def sync_launch_scheduled_execution(schedule_origin, system_tz=None):
check.inst_param(schedule_origin, "schedule_origin", ExternalJobOrigin)
with get_temp_file_name() as output_file:
parts = (
[sys.executable, "-m", "dagster", "api", "launch_scheduled_execution", output_file,]
+ xplat_shlex_split(schedule_origin.get_repo_cli_args())
+ ["--schedule_name={}".format(schedule_origin.job_name)]
+ (["--override-system-timezone={}".format(system_tz)] if system_tz else [])
)
subprocess.check_call(parts)
result = read_unary_response(output_file)
if isinstance(result, ScheduledExecutionResult):
return result
elif isinstance(result, IPCErrorMessage):
error = result.serializable_error_info
raise DagsterSubprocessError(
"Error in API subprocess: {message}\n\n{err}".format(
message=result.message, err=error.to_string()
),
subprocess_error_infos=[error],
)
else:
check.failed("Unexpected result {}".format(result))
def sync_launch_scheduled_execution(schedule_origin):
check.inst_param(schedule_origin, 'schedule_origin', SchedulePythonOrigin)
with get_temp_file_name() as output_file:
parts = ([
schedule_origin.executable_path,
'-m',
'dagster',
'api',
'launch_scheduled_execution',
output_file,
] + xplat_shlex_split(schedule_origin.get_repo_cli_args()) + [
'--schedule_name={}'.format(schedule_origin.schedule_name),
])
execute_command_in_subprocess(parts)
result = read_unary_response(output_file)
if isinstance(result, ScheduledExecutionResult):
return result
elif isinstance(result, IPCErrorMessage):
error = result.serializable_error_info
raise DagsterSubprocessError(
'Error in API subprocess: {message}\n\n{err}'.format(
message=result.message, err=error.to_string()),
subprocess_error_infos=[error],
)
else:
check.failed('Unexpected result {}'.format(result))
def sync_code():
# Sync remote dagster packages with local dagster code
sync_code_command = [
'rsync',
'-av',
'-progress',
"--exclude='scala_modules/'",
"--exclude='js_modules/'",
"--exclude='.git/'",
"--exclude='docs/'",
'-e',
'"ssh -i {aws_emr_pem_file}"'.format(aws_emr_pem_file=os.environ['AWS_EMR_PEM_FILE']),
os.environ['DAGSTER_DIR'],
os.environ['AWS_EMR_NODE_ADDRESS'] + ':~/',
]
if (
subprocess.call(
' '.join(sync_code_command), stdout=sys.stdout, stderr=sys.stderr, shell=True
)
!= 0
):
raise DagsterSubprocessError('Failed to sync code to EMR')
# Install dagster packages on remote node
remote_install_dagster_packages_command = ['sudo', 'python3', '-m', 'pip', 'install'] + [
token
for package_subpath in ['dagster', 'libraries/dagster-pyspark']
for token in ['-e', '/home/hadoop/dagster/python_modules/' + package_subpath]
]
install_dagster_packages_command = [
'ssh',
'-i',
os.environ['AWS_EMR_PEM_FILE'],
os.environ['AWS_EMR_NODE_ADDRESS'],
"'" + ' '.join(remote_install_dagster_packages_command) + "'",
]
if (
subprocess.call(
' '.join(install_dagster_packages_command),
stdout=sys.stdout,
stderr=sys.stderr,
shell=True,
)
!= 0
):
raise DagsterSubprocessError('Failed to install dagster packages on EMR')
def sync_code():
# Sync remote dagster packages with local dagster code
sync_code_command = [
"rsync",
"-av",
"-progress",
"--exclude='scala_modules/'",
"--exclude='js_modules/'",
"--exclude='.git/'",
"--exclude='docs/'",
"-e",
'"ssh -i {aws_emr_pem_file}"'.format(
aws_emr_pem_file=os.environ["AWS_EMR_PEM_FILE"]),
os.environ["DAGSTER_DIR"],
os.environ["AWS_EMR_NODE_ADDRESS"] + ":~/",
]
if (subprocess.call(" ".join(sync_code_command),
stdout=sys.stdout,
stderr=sys.stderr,
shell=True) != 0):
raise DagsterSubprocessError("Failed to sync code to EMR")
# Install dagster packages on remote node
remote_install_dagster_packages_command = [
"sudo", "python3", "-m", "pip", "install"
] + [
token for package_subpath in ["dagster", "libraries/dagster-pyspark"]
for token in
["-e", "/home/hadoop/dagster/python_modules/" + package_subpath]
]
install_dagster_packages_command = [
"ssh",
"-i",
os.environ["AWS_EMR_PEM_FILE"],
os.environ["AWS_EMR_NODE_ADDRESS"],
"'" + " ".join(remote_install_dagster_packages_command) + "'",
]
if (subprocess.call(
" ".join(install_dagster_packages_command),
stdout=sys.stdout,
stderr=sys.stderr,
shell=True,
) != 0):
raise DagsterSubprocessError(
"Failed to install dagster packages on EMR")
def bounded_parallel_executor(pipeline_context, step_contexts, limit):
pending_execution = list(step_contexts)
active_iters = {}
errors = {}
term_events = {}
stopping = False
while (not stopping and pending_execution) or active_iters:
try:
while len(active_iters
) < limit and pending_execution and not stopping:
step_context = pending_execution.pop(0)
step = step_context.step
term_events[step.key] = get_multiprocessing_context().Event()
active_iters[step.key] = execute_step_out_of_process(
step_context, step, errors, term_events)
empty_iters = []
for key, step_iter in active_iters.items():
try:
event_or_none = next(step_iter)
if event_or_none is None:
continue
else:
yield event_or_none
except StopIteration:
empty_iters.append(key)
for key in empty_iters:
del active_iters[key]
if term_events[key].is_set():
stopping = True
del term_events[key]
# In the very small chance that we get interrupted in this coordination section and not
# polling the subprocesses for events - try to clean up greacefully
except KeyboardInterrupt:
yield DagsterEvent.engine_event(
pipeline_context,
'Multiprocess engine: received KeyboardInterrupt - forwarding to active child processes',
EngineEventData.interrupted(list(term_events.keys())),
)
for event in term_events.values():
event.set()
errs = {pid: err for pid, err in errors.items() if err}
if errs:
raise DagsterSubprocessError(
'During multiprocess execution errors occured in child processes:\n{error_list}'
.format(error_list='\n'.join([
'In process {pid}: {err}'.format(pid=pid, err=err.to_string())
for pid, err in errs.items()
])),
subprocess_error_infos=list(errs.values()),
)
def sync_get_external_execution_plan_grpc(
api_client,
pipeline_origin,
run_config,
mode,
pipeline_snapshot_id,
solid_selection=None,
step_keys_to_execute=None,
known_state=None,
):
from dagster.grpc.client import DagsterGrpcClient
check.inst_param(api_client, "api_client", DagsterGrpcClient)
check.inst_param(pipeline_origin, "pipeline_origin",
ExternalPipelineOrigin)
check.opt_list_param(solid_selection, "solid_selection", of_type=str)
check.dict_param(run_config, "run_config")
check.str_param(mode, "mode")
check.opt_list_param(step_keys_to_execute,
"step_keys_to_execute",
of_type=str)
check.str_param(pipeline_snapshot_id, "pipeline_snapshot_id")
check.opt_inst_param(known_state, "known_state", KnownExecutionState)
result = check.inst(
api_client.execution_plan_snapshot(
execution_plan_snapshot_args=ExecutionPlanSnapshotArgs(
pipeline_origin=pipeline_origin,
solid_selection=solid_selection,
run_config=run_config,
mode=mode,
step_keys_to_execute=step_keys_to_execute,
pipeline_snapshot_id=pipeline_snapshot_id,
known_state=known_state,
)),
(ExecutionPlanSnapshot, ExecutionPlanSnapshotErrorData),
)
if isinstance(result, ExecutionPlanSnapshotErrorData):
raise DagsterSubprocessError(result.error.to_string(),
subprocess_error_infos=[result.error])
return result
def bounded_parallel_executor(step_contexts, limit):
pending_execution = list(step_contexts)
active_iters = {}
pid_tracker = {}
while pending_execution or active_iters:
while len(active_iters) < limit and pending_execution:
step_context = pending_execution.pop()
step = step_context.step
active_iters[step.key] = execute_step_out_of_process(step_context, step, pid_tracker)
empty_iters = []
for key, step_iter in active_iters.items():
try:
event_or_none = next(step_iter)
if event_or_none is None:
continue
else:
yield event_or_none
except StopIteration:
empty_iters.append(key)
for key in empty_iters:
del active_iters[key]
errs = {pid: err for pid, err in pid_tracker.items() if err}
if errs:
raise DagsterSubprocessError(
'During multiprocess execution errors occured in child processes:\n{error_list}'.format(
error_list='\n'.join(
[
'In process {pid}: {err}'.format(pid=pid, err=err.to_string())
for pid, err in errs.items()
]
)
),
subprocess_error_infos=list(errs.values()),
)
def sync_get_external_execution_plan(
pipeline_origin,
run_config,
mode,
pipeline_snapshot_id,
solid_selection=None,
step_keys_to_execute=None,
):
check.inst_param(pipeline_origin, "pipeline_origin", PipelinePythonOrigin)
check.opt_list_param(solid_selection, "solid_selection", of_type=str)
check.dict_param(run_config, "run_config")
check.str_param(mode, "mode")
check.opt_list_param(step_keys_to_execute,
"step_keys_to_execute",
of_type=str)
check.str_param(pipeline_snapshot_id, "pipeline_snapshot_id")
result = check.inst(
execute_unary_api_cli_command(
pipeline_origin.executable_path,
"execution_plan",
ExecutionPlanSnapshotArgs(
pipeline_origin=pipeline_origin,
solid_selection=solid_selection,
run_config=run_config,
mode=mode,
step_keys_to_execute=step_keys_to_execute,
pipeline_snapshot_id=pipeline_snapshot_id,
),
),
(ExecutionPlanSnapshot, ExecutionPlanSnapshotErrorData),
)
if isinstance(result, ExecutionPlanSnapshotErrorData):
raise DagsterSubprocessError(result.error.to_string(),
subprocess_error_infos=[result.error])
return result
def sync_get_external_execution_plan_grpc(
api_client,
pipeline_origin,
run_config,
mode,
pipeline_snapshot_id,
solid_selection=None,
step_keys_to_execute=None,
):
from dagster.grpc.client import DagsterGrpcClient
check.inst_param(api_client, 'api_client', DagsterGrpcClient)
check.inst_param(pipeline_origin, 'pipeline_origin', PipelineOrigin)
check.opt_list_param(solid_selection, 'solid_selection', of_type=str)
check.dict_param(run_config, 'run_config')
check.str_param(mode, 'mode')
check.opt_list_param(step_keys_to_execute,
'step_keys_to_execute',
of_type=str)
check.str_param(pipeline_snapshot_id, 'pipeline_snapshot_id')
result = check.inst(
api_client.execution_plan_snapshot(
execution_plan_snapshot_args=ExecutionPlanSnapshotArgs(
pipeline_origin=pipeline_origin,
solid_selection=solid_selection,
run_config=run_config,
mode=mode,
step_keys_to_execute=step_keys_to_execute,
pipeline_snapshot_id=pipeline_snapshot_id,
)),
(ExecutionPlanSnapshot, ExecutionPlanSnapshotErrorData),
)
if isinstance(result, ExecutionPlanSnapshotErrorData):
raise DagsterSubprocessError(result.error.to_string(),
subprocess_error_infos=[result.error])
return result
def execute(self, pipeline_context, execution_plan):
check.inst_param(pipeline_context, "pipeline_context",
SystemPipelineExecutionContext)
check.inst_param(execution_plan, "execution_plan", ExecutionPlan)
limit = self.max_concurrent
yield DagsterEvent.engine_event(
pipeline_context,
"Executing steps using multiprocess engine: parent process (pid: {pid})"
.format(pid=os.getpid()),
event_specific_data=EngineEventData.multiprocess(
os.getpid(),
step_keys_to_execute=execution_plan.step_keys_to_execute),
)
# It would be good to implement a reference tracking algorithm here so we could
# garbage collection results that are no longer needed by any steps
# https://github.com/dagster-io/dagster/issues/811
with time_execution_scope() as timer_result:
with execution_plan.start(
retries=self.retries) as active_execution:
active_iters = {}
errors = {}
term_events = {}
stopping = False
while (not stopping
and not active_execution.is_complete) or active_iters:
try:
# start iterators
while len(active_iters) < limit and not stopping:
steps = active_execution.get_steps_to_execute(
limit=(limit - len(active_iters)))
if not steps:
break
for step in steps:
step_context = pipeline_context.for_step(step)
term_events[step.key] = multiprocessing.Event()
active_iters[
step.
key] = self.execute_step_out_of_process(
step_context, step, errors,
term_events)
# process active iterators
empty_iters = []
for key, step_iter in active_iters.items():
try:
event_or_none = next(step_iter)
if event_or_none is None:
continue
else:
yield event_or_none
active_execution.handle_event(
event_or_none)
except ChildProcessCrashException as crash:
serializable_error = serializable_error_info_from_exc_info(
sys.exc_info())
yield DagsterEvent.engine_event(
pipeline_context,
("Multiprocess executor: child process for step {step_key} "
"unexpectedly exited with code {exit_code}"
).format(step_key=key,
exit_code=crash.exit_code),
EngineEventData.engine_error(
serializable_error),
step_key=key,
)
step_failure_event = DagsterEvent.step_failure_event(
step_context=pipeline_context.for_step(
active_execution.get_step_by_key(key)),
step_failure_data=StepFailureData(
error=serializable_error,
user_failure_data=None),
)
active_execution.handle_event(
step_failure_event)
yield step_failure_event
empty_iters.append(key)
except StopIteration:
empty_iters.append(key)
# clear and mark complete finished iterators
for key in empty_iters:
del active_iters[key]
if term_events[key].is_set():
stopping = True
del term_events[key]
active_execution.verify_complete(
pipeline_context, key)
# process skips from failures or uncovered inputs
for event in active_execution.skipped_step_events_iterator(
pipeline_context):
yield event
# In the very small chance that we get interrupted in this coordination section and not
# polling the subprocesses for events - try to clean up gracefully
except KeyboardInterrupt:
yield DagsterEvent.engine_event(
pipeline_context,
"Multiprocess engine: received KeyboardInterrupt - forwarding to active child processes",
EngineEventData.interrupted(
list(term_events.keys())),
)
stopping = True
for event in term_events.values():
event.set()
errs = {pid: err for pid, err in errors.items() if err}
if errs:
raise DagsterSubprocessError(
"During multiprocess execution errors occurred in child processes:\n{error_list}"
.format(error_list="\n".join([
"In process {pid}: {err}".format(
pid=pid, err=err.to_string())
for pid, err in errs.items()
])),
subprocess_error_infos=list(errs.values()),
)
yield DagsterEvent.engine_event(
pipeline_context,
"Multiprocess engine: parent process exiting after {duration} (pid: {pid})"
.format(duration=format_duration(timer_result.millis),
pid=os.getpid()),
event_specific_data=EngineEventData.multiprocess(os.getpid()),
)
def execute(pipeline_context, execution_plan):
check.inst_param(pipeline_context, 'pipeline_context',
SystemPipelineExecutionContext)
check.inst_param(execution_plan, 'execution_plan', ExecutionPlan)
check.param_invariant(
isinstance(pipeline_context.executor_config, CeleryConfig),
'pipeline_context',
'Expected executor_config to be CeleryConfig got {}'.format(
pipeline_context.executor_config),
)
celery_config = pipeline_context.executor_config
storage = pipeline_context.environment_dict.get('storage')
if (celery_config.broker and not is_local_uri(celery_config.broker)
) or (celery_config.backend
and not is_local_uri(celery_config.backend)):
check.invariant(
storage.get('s3') or storage.get('gcs'),
'Must use S3 or GCS storage with non-local Celery broker: {broker} '
'and backend: {backend}'.format(broker=celery_config.broker,
backend=celery_config.backend),
)
else:
check.invariant(
not storage.get('in_memory'),
'Cannot use in-memory storage with Celery, use filesystem, S3, or GCS',
)
pipeline_name = pipeline_context.pipeline_def.name
handle_dict = pipeline_context.execution_target_handle.to_dict()
instance_ref_dict = pipeline_context.instance.get_ref().to_dict()
environment_dict = dict(pipeline_context.environment_dict,
execution={'in_process': {}})
mode = pipeline_context.mode_def.name
run_id = pipeline_context.pipeline_run.run_id
app = make_app(celery_config)
task_signatures = {} # Dict[step_key, celery.Signature]
apply_kwargs = defaultdict(dict) # Dict[step_key, Dict[str, Any]]
priority_for_step = lambda step: (-1 * int(
step.tags.get('dagster-celery/priority', task_default_priority)))
priority_for_key = lambda step_key: (-1 * apply_kwargs[step_key][
'priority'])
_warn_on_priority_misuse(pipeline_context, execution_plan)
for step_key in execution_plan.step_keys_to_execute:
step = execution_plan.get_step_by_key(step_key)
priority = int(
step.tags.get('dagster-celery/priority',
task_default_priority))
queue = step.tags.get('dagster-celery/queue', task_default_queue)
task = create_task(app)
variables = {
'executionParams': {
'selector': {
'name': pipeline_name
},
'environmentConfigData': environment_dict,
'mode': mode,
'executionMetadata': {
'runId': run_id
},
'stepKeys': [step_key],
}
}
task_signatures[step_key] = task.si(handle_dict, variables,
instance_ref_dict)
apply_kwargs[step_key] = {
'priority': priority,
'queue': queue,
'routing_key': '{queue}.execute_query'.format(queue=queue),
}
step_results = {} # Dict[ExecutionStep, celery.AsyncResult]
step_success = {}
step_errors = {}
completed_steps = set({}) # Set[step_key]
active_execution = execution_plan.start(sort_key_fn=priority_for_step)
stopping = False
while (not active_execution.is_complete
and not stopping) or step_results:
results_to_pop = []
for step_key, result in sorted(
step_results.items(),
key=lambda x: priority_for_key(x[0])):
if result.ready():
try:
step_events = result.get()
except Exception as e: # pylint: disable=broad-except
# We will want to do more to handle the exception here.. maybe subclass Task
# Certainly yield an engine or pipeline event
step_events = []
step_errors[
step_key] = serializable_error_info_from_exc_info(
sys.exc_info())
stopping = True
for step_event in step_events:
event = deserialize_json_to_dagster_namedtuple(
step_event)
yield event
if event.is_step_success:
step_success[step_key] = True
elif event.is_step_failure:
step_success[step_key] = False
results_to_pop.append(step_key)
completed_steps.add(step_key)
for step_key in results_to_pop:
if step_key in step_results:
del step_results[step_key]
was_success = step_success.get(step_key)
if was_success == True:
active_execution.mark_success(step_key)
elif was_success == False:
active_execution.mark_failed(step_key)
else:
# check errors list?
pipeline_context.log.error(
'Step {key} finished without success or failure event, assuming failure.'
.format(key=step_key))
active_execution.mark_failed(step_key)
# process skips from failures or uncovered inputs
for event in active_execution.skipped_step_events_iterator(
pipeline_context):
yield event
# dont add any new steps if we are stopping
if stopping:
continue
# This is a slight refinement. If we have n workers idle and schedule m > n steps for
# execution, the first n steps will be picked up by the idle workers in the order in
# which they are scheduled (and the following m-n steps will be executed in priority
# order, provided that it takes longer to execute a step than to schedule it). The test
# case has m >> n to exhibit this behavior in the absence of this sort step.
for step in active_execution.get_steps_to_execute():
try:
step_results[step.key] = task_signatures[
step.key].apply_async(**apply_kwargs[step.key])
except Exception:
yield DagsterEvent.engine_event(
pipeline_context,
'Encountered error during celery task submission.'.
format(),
event_specific_data=EngineEventData.engine_error(
serializable_error_info_from_exc_info(
sys.exc_info()), ),
)
raise
time.sleep(TICK_SECONDS)
if step_errors:
raise DagsterSubprocessError(
'During celery execution errors occured in workers:\n{error_list}'
.format(error_list='\n'.join([
'[{step}]: {err}'.format(step=key, err=err.to_string())
for key, err in step_errors.items()
])),
subprocess_error_infos=list(step_errors.values()),
)
def execute(pipeline_context, execution_plan):
check.inst_param(pipeline_context, 'pipeline_context',
SystemPipelineExecutionContext)
check.inst_param(execution_plan, 'execution_plan', ExecutionPlan)
intermediates_manager = pipeline_context.intermediates_manager
limit = pipeline_context.executor_config.max_concurrent
yield DagsterEvent.engine_event(
pipeline_context,
'Executing steps using multiprocess engine: parent process (pid: {pid})'
.format(pid=os.getpid()),
event_specific_data=EngineEventData.multiprocess(
os.getpid(),
step_keys_to_execute=execution_plan.step_keys_to_execute),
)
# It would be good to implement a reference tracking algorithm here so we could
# garbage collection results that are no longer needed by any steps
# https://github.com/dagster-io/dagster/issues/811
with time_execution_scope() as timer_result:
active_execution = execution_plan.start(
retries=pipeline_context.executor_config.retries)
active_iters = {}
errors = {}
term_events = {}
stopping = False
while (not stopping
and not active_execution.is_complete) or active_iters:
try:
# start iterators
while len(active_iters) < limit and not stopping:
steps = active_execution.get_steps_to_execute(
limit=(limit - len(active_iters)))
if not steps:
break
for step in steps:
step_context = pipeline_context.for_step(step)
term_events[
step.key] = get_multiprocessing_context(
).Event()
active_iters[
step.key] = execute_step_out_of_process(
step_context, step, errors, term_events)
# process active iterators
empty_iters = []
for key, step_iter in active_iters.items():
try:
event_or_none = next(step_iter)
if event_or_none is None:
continue
else:
yield event_or_none
active_execution.handle_event(event_or_none)
except StopIteration:
empty_iters.append(key)
# clear and mark complete finished iterators
for key in empty_iters:
del active_iters[key]
if term_events[key].is_set():
stopping = True
del term_events[key]
active_execution.verify_complete(pipeline_context, key)
# process skips from failures or uncovered inputs
for event in active_execution.skipped_step_events_iterator(
pipeline_context):
yield event
# In the very small chance that we get interrupted in this coordination section and not
# polling the subprocesses for events - try to clean up gracefully
except KeyboardInterrupt:
yield DagsterEvent.engine_event(
pipeline_context,
'Multiprocess engine: received KeyboardInterrupt - forwarding to active child processes',
EngineEventData.interrupted(list(term_events.keys())),
)
stopping = True
for event in term_events.values():
event.set()
errs = {pid: err for pid, err in errors.items() if err}
if errs:
raise DagsterSubprocessError(
'During multiprocess execution errors occurred in child processes:\n{error_list}'
.format(error_list='\n'.join([
'In process {pid}: {err}'.format(pid=pid,
err=err.to_string())
for pid, err in errs.items()
])),
subprocess_error_infos=list(errs.values()),
)
yield DagsterEvent.engine_event(
pipeline_context,
'Multiprocess engine: parent process exiting after {duration} (pid: {pid})'
.format(duration=format_duration(timer_result.millis),
pid=os.getpid()),
event_specific_data=EngineEventData.multiprocess(os.getpid()),
)
def execute(self, pipeline_context, execution_plan):
check.inst_param(pipeline_context, "pipeline_context",
SystemPipelineExecutionContext)
check.inst_param(execution_plan, "execution_plan", ExecutionPlan)
limit = self.max_concurrent
yield DagsterEvent.engine_event(
pipeline_context,
"Executing steps using multiprocess executor: parent process (pid: {pid})"
.format(pid=os.getpid()),
event_specific_data=EngineEventData.multiprocess(
os.getpid(),
step_keys_to_execute=execution_plan.step_keys_to_execute),
)
# It would be good to implement a reference tracking algorithm here so we could
# garbage collect results that are no longer needed by any steps
# https://github.com/dagster-io/dagster/issues/811
with time_execution_scope() as timer_result:
with execution_plan.start(
retry_mode=self.retries) as active_execution:
active_iters = {}
errors = {}
term_events = {}
stopping = False
while (not stopping
and not active_execution.is_complete) or active_iters:
if active_execution.check_for_interrupts():
yield DagsterEvent.engine_event(
pipeline_context,
"Multiprocess executor: received termination signal - "
"forwarding to active child processes",
EngineEventData.interrupted(
list(term_events.keys())),
)
stopping = True
active_execution.mark_interrupted()
for key, event in term_events.items():
event.set()
# start iterators
while len(active_iters) < limit and not stopping:
steps = active_execution.get_steps_to_execute(
limit=(limit - len(active_iters)))
if not steps:
break
for step in steps:
step_context = pipeline_context.for_step(step)
term_events[step.key] = multiprocessing.Event()
active_iters[
step.key] = self.execute_step_out_of_process(
step_context,
step,
errors,
term_events,
active_execution.get_known_state(),
)
# process active iterators
empty_iters = []
for key, step_iter in active_iters.items():
try:
event_or_none = next(step_iter)
if event_or_none is None:
continue
else:
yield event_or_none
active_execution.handle_event(event_or_none)
except ChildProcessCrashException as crash:
serializable_error = serializable_error_info_from_exc_info(
sys.exc_info())
yield DagsterEvent.engine_event(
pipeline_context,
("Multiprocess executor: child process for step {step_key} "
"unexpectedly exited with code {exit_code}"
).format(step_key=key,
exit_code=crash.exit_code),
EngineEventData.engine_error(
serializable_error),
step_handle=active_execution.get_step_by_key(
key).handle,
)
step_failure_event = DagsterEvent.step_failure_event(
step_context=pipeline_context.for_step(
active_execution.get_step_by_key(key)),
step_failure_data=StepFailureData(
error=serializable_error,
user_failure_data=None),
)
active_execution.handle_event(step_failure_event)
yield step_failure_event
empty_iters.append(key)
except StopIteration:
empty_iters.append(key)
# clear and mark complete finished iterators
for key in empty_iters:
del active_iters[key]
del term_events[key]
active_execution.verify_complete(pipeline_context, key)
# process skipped and abandoned steps
yield from active_execution.plan_events_iterator(
pipeline_context)
errs = {pid: err for pid, err in errors.items() if err}
# After termination starts, raise an interrupted exception once all subprocesses
# have finished cleaning up (and the only errors were from being interrupted)
if (stopping and (not active_iters) and all([
err_info.cls_name == "DagsterExecutionInterruptedError"
for err_info in errs.values()
])):
yield DagsterEvent.engine_event(
pipeline_context,
"Multiprocess executor: interrupted all active child processes",
event_specific_data=EngineEventData(),
)
raise DagsterExecutionInterruptedError()
elif errs:
raise DagsterSubprocessError(
"During multiprocess execution errors occurred in child processes:\n{error_list}"
.format(error_list="\n".join([
"In process {pid}: {err}".format(
pid=pid, err=err.to_string())
for pid, err in errs.items()
])),
subprocess_error_infos=list(errs.values()),
)
yield DagsterEvent.engine_event(
pipeline_context,
"Multiprocess executor: parent process exiting after {duration} (pid: {pid})"
.format(duration=format_duration(timer_result.millis),
pid=os.getpid()),
event_specific_data=EngineEventData.multiprocess(os.getpid()),
)
def core_celery_execution_loop(pipeline_context, execution_plan,
step_execution_fn):
check.inst_param(pipeline_context, "pipeline_context",
SystemPipelineExecutionContext)
check.inst_param(execution_plan, "execution_plan", ExecutionPlan)
check.callable_param(step_execution_fn, "step_execution_fn")
executor = pipeline_context.executor
# https://github.com/dagster-io/dagster/issues/2440
check.invariant(
execution_plan.artifacts_persisted,
"Cannot use in-memory storage with Celery, use filesystem (on top of NFS or "
"similar system that allows files to be available to all nodes), S3, or GCS",
)
app = make_app(executor.app_args())
priority_for_step = lambda step: (-1 * int(
step.tags.get(DAGSTER_CELERY_STEP_PRIORITY_TAG, task_default_priority)
) + -1 * _get_run_priority(pipeline_context))
priority_for_key = lambda step_key: (priority_for_step(
execution_plan.get_step_by_key(step_key)))
_warn_on_priority_misuse(pipeline_context, execution_plan)
step_results = {} # Dict[ExecutionStep, celery.AsyncResult]
step_errors = {}
with execution_plan.start(
retries=pipeline_context.executor.retries,
sort_key_fn=priority_for_step,
) as active_execution:
stopping = False
while (not active_execution.is_complete
and not stopping) or step_results:
if active_execution.check_for_interrupts():
yield DagsterEvent.engine_event(
pipeline_context,
"Celery executor: received termination signal - revoking active tasks from workers",
EngineEventData.interrupted(list(step_results.keys())),
)
stopping = True
active_execution.mark_interrupted()
for result in step_results.values():
result.revoke()
results_to_pop = []
for step_key, result in sorted(
step_results.items(),
key=lambda x: priority_for_key(x[0])):
if result.ready():
try:
step_events = result.get()
except TaskRevokedError:
step_events = []
yield DagsterEvent.engine_event(
pipeline_context,
'celery task for running step "{step_key}" was revoked.'
.format(step_key=step_key, ),
EngineEventData(marker_end=DELEGATE_MARKER),
step_handle=active_execution.get_step_by_key(
step_key).handle,
)
except Exception: # pylint: disable=broad-except
# We will want to do more to handle the exception here.. maybe subclass Task
# Certainly yield an engine or pipeline event
step_events = []
step_errors[
step_key] = serializable_error_info_from_exc_info(
sys.exc_info())
for step_event in step_events:
event = deserialize_json_to_dagster_namedtuple(
step_event)
yield event
active_execution.handle_event(event)
results_to_pop.append(step_key)
for step_key in results_to_pop:
if step_key in step_results:
del step_results[step_key]
active_execution.verify_complete(pipeline_context,
step_key)
# process skips from failures or uncovered inputs
for event in active_execution.plan_events_iterator(
pipeline_context):
yield event
# don't add any new steps if we are stopping
if stopping or step_errors:
continue
# This is a slight refinement. If we have n workers idle and schedule m > n steps for
# execution, the first n steps will be picked up by the idle workers in the order in
# which they are scheduled (and the following m-n steps will be executed in priority
# order, provided that it takes longer to execute a step than to schedule it). The test
# case has m >> n to exhibit this behavior in the absence of this sort step.
for step in active_execution.get_steps_to_execute():
try:
queue = step.tags.get(DAGSTER_CELERY_QUEUE_TAG,
task_default_queue)
yield DagsterEvent.engine_event(
pipeline_context,
'Submitting celery task for step "{step_key}" to queue "{queue}".'
.format(step_key=step.key, queue=queue),
EngineEventData(marker_start=DELEGATE_MARKER),
step_handle=step.handle,
)
# Get the Celery priority for this step
priority = _get_step_priority(pipeline_context, step)
# Submit the Celery tasks
step_results[step.key] = step_execution_fn(
app, pipeline_context, step, queue, priority)
except Exception:
yield DagsterEvent.engine_event(
pipeline_context,
"Encountered error during celery task submission.".
format(),
event_specific_data=EngineEventData.engine_error(
serializable_error_info_from_exc_info(
sys.exc_info()), ),
)
raise
time.sleep(TICK_SECONDS)
if step_errors:
raise DagsterSubprocessError(
"During celery execution errors occurred in workers:\n{error_list}"
.format(error_list="\n".join([
"[{step}]: {err}".format(step=key, err=err.to_string())
for key, err in step_errors.items()
])),
subprocess_error_infos=list(step_errors.values()),
)
def _core_celery_execution_loop(pipeline_context, execution_plan, step_execution_fn):
from .tasks import make_app
check.inst_param(pipeline_context, 'pipeline_context', SystemPipelineExecutionContext)
check.inst_param(execution_plan, 'execution_plan', ExecutionPlan)
check.callable_param(step_execution_fn, 'step_execution_fn')
check.param_invariant(
isinstance(pipeline_context.executor_config, (CeleryConfig, CeleryK8sJobConfig)),
'pipeline_context',
'Expected executor_config to be Celery config got {}'.format(
pipeline_context.executor_config
),
)
celery_config = pipeline_context.executor_config
# https://github.com/dagster-io/dagster/issues/2440
check.invariant(
pipeline_context.system_storage_def.is_persistent,
'Cannot use in-memory storage with Celery, use filesystem (on top of NFS or '
'similar system that allows files to be available to all nodes), S3, or GCS',
)
app = make_app(celery_config)
priority_for_step = lambda step: (
-1 * int(step.tags.get(DAGSTER_CELERY_STEP_PRIORITY_TAG, task_default_priority))
+ -1 * _get_run_priority(pipeline_context)
)
priority_for_key = lambda step_key: (
priority_for_step(execution_plan.get_step_by_key(step_key))
)
_warn_on_priority_misuse(pipeline_context, execution_plan)
step_results = {} # Dict[ExecutionStep, celery.AsyncResult]
step_errors = {}
completed_steps = set({}) # Set[step_key]
active_execution = execution_plan.start(
retries=pipeline_context.executor_config.retries, sort_key_fn=priority_for_step
)
stopping = False
while (not active_execution.is_complete and not stopping) or step_results:
results_to_pop = []
for step_key, result in sorted(step_results.items(), key=lambda x: priority_for_key(x[0])):
if result.ready():
try:
step_events = result.get()
except Exception: # pylint: disable=broad-except
# We will want to do more to handle the exception here.. maybe subclass Task
# Certainly yield an engine or pipeline event
step_events = []
step_errors[step_key] = serializable_error_info_from_exc_info(sys.exc_info())
stopping = True
for step_event in step_events:
event = deserialize_json_to_dagster_namedtuple(step_event)
yield event
active_execution.handle_event(event)
results_to_pop.append(step_key)
completed_steps.add(step_key)
for step_key in results_to_pop:
if step_key in step_results:
del step_results[step_key]
active_execution.verify_complete(pipeline_context, step_key)
# process skips from failures or uncovered inputs
for event in active_execution.skipped_step_events_iterator(pipeline_context):
yield event
# don't add any new steps if we are stopping
if stopping:
continue
# This is a slight refinement. If we have n workers idle and schedule m > n steps for
# execution, the first n steps will be picked up by the idle workers in the order in
# which they are scheduled (and the following m-n steps will be executed in priority
# order, provided that it takes longer to execute a step than to schedule it). The test
# case has m >> n to exhibit this behavior in the absence of this sort step.
for step in active_execution.get_steps_to_execute():
try:
queue = step.tags.get(DAGSTER_CELERY_QUEUE_TAG, task_default_queue)
yield DagsterEvent.engine_event(
pipeline_context,
'Submitting celery task for step "{step_key}" to queue "{queue}".'.format(
step_key=step.key, queue=queue
),
EngineEventData(marker_start=DELEGATE_MARKER),
step_key=step.key,
)
# Get the Celery priority for this step
priority = _get_step_priority(pipeline_context, step)
# Submit the Celery tasks
step_results[step.key] = step_execution_fn(
app, pipeline_context, step, queue, priority
)
except Exception:
yield DagsterEvent.engine_event(
pipeline_context,
'Encountered error during celery task submission.'.format(),
event_specific_data=EngineEventData.engine_error(
serializable_error_info_from_exc_info(sys.exc_info()),
),
)
raise
time.sleep(TICK_SECONDS)
if step_errors:
raise DagsterSubprocessError(
'During celery execution errors occurred in workers:\n{error_list}'.format(
error_list='\n'.join(
[
'[{step}]: {err}'.format(step=key, err=err.to_string())
for key, err in step_errors.items()
]
)
),
subprocess_error_infos=list(step_errors.values()),
)
def execute(pipeline_context, execution_plan):
from .tasks import make_app
check.inst_param(pipeline_context, 'pipeline_context',
SystemPipelineExecutionContext)
check.inst_param(execution_plan, 'execution_plan', ExecutionPlan)
check.param_invariant(
isinstance(pipeline_context.executor_config, CeleryConfig),
'pipeline_context',
'Expected executor_config to be CeleryConfig got {}'.format(
pipeline_context.executor_config),
)
celery_config = pipeline_context.executor_config
storage = pipeline_context.environment_dict.get('storage')
if (celery_config.broker and not is_local_uri(celery_config.broker)
) or (celery_config.backend
and not is_local_uri(celery_config.backend)):
check.invariant(
storage.get('s3') or storage.get('gcs'),
'Must use S3 or GCS storage with non-local Celery broker: {broker} '
'and backend: {backend}'.format(broker=celery_config.broker,
backend=celery_config.backend),
)
else:
check.invariant(
not storage.get('in_memory'),
'Cannot use in-memory storage with Celery, use filesystem, S3, or GCS',
)
app = make_app(celery_config)
priority_for_step = lambda step: (-1 * int(
step.tags.get('dagster-celery/priority', task_default_priority)
) + -1 * _get_run_priority(pipeline_context))
priority_for_key = lambda step_key: (priority_for_step(
execution_plan.get_step_by_key(step_key)))
_warn_on_priority_misuse(pipeline_context, execution_plan)
step_results = {} # Dict[ExecutionStep, celery.AsyncResult]
step_errors = {}
completed_steps = set({}) # Set[step_key]
active_execution = execution_plan.start(
retries=pipeline_context.executor_config.retries,
sort_key_fn=priority_for_step)
stopping = False
while (not active_execution.is_complete
and not stopping) or step_results:
results_to_pop = []
for step_key, result in sorted(
step_results.items(),
key=lambda x: priority_for_key(x[0])):
if result.ready():
try:
step_events = result.get()
except Exception as e: # pylint: disable=broad-except
# We will want to do more to handle the exception here.. maybe subclass Task
# Certainly yield an engine or pipeline event
step_events = []
step_errors[
step_key] = serializable_error_info_from_exc_info(
sys.exc_info())
stopping = True
for step_event in step_events:
event = deserialize_json_to_dagster_namedtuple(
step_event)
yield event
active_execution.handle_event(event)
results_to_pop.append(step_key)
completed_steps.add(step_key)
for step_key in results_to_pop:
if step_key in step_results:
del step_results[step_key]
active_execution.verify_complete(pipeline_context,
step_key)
# process skips from failures or uncovered inputs
for event in active_execution.skipped_step_events_iterator(
pipeline_context):
yield event
# don't add any new steps if we are stopping
if stopping:
continue
# This is a slight refinement. If we have n workers idle and schedule m > n steps for
# execution, the first n steps will be picked up by the idle workers in the order in
# which they are scheduled (and the following m-n steps will be executed in priority
# order, provided that it takes longer to execute a step than to schedule it). The test
# case has m >> n to exhibit this behavior in the absence of this sort step.
for step in active_execution.get_steps_to_execute():
try:
queue = step.tags.get('dagster-celery/queue',
task_default_queue)
yield DagsterEvent.engine_event(
pipeline_context,
'Submitting celery task for step "{step_key}" to queue "{queue}".'
.format(step_key=step.key, queue=queue),
EngineEventData(marker_start=DELEGATE_MARKER),
step_key=step.key,
)
step_results[step.key] = _submit_task(
app, pipeline_context, step, queue)
except Exception:
yield DagsterEvent.engine_event(
pipeline_context,
'Encountered error during celery task submission.'.
format(),
event_specific_data=EngineEventData.engine_error(
serializable_error_info_from_exc_info(
sys.exc_info()), ),
)
raise
time.sleep(TICK_SECONDS)
if step_errors:
raise DagsterSubprocessError(
'During celery execution errors occurred in workers:\n{error_list}'
.format(error_list='\n'.join([
'[{step}]: {err}'.format(step=key, err=err.to_string())
for key, err in step_errors.items()
])),
subprocess_error_infos=list(step_errors.values()),
)
