def track_finished_pipelines():
"""when all nodes of a pipeline have been processed, then emit events"""
for running_pipeline, (start_time, running_children) \
in dict(running_pipelines).items(): # type: pipelines.Pipeline
if len(
set(running_pipeline.nodes.values())
& processed_nodes) == len(running_pipeline.nodes):
succeeded = running_pipeline not in failed_pipelines
event_queue.put(
events.Output(
node_path=running_pipeline.path(),
format=logger.Format.ITALICS,
is_error=not succeeded,
message=
f'{"succeeded" if succeeded else "failed"}, {logger.format_time_difference(run_start_time, datetime.datetime.now())}'
))
event_queue.put(
events.NodeFinished(
node_path=running_pipeline.path(),
start_time=start_time,
end_time=datetime.datetime.now(),
is_pipeline=True,
succeeded=succeeded))
del running_pipelines[running_pipeline]
processed_nodes.add(running_pipeline)
def log(message: str,
format: events.Output.Format = Format.STANDARD,
is_error: bool = False) -> None:
"""
Logs text messages.
When run inside a pipeline, this will send a log message to the parent process.
Otherwise, messages will be printed to `sys.stdout` and `sys.stderr`
Args:
message: The message to display
format: How to format the message
is_error: Whether the message is considered an error message
"""
message = message.rstrip()
if message:
if _event_queue:
_event_queue.put(
events.Output(_node_path, message, format, is_error))
elif is_error:
sys.stderr.write(message + '\n')
else:
sys.stdout.write(message + '\n')
def run():
# collect system stats in a separate Process
statistics_process = multiprocessing.Process(
target=lambda: system_statistics.generate_system_statistics(
event_queue),
name='system_statistics')
statistics_process.start()
try:
# capture output of print statements and other unplanned output
logger.redirect_output(event_queue, pipeline.path())
# all nodes that have not run yet, ordered by priority
node_queue: [pipelines.Node] = []
# data needed for computing cost
node_durations_and_run_times = node_cost.node_durations_and_run_times(
pipeline.path())
# Putting nodes into the node queue
def queue(nodes: [pipelines.Node]):
for node in nodes:
node_cost.compute_cost(node, node_durations_and_run_times)
node_queue.append(node)
node_queue.sort(key=lambda node: node.cost, reverse=True)
if nodes: # only run a set of child nodes
def with_all_upstreams(nodes: {pipelines.Node}):
"""recursively find all upstreams of a list of nodes"""
return functools.reduce(
set.union,
[with_all_upstreams(node.upstreams)
for node in nodes], nodes)
# when requested, include all upstreams of nodes, otherwise just use provided nodes
nodes_to_run = with_all_upstreams(
set(nodes)) if with_upstreams else set(nodes)
# remove everything from pipeline that should not be run
# (that's makes updating dependencies between nodes easier)
for node in set(pipeline.nodes.values()) - nodes_to_run:
pipeline.remove(node)
# queue remaining nodes
queue(list((pipeline.nodes).values()))
else:
# remove dependencies to siblings
pipeline.upstreams = set()
pipeline.downstreams = set()
# queue whole pipeline
queue([pipeline])
# book keeping
run_start_time = datetime.datetime.now()
# all nodes that already ran or that won't be run anymore
processed_nodes: {pipelines.Node} = set()
# running pipelines with start times and number of running children
running_pipelines: {
pipelines.Pipeline: [datetime.datetime, int]
} = {}
failed_pipelines: {pipelines.Pipeline
} = set() # pipelines with failed tasks
running_task_processes: {pipelines.Task: TaskProcess} = {}
def dequeue() -> pipelines.Node:
"""
Finds the next task in the queue
- without upstreams or where all upstreams have been run already
- where the pipeline specific maximum number of parallel tasks per pipeline is not reached
"""
for node in node_queue: # type: pipelines.Node
if ((not node.upstreams
or len(node.upstreams & processed_nodes) == len(
node.upstreams)) and
(not isinstance(node.parent, pipelines.Pipeline) or
(not node.parent.max_number_of_parallel_tasks) or
(not node.parent in running_pipelines) or
(running_pipelines[node.parent][1] <
node.parent.max_number_of_parallel_tasks))):
node_queue.remove(node)
if node.parent in failed_pipelines:
processed_nodes.add(
node
) # if the parent pipeline failed, don't launch new nodes
else:
return node
def track_finished_pipelines():
"""when all nodes of a pipeline have been processed, then emit events"""
for running_pipeline, (start_time, running_children) \
in dict(running_pipelines).items(): # type: pipelines.Pipeline
if len(
set(running_pipeline.nodes.values())
& processed_nodes) == len(running_pipeline.nodes):
succeeded = running_pipeline not in failed_pipelines
event_queue.put(
events.Output(
node_path=running_pipeline.path(),
format=logger.Format.ITALICS,
is_error=not succeeded,
message=
f'{"succeeded" if succeeded else "failed"}, {logger.format_time_difference(run_start_time, datetime.datetime.now())}'
))
event_queue.put(
events.NodeFinished(
node_path=running_pipeline.path(),
start_time=start_time,
end_time=datetime.datetime.now(),
is_pipeline=True,
succeeded=succeeded))
del running_pipelines[running_pipeline]
processed_nodes.add(running_pipeline)
# announce run start
event_queue.put(
events.RunStarted(node_path=pipeline.path(),
start_time=run_start_time,
pid=os.getpid()))
# run as long
# - as task processes are still running
# - as there is still stuff in the node queue
while running_task_processes or node_queue:
# don't do anything if the maximum number of parallel tasks is currently running
if len(running_task_processes
) < config.max_number_of_parallel_tasks():
next_node = dequeue(
) # get the next runnable node from the queue
if next_node:
if isinstance(next_node, pipelines.Pipeline):
# connect pipeline nodes without upstreams to upstreams of pipeline
for upstream in next_node.upstreams:
for pipeline_node in next_node.nodes.values():
if not pipeline_node.upstreams:
next_node.add_dependency(
upstream, pipeline_node)
# connect pipeline nodes without downstreams to downstream of pipeline
for downstream in next_node.downstreams:
for pipeline_node in next_node.nodes.values():
if not pipeline_node.downstreams:
next_node.add_dependency(
pipeline_node, downstream)
# queue all child nodes
queue(list(next_node.nodes.values()))
# book keeping and event emission
pipeline_start_time = datetime.datetime.now()
running_pipelines[next_node] = [
pipeline_start_time, 0
]
event_queue.put(
events.NodeStarted(next_node.path(),
pipeline_start_time, True))
event_queue.put(
events.Output(
node_path=next_node.path(),
format=logger.Format.ITALICS,
message='★ ' + node_cost.format_duration(
node_durations_and_run_times.get(
tuple(next_node.path()),
[0, 0])[0])))
elif isinstance(next_node, pipelines.ParallelTask):
# create sub tasks and queue them
try:
logger.redirect_output(event_queue,
next_node.path())
logger.log('☆ Launching tasks',
format=logger.Format.ITALICS)
sub_pipeline = next_node.launch()
next_node.parent.replace(
next_node, sub_pipeline)
queue([sub_pipeline])
except Exception as e:
logger.log(
message=f'Could not launch parallel tasks',
format=logger.Format.ITALICS,
is_error=True)
logger.log(
message=traceback.format_exc(),
format=events.Output.Format.VERBATIM,
is_error=True)
failed_pipelines.add(next_node.parent)
finally:
logger.redirect_output(event_queue,
pipeline.path())
else:
# run a task in a subprocess
if next_node.parent in running_pipelines:
running_pipelines[next_node.parent][1] += 1
event_queue.put(
events.NodeStarted(next_node.path(),
datetime.datetime.now(),
False))
event_queue.put(
events.Output(
node_path=next_node.path(),
format=logger.Format.ITALICS,
message='★ ' + node_cost.format_duration(
node_durations_and_run_times.get(
tuple(next_node.path()),
[0, 0])[0])))
status_queue = multiprocessing.Queue()
process = TaskProcess(next_node, event_queue,
status_queue)
process.start()
running_task_processes[next_node] = process
# check whether some of the running processes finished
for task_process in list(
running_task_processes.values()): # type: TaskProcess
if task_process.is_alive():
pass
else:
del running_task_processes[task_process.task]
if task_process.task.parent in running_pipelines:
running_pipelines[task_process.task.parent][1] -= 1
processed_nodes.add(task_process.task)
succeeded = not (task_process.status_queue.get()
== False
or task_process.exitcode != 0)
if not succeeded:
for parent in task_process.task.parents()[:-1]:
failed_pipelines.add(parent)
end_time = datetime.datetime.now()
event_queue.put(
events.Output(
task_process.task.path(),
('succeeded' if succeeded else 'failed') +
', ' + logger.format_time_difference(
task_process.start_time, end_time),
format=logger.Format.ITALICS,
is_error=not succeeded))
event_queue.put(
events.NodeFinished(task_process.task.path(),
task_process.start_time,
end_time, False, succeeded))
# check if some pipelines finished
track_finished_pipelines()
# don't busy-wait
time.sleep(0.001)
except:
event_queue.put(
events.Output(node_path=pipeline.path(),
message=traceback.format_exc(),
format=logger.Format.ITALICS,
is_error=True))
# run again because `dequeue` might have moved more nodes to `finished_nodes`
track_finished_pipelines()
# kill the stats process (joining or terminating does not work in gunicorn)
os.kill(statistics_process.pid, signal.SIGKILL)
statistics_process.join()
# run finished
event_queue.put(
events.RunFinished(node_path=pipeline.path(),
end_time=datetime.datetime.now(),
succeeded=not failed_pipelines))
