
self._data_channel_factory = data_plane.GrpcClientDataChannelFactory(
credentials, self._worker_id, data_buffer_time_limit_ms)
self._state_handler_factory = GrpcStateHandlerFactory(
self._state_cache, credentials)
self._profiler_factory = profiler_factory
def default_factory(id):
# type: (str) -> beam_fn_api_pb2.ProcessBundleDescriptor
return self._control_stub.GetProcessBundleDescriptor(
beam_fn_api_pb2.GetProcessBundleDescriptorRequest(
process_bundle_descriptor_id=id))
self._fns = KeyedDefaultDict(default_factory)
# BundleProcessor cache across all workers.
self._bundle_processor_cache = BundleProcessorCache(
state_handler_factory=self._state_handler_factory,
data_channel_factory=self._data_channel_factory,
fns=self._fns)
if status_address:
try:
self._status_handler = FnApiWorkerStatusHandler(
status_address, self._bundle_processor_cache,
enable_heap_dump) # type: Optional[FnApiWorkerStatusHandler]
except Exception:
traceback_string = traceback.format_exc()
_LOGGER.warning(
'Error creating worker status request handler, '
'skipping status report. Trace back: %s' % traceback_string)
else:
self._status_handler = None
# TODO(BEAM-8998) use common
# thread_pool_executor.shared_unbounded_instance() to process bundle
# progress once dataflow runner's excessive progress polling is removed.
self._report_progress_executor = futures.ThreadPoolExecutor(max_workers=1)
self._worker_thread_pool = thread_pool_executor.shared_unbounded_instance()
self._responses = queue.Queue(
) # type: queue.Queue[Union[beam_fn_api_pb2.InstructionResponse, Sentinel]]
_LOGGER.info('Initializing SDKHarness with unbounded number of workers.')
def __init__(self, control_address):
self._control_channel = grpc.insecure_channel(control_address)
self._data_channel_factory = data_plane.GrpcClientDataChannelFactory()
def __init__(self, control_channel):
self._control_channel = control_channel
self._data_channel_factory = data_plane.GrpcClientDataChannelFactory()