def read(args, options):
"""Replay a thermos checkpoint.
Usage: thermos read [options] checkpoint_filename
Options:
--simple Do not replay the full task state machine. Only print out the contents of
each checkpoint log message.
"""
if len(args) != 1:
app.error('Expected one checkpoint file, got %s' % len(args))
if not os.path.exists(args[0]):
app.error('Could not find %s' % args[0])
dispatcher = CheckpointDispatcher()
state = RunnerState(processes={})
with open(args[0], 'r') as fp:
try:
for record in ThriftRecordReader(fp, RunnerCkpt):
if not options.simple:
dispatcher.dispatch(state, record)
else:
print('CKPT: %s' % record)
except RecordIO.Error as err:
print("Failed to recover from %s: %s" % (fp.name, err))
return
if not options.simple:
if state is None or state.header is None:
print('Checkpoint stream CORRUPT or outdated format')
return
print('Recovered Task Header:')
print(' id: %s' % state.header.task_id)
print(' user: %s' % state.header.user)
print(' host: %s' % state.header.hostname)
print(' sandbox: %s' % state.header.sandbox)
if state.header.ports:
print(' ports: %s' %
' '.join('%s->%s' % (name, port)
for (name, port) in state.header.ports.items()))
print('Recovered Task States:')
for task_status in state.statuses:
print(
' %s [pid: %d] => %s' %
(time.asctime(time.localtime(task_status.timestamp_ms /
1000.0)), task_status.runner_pid,
TaskState._VALUES_TO_NAMES[task_status.state]))
print('Recovered Processes:')
for process, process_history in state.processes.items():
print(' %s runs: %s' % (process, len(process_history)))
for k in reversed(range(len(process_history))):
run = process_history[k]
print(
' %2d: pid=%d, rc=%s, finish:%s, state:%s' %
(k, run.pid, run.return_code if run.return_code is not None
else '', time.asctime(time.localtime(run.stop_time))
if run.stop_time else 'None',
ProcessState._VALUES_TO_NAMES.get(run.state, 'Unknown')))
def read(args, options):
"""Replay a thermos checkpoint.
Usage: thermos read [options] checkpoint_filename
Options:
--simple Do not replay the full task state machine. Only print out the contents of
each checkpoint log message.
"""
if len(args) != 1:
app.error('Expected one checkpoint file, got %s' % len(args))
if not os.path.exists(args[0]):
app.error('Could not find %s' % args[0])
dispatcher = CheckpointDispatcher()
state = RunnerState(processes={})
with open(args[0], 'r') as fp:
try:
for record in ThriftRecordReader(fp, RunnerCkpt):
if not options.simple:
dispatcher.dispatch(state, record)
else:
print('CKPT: %s' % record)
except RecordIO.Error as err:
print("Failed to recover from %s: %s" % (fp.name, err))
return
if not options.simple:
if state is None or state.header is None:
print('Checkpoint stream CORRUPT or outdated format')
return
print('Recovered Task Header:')
print(' id: %s' % state.header.task_id)
print(' user: %s' % state.header.user)
print(' host: %s' % state.header.hostname)
print(' sandbox: %s' % state.header.sandbox)
if state.header.ports:
print(' ports: %s' % ' '.join(
'%s->%s' % (name, port) for (name, port) in state.header.ports.items()))
print('Recovered Task States:')
for task_status in state.statuses:
print(' %s [pid: %d] => %s' % (
time.asctime(time.localtime(task_status.timestamp_ms/1000.0)),
task_status.runner_pid,
TaskState._VALUES_TO_NAMES[task_status.state]))
print('Recovered Processes:')
for process, process_history in state.processes.items():
print(' %s runs: %s' % (process, len(process_history)))
for k in reversed(range(len(process_history))):
run = process_history[k]
print(' %2d: pid=%d, rc=%s, finish:%s, state:%s' % (
k,
run.pid,
run.return_code if run.return_code is not None else '',
time.asctime(time.localtime(run.stop_time)) if run.stop_time else 'None',
ProcessState._VALUES_TO_NAMES.get(run.state, 'Unknown')))
def __init__(self, pathspec, task_id):
self._task_id = task_id
self._dispatcher = CheckpointDispatcher()
self._runnerstate = RunnerState(processes={})
self._runner_ckpt = pathspec.given(
task_id=task_id).getpath('runner_checkpoint')
self._active_file, self._finished_file = (pathspec.given(
task_id=task_id, state=state).getpath('task_path')
for state in ('active',
'finished'))
self._ckpt_head = 0
self._apply_states()
self._lock = threading.Lock()
def state(self):
"""Return final state of Task (RunnerState, read from disk and cached for future access)"""
if self._state is None:
path = self._pathspec.given(
task_id=self._task_id).getpath('runner_checkpoint')
self._state = CheckpointDispatcher.from_file(path)
return self._state
def open_checkpoint(cls, filename, force=False, state=None):
"""
Acquire a locked checkpoint stream.
"""
safe_mkdir(os.path.dirname(filename))
fp = lock_file(filename, "a+")
if fp in (None, False):
if force:
log.info('Found existing runner, forcing leadership forfeit.')
state = state or CheckpointDispatcher.from_file(filename)
if cls.kill_runner(state):
log.info('Successfully killed leader.')
# TODO(wickman) Blocking may not be the best idea here. Perhaps block up to
# a maximum timeout. But blocking is necessary because os.kill does not immediately
# release the lock if we're in force mode.
fp = lock_file(filename, "a+", blocking=True)
else:
log.error('Found existing runner, cannot take control.')
if fp in (None, False):
raise cls.PermissionError(
'Could not open locked checkpoint: %s, lock_file = %s' %
(filename, fp))
ckpt = ThriftRecordWriter(fp)
ckpt.set_sync(True)
return ckpt
def get_states(self, task_id):
"""Returns the (timestamp, status) tuples of the task or [] if could not replay."""
statuses = CheckpointDispatcher.iter_statuses(self._runner_ckpt(task_id))
try:
return [(state.timestamp_ms / 1000.0, state.state) for state in statuses]
except CheckpointDispatcher.ErrorRecoveringState:
return []
def run(self):
self._run_count += 1
atexit.register(self.cleanup)
if self.script_filename:
os.unlink(self.script_filename)
with temporary_file(cleanup=False) as fp:
self.script_filename = fp.name
fp.write(
self.RUN_JOB_SCRIPT % {
'filename': self.job_filename,
'sandbox': self.sandbox,
'root': self.tempdir,
'task_id': self.task_id,
'state_filename': self.state_filename,
'portmap': repr(self.portmap),
'success_rate': self.success_rate,
'random_seed': self.random_seed + self._run_count,
})
with environment_as(PYTHONPATH=os.pathsep.join(sys.path)):
self.po = subprocess.Popen([sys.executable, self.script_filename],
stdout=subprocess.PIPE,
stderr=subprocess.PIPE)
try:
so, se = self.po.communicate()
except OSError as e:
if e.errno == errno.ECHILD:
so = se = 'Killed'
else:
raise
rc = self.po.returncode
if rc != 0:
if os.path.exists(self.job_filename):
config = open(self.job_filename).read()
else:
config = 'Nonexistent!'
if 'THERMOS_DEBUG' in os.environ:
print(
"Runner failed!\n\n\nconfig:%s\n\n\nstdout:%s\n\n\nstderr:%s\n\n\n"
% (config, so, se))
try:
with open(self.state_filename, 'r') as fp:
self.state = thrift_deserialize(RunnerState(), fp.read())
except Exception as e:
if 'THERMOS_DEBUG' in os.environ:
print('Failed to load Runner state: %s' % e, file=sys.stderr)
self.state = RunnerState()
try:
self.reconstructed_state = CheckpointDispatcher.from_file(
self.pathspec.getpath('runner_checkpoint'))
except:
self.reconstructed_state = None
self.initialized = True
return rc
def get_sandbox(self, task_id):
"""Returns the sandbox of the task, or None if it has not yet been initialized."""
try:
for update in CheckpointDispatcher.iter_updates(self._runner_ckpt(task_id)):
if update.runner_header and update.runner_header.sandbox:
return update.runner_header.sandbox
except CheckpointDispatcher.ErrorRecoveringState:
return None
def tail(args, options):
"""Tail the logs of a task process.
Usage: thermos tail task_name [process_name]
"""
if len(args) == 0:
app.error('Expected a task to tail, got nothing!')
if len(args) not in (1, 2):
app.error('Expected at most two arguments (task and optional process), got %d' % len(args))
task_id = args[0]
detector = TaskDetector(root=options.root)
checkpoint = CheckpointDispatcher.from_file(detector.get_checkpoint(task_id))
log_dir = checkpoint.header.log_dir
process_runs = [(process, run) for (process, run) in detector.get_process_runs(task_id, log_dir)]
if len(args) == 2:
process_runs = [(process, run) for (process, run) in process_runs if process == args[1]]
if len(process_runs) == 0:
print('ERROR: No processes found.', file=sys.stderr)
sys.exit(1)
processes = set([process for process, _ in process_runs])
if len(processes) != 1:
print('ERROR: More than one process matches query.', file=sys.stderr)
sys.exit(1)
process = processes.pop()
run = max([run for _, run in process_runs])
logdir = TaskPath(root=options.root, task_id=args[0], process=process,
run=run, log_dir=log_dir).getpath('process_logdir')
logfile = os.path.join(logdir, 'stderr' if options.use_stderr else 'stdout')
monitor = TaskMonitor(TaskPath(root = options.root), args[0])
def log_is_active():
active_processes = monitor.get_active_processes()
for process_status, process_run in active_processes:
if process_status.process == process and process_run == run:
return True
return False
if not log_is_active():
print('Tail of terminal log %s' % logfile)
for line in tail_closed(logfile):
print(line.rstrip())
return
now = time.time()
next_check = now + 5.0
print('Tail of active log %s' % logfile)
for line in tail_f(logfile, include_last=True, forever=False):
print(line.rstrip())
if time.time() > next_check:
if not log_is_active():
break
else:
next_check = time.time() + 5.0
def run(self):
self._run_count += 1
atexit.register(self.cleanup)
if self.script_filename:
os.unlink(self.script_filename)
with temporary_file(cleanup=False) as fp:
self.script_filename = fp.name
fp.write(self.RUN_JOB_SCRIPT % {
'filename': self.job_filename,
'sandbox': self.sandbox,
'root': self.tempdir,
'task_id': self.task_id,
'state_filename': self.state_filename,
'portmap': repr(self.portmap),
'success_rate': self.success_rate,
'random_seed': self.random_seed + self._run_count,
})
with environment_as(PYTHONPATH=os.pathsep.join(sys.path)):
self.po = subprocess.Popen([sys.executable, self.script_filename],
stdout=subprocess.PIPE, stderr=subprocess.PIPE)
try:
so, se = self.po.communicate()
except OSError as e:
if e.errno == errno.ECHILD:
so = se = 'Killed'
else:
raise
rc = self.po.returncode
if rc != 0:
if os.path.exists(self.job_filename):
config = open(self.job_filename).read()
else:
config = 'Nonexistent!'
if 'THERMOS_DEBUG' in os.environ:
print("Runner failed!\n\n\nconfig:%s\n\n\nstdout:%s\n\n\nstderr:%s\n\n\n" % (
config, so, se))
try:
with open(self.state_filename, 'r') as fp:
self.state = thrift_deserialize(RunnerState(), fp.read())
except Exception as e:
if 'THERMOS_DEBUG' in os.environ:
print('Failed to load Runner state: %s' % e, file=sys.stderr)
self.state = RunnerState()
try:
self.reconstructed_state = CheckpointDispatcher.from_file(
self.pathspec.getpath('runner_checkpoint'))
except:
self.reconstructed_state = None
self.initialized = True
return rc
def __init__(self, pathspec, task_id):
self._task_id = task_id
self._dispatcher = CheckpointDispatcher()
self._runnerstate = RunnerState(processes={})
self._runner_ckpt = pathspec.given(task_id=task_id).getpath('runner_checkpoint')
self._active_file, self._finished_file = (
pathspec.given(task_id=task_id, state=state).getpath('task_path')
for state in ('active', 'finished'))
self._ckpt_head = 0
self._apply_states()
self._lock = threading.Lock()
def main(args):
values = app.get_options()
if len(args) > 0:
print >> sys.stderr, "ERROR: unrecognized arguments: %s\n" % (
" ".join(args))
app.help()
sys.exit(1)
if not values.ckpt:
print >> sys.stderr, "ERROR: must supply --checkpoint"
app.help()
sys.exit(1)
fp = file(values.ckpt, "r")
rr = ThriftRecordReader(fp, RunnerCkpt)
wrs = RunnerState(processes={})
dispatcher = CheckpointDispatcher()
try:
for wts in rr:
print 'Recovering: ', wts
if values.assemble is True:
dispatcher.dispatch(wrs, wts)
except RecordIO.Error as err:
print 'Error recovering checkpoint stream: %s' % err
return
print '\n\n\n'
if values.assemble:
print 'Recovered Task Header'
pprint.pprint(wrs.header, indent=4)
print '\nRecovered Task States'
for task_status in wrs.statuses:
print ' %s [pid: %d] => %s' % (
time.asctime(time.localtime(task_status.timestamp_ms /
1000.0)), task_status.runner_pid,
TaskState._VALUES_TO_NAMES[task_status.state])
print '\nRecovered Processes'
pprint.pprint(wrs.processes, indent=4)
def main(args):
values = app.get_options()
if len(args) > 0:
print >> sys.stderr, "ERROR: unrecognized arguments: %s\n" % (" ".join(args))
app.help()
sys.exit(1)
if not values.ckpt:
print >> sys.stderr, "ERROR: must supply --checkpoint"
app.help()
sys.exit(1)
fp = file(values.ckpt, "r")
rr = ThriftRecordReader(fp, RunnerCkpt)
wrs = RunnerState(processes = {})
dispatcher = CheckpointDispatcher()
try:
for wts in rr:
print 'Recovering: ', wts
if values.assemble is True:
dispatcher.dispatch(wrs, wts)
except RecordIO.Error as err:
print 'Error recovering checkpoint stream: %s' % err
return
print '\n\n\n'
if values.assemble:
print 'Recovered Task Header'
pprint.pprint(wrs.header, indent=4)
print '\nRecovered Task States'
for task_status in wrs.statuses:
print ' %s [pid: %d] => %s' % (time.asctime(time.localtime(task_status.timestamp_ms/1000.0)),
task_status.runner_pid, TaskState._VALUES_TO_NAMES[task_status.state])
print '\nRecovered Processes'
pprint.pprint(wrs.processes, indent=4)
def format_task(task_id):
checkpoint_filename = detector.get_checkpoint(task_id)
checkpoint_stat = os.stat(checkpoint_filename)
try:
checkpoint_owner = pwd.getpwuid(checkpoint_stat.st_uid).pw_name
except:
checkpoint_owner = 'uid:%s' % checkpoint_stat.st_uid
print(' %-20s [owner: %8s]' % (task_id, checkpoint_owner), end='')
if options.verbose == 0:
print()
if options.verbose > 0:
state = CheckpointDispatcher.from_file(checkpoint_filename)
if state is None or state.header is None:
print(' - checkpoint stream CORRUPT or outdated format')
return
print(' state: %8s' % TaskState._VALUES_TO_NAMES.get(
state.statuses[-1].state, 'Unknown'),
end='')
print(' start: %25s' % time.asctime(
time.localtime(state.header.launch_time_ms / 1000.0)))
if options.verbose > 1:
print(' user: %s' % state.header.user, end='')
if state.header.ports:
print(' ports: %s' %
' '.join('%s -> %s' % (key, val)
for key, val in state.header.ports.items()))
else:
print(' ports: None')
print(' sandbox: %s' % state.header.sandbox)
if options.verbose > 2:
print(' process table:')
for process, process_history in state.processes.items():
print(' - %s runs: %s' % (process, len(process_history)),
end='')
last_run = process_history[-1]
print(' last: pid=%s, rc=%s, finish:%s, state:%s' %
(last_run.pid or 'None', last_run.return_code
if last_run.return_code is not None else '',
time.asctime(time.localtime(last_run.stop_time))
if last_run.stop_time else 'None',
ProcessState._VALUES_TO_NAMES.get(
last_run.state, 'Unknown')))
print()
def open_checkpoint(cls, filename, force=False, state=None):
"""
Acquire a locked checkpoint stream.
"""
safe_mkdir(os.path.dirname(filename))
fp = lock_file(filename, "a+")
if fp in (None, False):
if force:
log.info("Found existing runner, forcing leadership forfeit.")
state = state or CheckpointDispatcher.from_file(filename)
if cls.kill_runner(state):
log.info("Successfully killed leader.")
# TODO(wickman) Blocking may not be the best idea here. Perhaps block up to
# a maximum timeout. But blocking is necessary because os.kill does not immediately
# release the lock if we're in force mode.
fp = lock_file(filename, "a+", blocking=True)
else:
log.error("Found existing runner, cannot take control.")
if fp in (None, False):
raise cls.PermissionError("Could not open locked checkpoint: %s, lock_file = %s" % (filename, fp))
ckpt = ThriftRecordWriter(fp)
ckpt.set_sync(True)
return ckpt
def format_task(task_id):
checkpoint_filename = detector.get_checkpoint(task_id)
checkpoint_stat = os.stat(checkpoint_filename)
try:
checkpoint_owner = pwd.getpwuid(checkpoint_stat.st_uid).pw_name
except:
checkpoint_owner = 'uid:%s' % checkpoint_stat.st_uid
print(' %-20s [owner: %8s]' % (task_id, checkpoint_owner), end='')
if options.verbose == 0:
print()
if options.verbose > 0:
state = CheckpointDispatcher.from_file(checkpoint_filename)
if state is None or state.header is None:
print(' - checkpoint stream CORRUPT or outdated format')
return
print(' state: %8s' % TaskState._VALUES_TO_NAMES.get(state.statuses[-1].state, 'Unknown'),
end='')
print(' start: %25s' % time.asctime(time.localtime(state.header.launch_time_ms/1000.0)))
if options.verbose > 1:
print(' user: %s' % state.header.user, end='')
if state.header.ports:
print(' ports: %s' % ' '.join('%s -> %s' % (key, val)
for key, val in state.header.ports.items()))
else:
print(' ports: None')
print(' sandbox: %s' % state.header.sandbox)
if options.verbose > 2:
print(' process table:')
for process, process_history in state.processes.items():
print(' - %s runs: %s' % (process, len(process_history)), end='')
last_run = process_history[-1]
print(' last: pid=%s, rc=%s, finish:%s, state:%s' % (
last_run.pid or 'None',
last_run.return_code if last_run.return_code is not None else '',
time.asctime(time.localtime(last_run.stop_time)) if last_run.stop_time else 'None',
ProcessState._VALUES_TO_NAMES.get(last_run.state, 'Unknown')))
print()
def get(cls, task_id, checkpoint_root):
"""
Get a TaskRunner bound to the task_id in checkpoint_root.
"""
path = TaskPath(root=checkpoint_root, task_id=task_id, state='active')
task_json = path.getpath('task_path')
task_checkpoint = path.getpath('runner_checkpoint')
if not os.path.exists(task_json):
return None
task = ThermosConfigLoader.load_json(task_json)
if task is None:
return None
if len(task.tasks()) == 0:
return None
try:
checkpoint = CheckpointDispatcher.from_file(task_checkpoint)
if checkpoint is None or checkpoint.header is None:
return None
return cls(task.tasks()[0].task(), checkpoint_root, checkpoint.header.sandbox,
log_dir=checkpoint.header.log_dir, task_id=task_id,
portmap=checkpoint.header.ports)
except Exception as e:
log.error('Failed to reconstitute checkpoint in TaskRunner.get: %s' % e, exc_info=True)
return None
def state(self, task_id):
if task_id not in self._states:
self._states[task_id] = CheckpointDispatcher.from_file(self._detector.get_checkpoint(task_id))
return self._states[task_id]
def kill(cls,
task_id,
checkpoint_root,
force=False,
terminal_status=TaskState.KILLED,
clock=time):
"""
An implementation of Task killing that doesn't require a fully hydrated TaskRunner object.
Terminal status must be either KILLED or LOST state.
"""
if terminal_status not in (TaskState.KILLED, TaskState.LOST):
raise cls.Error('terminal_status must be KILLED or LOST (got %s)' %
TaskState._VALUES_TO_NAMES.get(terminal_status)
or terminal_status)
pathspec = TaskPath(root=checkpoint_root, task_id=task_id)
checkpoint = pathspec.getpath('runner_checkpoint')
state = CheckpointDispatcher.from_file(checkpoint)
if state is None or state.header is None or state.statuses is None:
if force:
log.error(
'Task has uninitialized TaskState - forcibly finalizing')
cls.finalize_task(pathspec)
return
else:
log.error('Cannot update states in uninitialized TaskState!')
return
ckpt = cls.open_checkpoint(checkpoint, force=force, state=state)
def write_task_state(state):
update = TaskStatus(state=state,
timestamp_ms=int(clock.time() * 1000),
runner_pid=os.getpid(),
runner_uid=os.getuid())
ckpt.write(RunnerCkpt(task_status=update))
def write_process_status(status):
ckpt.write(RunnerCkpt(process_status=status))
if cls.is_task_terminal(state.statuses[-1].state):
log.info('Task is already in terminal state! Finalizing.')
cls.finalize_task(pathspec)
return
with closing(ckpt):
write_task_state(TaskState.ACTIVE)
for process, history in state.processes.items():
process_status = history[-1]
if not cls.is_process_terminal(process_status.state):
if cls.kill_process(state, process):
write_process_status(
ProcessStatus(process=process,
state=ProcessState.KILLED,
seq=process_status.seq + 1,
return_code=-9,
stop_time=clock.time()))
else:
if process_status.state is not ProcessState.WAITING:
write_process_status(
ProcessStatus(process=process,
state=ProcessState.LOST,
seq=process_status.seq + 1))
write_task_state(terminal_status)
cls.finalize_task(pathspec)
def tail(args, options):
"""Tail the logs of a task process.
Usage: thermos tail task_name [process_name]
"""
if len(args) == 0:
app.error('Expected a task to tail, got nothing!')
if len(args) not in (1, 2):
app.error(
'Expected at most two arguments (task and optional process), got %d'
% len(args))
task_id = args[0]
detector = TaskDetector(root=options.root)
checkpoint = CheckpointDispatcher.from_file(
detector.get_checkpoint(task_id))
log_dir = checkpoint.header.log_dir
process_runs = [(process, run)
for (process,
run) in detector.get_process_runs(task_id, log_dir)]
if len(args) == 2:
process_runs = [(process, run) for (process, run) in process_runs
if process == args[1]]
if len(process_runs) == 0:
print('ERROR: No processes found.', file=sys.stderr)
sys.exit(1)
processes = set([process for process, _ in process_runs])
if len(processes) != 1:
print('ERROR: More than one process matches query.', file=sys.stderr)
sys.exit(1)
process = processes.pop()
run = max([run for _, run in process_runs])
logdir = TaskPath(root=options.root,
task_id=args[0],
process=process,
run=run,
log_dir=log_dir).getpath('process_logdir')
logfile = os.path.join(logdir,
'stderr' if options.use_stderr else 'stdout')
monitor = TaskMonitor(TaskPath(root=options.root), args[0])
def log_is_active():
active_processes = monitor.get_active_processes()
for process_status, process_run in active_processes:
if process_status.process == process and process_run == run:
return True
return False
if not log_is_active():
print('Tail of terminal log %s' % logfile)
for line in tail_closed(logfile):
print(line.rstrip())
return
now = time.time()
next_check = now + 5.0
print('Tail of active log %s' % logfile)
for line in tail_f(logfile, include_last=True, forever=False):
print(line.rstrip())
if time.time() > next_check:
if not log_is_active():
break
else:
next_check = time.time() + 5.0
def __init__(self, task, checkpoint_root, sandbox, log_dir=None,
task_id=None, portmap=None, user=None, chroot=False, clock=time,
universal_handler=None, planner_class=TaskPlanner):
"""
required:
task (config.Task) = the task to run
checkpoint_root (path) = the checkpoint root
sandbox (path) = the sandbox in which the path will be run
[if None, cwd will be assumed, but garbage collection will be
disabled for this task.]
optional:
log_dir (string) = directory to house stdout/stderr logs. If not specified, logs will be
written into the sandbox directory under .logs/
task_id (string) = bind to this task id. if not specified, will synthesize an id based
upon task.name()
portmap (dict) = a map (string => integer) from name to port, e.g. { 'http': 80 }
user (string) = the user to run the task as. if not current user, requires setuid
privileges.
chroot (boolean) = whether or not to chroot into the sandbox prior to exec.
clock (time interface) = the clock to use throughout
universal_handler = checkpoint record handler (only used for testing)
planner_class (TaskPlanner class) = TaskPlanner class to use for constructing the task
planning policy.
"""
if not issubclass(planner_class, TaskPlanner):
raise TypeError('planner_class must be a TaskPlanner.')
self._clock = clock
launch_time = self._clock.time()
launch_time_ms = '%06d' % int((launch_time - int(launch_time)) * 10**6)
if not task_id:
self._task_id = '%s-%s.%s' % (task.name(),
time.strftime('%Y%m%d-%H%M%S', time.localtime(launch_time)),
launch_time_ms)
else:
self._task_id = task_id
current_user = TaskRunnerHelper.get_actual_user()
self._user = user or current_user
# TODO(wickman) This should be delegated to the ProcessPlatform / Helper
if self._user != current_user:
if os.geteuid() != 0:
raise ValueError('task specifies user as %s, but %s does not have setuid permission!' % (
self._user, current_user))
self._portmap = portmap or {}
self._launch_time = launch_time
self._log_dir = log_dir or os.path.join(sandbox, '.logs')
self._pathspec = TaskPath(root=checkpoint_root, task_id=self._task_id, log_dir=self._log_dir)
try:
ThermosTaskValidator.assert_valid_task(task)
ThermosTaskValidator.assert_valid_ports(task, self._portmap)
except ThermosTaskValidator.InvalidTaskError as e:
raise self.InvalidTask('Invalid task: %s' % e)
context = ThermosContext(
task_id=self._task_id,
ports=self._portmap,
user=self._user)
self._task, uninterp = (task % Environment(thermos=context)).interpolate()
if len(uninterp) > 0:
raise self.InvalidTask('Failed to interpolate task, missing: %s' %
', '.join(str(ref) for ref in uninterp))
try:
ThermosTaskValidator.assert_same_task(self._pathspec, self._task)
except ThermosTaskValidator.InvalidTaskError as e:
raise self.InvalidTask('Invalid task: %s' % e)
self._plan = None # plan currently being executed (updated by Handlers)
self._regular_plan = planner_class(self._task, clock=clock,
process_filter=lambda proc: proc.final().get() == False)
self._finalizing_plan = planner_class(self._task, clock=clock,
process_filter=lambda proc: proc.final().get() == True)
self._chroot = chroot
self._sandbox = sandbox
self._terminal_state = None
self._ckpt = None
self._process_map = dict((p.name().get(), p) for p in self._task.processes())
self._task_processes = {}
self._stages = dict((state, stage(self)) for state, stage in self.STAGES.items())
self._finalization_start = None
self._preemption_deadline = None
self._watcher = ProcessMuxer(self._pathspec)
self._state = RunnerState(processes = {})
# create runner state
universal_handler = universal_handler or TaskRunnerUniversalHandler
self._dispatcher = CheckpointDispatcher()
self._dispatcher.register_handler(universal_handler(self))
self._dispatcher.register_handler(TaskRunnerProcessHandler(self))
self._dispatcher.register_handler(TaskRunnerTaskHandler(self))
# recover checkpointed runner state and update plan
self._recovery = True
self._replay_runner_ckpt()
def state(self):
"""Return final state of Task (RunnerState, read from disk and cached for future access)"""
if self._state is None:
path = self._pathspec.given(task_id=self._task_id).getpath('runner_checkpoint')
self._state = CheckpointDispatcher.from_file(path)
return self._state
class TaskMonitor(object):
"""
Class responsible for reconstructing and monitoring the state of an individual Thermos task via
its runner checkpoint. Also exports information on active processes in the task.
"""
def __init__(self, pathspec, task_id):
self._task_id = task_id
self._dispatcher = CheckpointDispatcher()
self._runnerstate = RunnerState(processes={})
self._runner_ckpt = pathspec.given(
task_id=task_id).getpath('runner_checkpoint')
self._active_file, self._finished_file = (pathspec.given(
task_id=task_id, state=state).getpath('task_path')
for state in ('active',
'finished'))
self._ckpt_head = 0
self._apply_states()
self._lock = threading.Lock()
def _apply_states(self):
"""
os.stat() the corresponding checkpoint stream of this task and determine if there are new ckpt
records. Attempt to read those records and update the high watermark for that stream.
Returns True if new states were applied, False otherwise.
"""
ckpt_offset = None
try:
ckpt_offset = os.stat(self._runner_ckpt).st_size
updated = False
if self._ckpt_head < ckpt_offset:
with open(self._runner_ckpt, 'r') as fp:
fp.seek(self._ckpt_head)
rr = ThriftRecordReader(fp, RunnerCkpt)
while True:
runner_update = rr.try_read()
if not runner_update:
break
try:
self._dispatcher.dispatch(self._runnerstate,
runner_update)
except CheckpointDispatcher.InvalidSequenceNumber as e:
log.error('Checkpoint stream is corrupt: %s' % e)
break
new_ckpt_head = fp.tell()
updated = self._ckpt_head != new_ckpt_head
self._ckpt_head = new_ckpt_head
return updated
except OSError as e:
if e.errno == errno.ENOENT:
# The log doesn't yet exist, will retry later.
log.warning('Could not read from discovered task %s.' %
self._task_id)
return False
else:
raise
def refresh(self):
"""
Check to see if there are new updates and apply them. Return true if
updates were applied, false otherwise.
"""
with self._lock:
return self._apply_states()
def get_state(self):
"""
Get the latest state of this Task.
"""
with self._lock:
self._apply_states()
return copy.deepcopy(self._runnerstate)
def task_state(self):
state = self.get_state()
return state.statuses[-1].state if state.statuses else TaskState.ACTIVE
@property
def active(self):
return os.path.exists(self._active_file)
@property
def finished(self):
return os.path.exists(self._finished_file)
def get_active_processes(self):
"""
Get active processes. Returned is a list of tuples of the form:
(ProcessStatus object of running object, its run number)
"""
active_processes = []
with self._lock:
self._apply_states()
state = self._runnerstate
for process, runs in state.processes.items():
if len(runs) == 0:
continue
last_run = runs[-1]
if last_run.state == ProcessState.RUNNING:
active_processes.append((last_run, len(runs) - 1))
return active_processes
def kill(cls, task_id, checkpoint_root, force=False, terminal_status=TaskState.KILLED, clock=time):
"""
An implementation of Task killing that doesn't require a fully hydrated TaskRunner object.
Terminal status must be either KILLED or LOST state.
"""
if terminal_status not in (TaskState.KILLED, TaskState.LOST):
raise cls.Error(
"terminal_status must be KILLED or LOST (got %s)" % TaskState._VALUES_TO_NAMES.get(terminal_status)
or terminal_status
)
pathspec = TaskPath(root=checkpoint_root, task_id=task_id)
checkpoint = pathspec.getpath("runner_checkpoint")
state = CheckpointDispatcher.from_file(checkpoint)
if state is None or state.header is None or state.statuses is None:
if force:
log.error("Task has uninitialized TaskState - forcibly finalizing")
cls.finalize_task(pathspec)
return
else:
log.error("Cannot update states in uninitialized TaskState!")
return
ckpt = cls.open_checkpoint(checkpoint, force=force, state=state)
def write_task_state(state):
update = TaskStatus(
state=state, timestamp_ms=int(clock.time() * 1000), runner_pid=os.getpid(), runner_uid=os.getuid()
)
ckpt.write(RunnerCkpt(task_status=update))
def write_process_status(status):
ckpt.write(RunnerCkpt(process_status=status))
if cls.is_task_terminal(state.statuses[-1].state):
log.info("Task is already in terminal state! Finalizing.")
cls.finalize_task(pathspec)
return
with closing(ckpt):
write_task_state(TaskState.ACTIVE)
for process, history in state.processes.items():
process_status = history[-1]
if not cls.is_process_terminal(process_status.state):
if cls.kill_process(state, process):
write_process_status(
ProcessStatus(
process=process,
state=ProcessState.KILLED,
seq=process_status.seq + 1,
return_code=-9,
stop_time=clock.time(),
)
)
else:
if process_status.state is not ProcessState.WAITING:
write_process_status(
ProcessStatus(process=process, state=ProcessState.LOST, seq=process_status.seq + 1)
)
write_task_state(terminal_status)
cls.finalize_task(pathspec)
def inspect(self, task_id):
"""
Reconstructs the checkpoint stream and returns a CheckpointInspection.
"""
dispatcher = CheckpointDispatcher()
state = RunnerState(processes = {})
muxer = ProcessMuxer(self._path.given(task_id=task_id))
runner_processes = []
coordinator_processes = set()
processes = set()
def consume_process_record(record):
if not record.process_status:
return
try:
user_uid = pwd.getpwnam(state.header.user).pw_uid
except KeyError:
log.error('Could not find user: %s' % state.header.user)
return
if record.process_status.state == ProcessState.FORKED:
coordinator_processes.add((record.process_status.coordinator_pid, user_uid,
record.process_status.fork_time))
elif record.process_status.state == ProcessState.RUNNING:
processes.add((record.process_status.pid, user_uid,
record.process_status.start_time))
# replay runner checkpoint
runner_pid = None
runner_latest_update = 0
try:
with open(self._path.given(task_id=task_id).getpath('runner_checkpoint')) as fp:
with closing(ThriftRecordReader(fp, RunnerCkpt)) as ckpt:
for record in ckpt:
dispatcher.dispatch(state, record)
runner_latest_update = max(runner_latest_update,
self.get_timestamp(record.process_status))
# collect all bound runners
if record.task_status:
if record.task_status.runner_pid != runner_pid:
runner_processes.append((record.task_status.runner_pid,
record.task_status.runner_uid or 0,
record.task_status.timestamp_ms))
runner_pid = record.task_status.runner_pid
elif record.process_status:
consume_process_record(record)
except (IOError, OSError, RecordIO.Error) as err:
log.debug('Error inspecting task runner checkpoint: %s' % err)
return
# register existing processes in muxer
for process_name in state.processes:
muxer.register(process_name)
# read process checkpoints
process_latest_update = runner_latest_update
for record in muxer.select():
process_latest_update = max(process_latest_update, self.get_timestamp(record.process_status))
consume_process_record(record)
return CheckpointInspection(
runner_latest_update=runner_latest_update,
process_latest_update=process_latest_update,
runner_processes=runner_processes,
coordinator_processes=coordinator_processes,
processes=processes)
class TaskMonitor(object):
"""
Class responsible for reconstructing and monitoring the state of an individual Thermos task via
its runner checkpoint. Also exports information on active processes in the task.
"""
def __init__(self, pathspec, task_id):
self._task_id = task_id
self._dispatcher = CheckpointDispatcher()
self._runnerstate = RunnerState(processes={})
self._runner_ckpt = pathspec.given(task_id=task_id).getpath('runner_checkpoint')
self._active_file, self._finished_file = (
pathspec.given(task_id=task_id, state=state).getpath('task_path')
for state in ('active', 'finished'))
self._ckpt_head = 0
self._apply_states()
self._lock = threading.Lock()
def _apply_states(self):
"""
os.stat() the corresponding checkpoint stream of this task and determine if there are new ckpt
records. Attempt to read those records and update the high watermark for that stream.
Returns True if new states were applied, False otherwise.
"""
ckpt_offset = None
try:
ckpt_offset = os.stat(self._runner_ckpt).st_size
updated = False
if self._ckpt_head < ckpt_offset:
with open(self._runner_ckpt, 'r') as fp:
fp.seek(self._ckpt_head)
rr = ThriftRecordReader(fp, RunnerCkpt)
while True:
runner_update = rr.try_read()
if not runner_update:
break
try:
self._dispatcher.dispatch(self._runnerstate, runner_update)
except CheckpointDispatcher.InvalidSequenceNumber as e:
log.error('Checkpoint stream is corrupt: %s' % e)
break
new_ckpt_head = fp.tell()
updated = self._ckpt_head != new_ckpt_head
self._ckpt_head = new_ckpt_head
return updated
except OSError as e:
if e.errno == errno.ENOENT:
# The log doesn't yet exist, will retry later.
log.warning('Could not read from discovered task %s.' % self._task_id)
return False
else:
raise
def refresh(self):
"""
Check to see if there are new updates and apply them. Return true if
updates were applied, false otherwise.
"""
with self._lock:
return self._apply_states()
def get_state(self):
"""
Get the latest state of this Task.
"""
with self._lock:
self._apply_states()
return copy.deepcopy(self._runnerstate)
def task_state(self):
state = self.get_state()
return state.statuses[-1].state if state.statuses else TaskState.ACTIVE
@property
def active(self):
return os.path.exists(self._active_file)
@property
def finished(self):
return os.path.exists(self._finished_file)
def get_active_processes(self):
"""
Get active processes. Returned is a list of tuples of the form:
(ProcessStatus object of running object, its run number)
"""
active_processes = []
with self._lock:
self._apply_states()
state = self._runnerstate
for process, runs in state.processes.items():
if len(runs) == 0:
continue
last_run = runs[-1]
if last_run.state == ProcessState.RUNNING:
active_processes.append((last_run, len(runs) - 1))
return active_processes
class TaskRunner(object):
"""
Run a ThermosTask.
This class encapsulates the core logic to run and control the state of a Thermos task.
Typically, it will be instantiated directly to control a new task, but a TaskRunner can also be
synthesised from an existing task's checkpoint root
"""
class Error(Exception): pass
class InvalidTask(Error): pass
class InternalError(Error): pass
class PermissionError(Error): pass
class StateError(Error): pass
# Maximum amount of time we spend waiting for new updates from the checkpoint streams
# before doing housecleaning (checking for LOST tasks, dead PIDs.)
MAX_ITERATION_TIME = Amount(10, Time.SECONDS)
# Minimum amount of time we wait between polls for updates on coordinator checkpoints.
COORDINATOR_INTERVAL_SLEEP = Amount(1, Time.SECONDS)
# Amount of time we're willing to wait after forking before we expect the runner to have
# exec'ed the child process.
LOST_TIMEOUT = Amount(60, Time.SECONDS)
# Active task stages
STAGES = {
TaskState.ACTIVE: TaskRunnerStage_ACTIVE,
TaskState.CLEANING: TaskRunnerStage_CLEANING,
TaskState.FINALIZING: TaskRunnerStage_FINALIZING
}
@classmethod
def get(cls, task_id, checkpoint_root):
"""
Get a TaskRunner bound to the task_id in checkpoint_root.
"""
path = TaskPath(root=checkpoint_root, task_id=task_id, state='active')
task_json = path.getpath('task_path')
task_checkpoint = path.getpath('runner_checkpoint')
if not os.path.exists(task_json):
return None
task = ThermosConfigLoader.load_json(task_json)
if task is None:
return None
if len(task.tasks()) == 0:
return None
try:
checkpoint = CheckpointDispatcher.from_file(task_checkpoint)
if checkpoint is None or checkpoint.header is None:
return None
return cls(task.tasks()[0].task(), checkpoint_root, checkpoint.header.sandbox,
log_dir=checkpoint.header.log_dir, task_id=task_id,
portmap=checkpoint.header.ports)
except Exception as e:
log.error('Failed to reconstitute checkpoint in TaskRunner.get: %s' % e, exc_info=True)
return None
def __init__(self, task, checkpoint_root, sandbox, log_dir=None,
task_id=None, portmap=None, user=None, chroot=False, clock=time,
universal_handler=None, planner_class=TaskPlanner):
"""
required:
task (config.Task) = the task to run
checkpoint_root (path) = the checkpoint root
sandbox (path) = the sandbox in which the path will be run
[if None, cwd will be assumed, but garbage collection will be
disabled for this task.]
optional:
log_dir (string) = directory to house stdout/stderr logs. If not specified, logs will be
written into the sandbox directory under .logs/
task_id (string) = bind to this task id. if not specified, will synthesize an id based
upon task.name()
portmap (dict) = a map (string => integer) from name to port, e.g. { 'http': 80 }
user (string) = the user to run the task as. if not current user, requires setuid
privileges.
chroot (boolean) = whether or not to chroot into the sandbox prior to exec.
clock (time interface) = the clock to use throughout
universal_handler = checkpoint record handler (only used for testing)
planner_class (TaskPlanner class) = TaskPlanner class to use for constructing the task
planning policy.
"""
if not issubclass(planner_class, TaskPlanner):
raise TypeError('planner_class must be a TaskPlanner.')
self._clock = clock
launch_time = self._clock.time()
launch_time_ms = '%06d' % int((launch_time - int(launch_time)) * 10**6)
if not task_id:
self._task_id = '%s-%s.%s' % (task.name(),
time.strftime('%Y%m%d-%H%M%S', time.localtime(launch_time)),
launch_time_ms)
else:
self._task_id = task_id
current_user = TaskRunnerHelper.get_actual_user()
self._user = user or current_user
# TODO(wickman) This should be delegated to the ProcessPlatform / Helper
if self._user != current_user:
if os.geteuid() != 0:
raise ValueError('task specifies user as %s, but %s does not have setuid permission!' % (
self._user, current_user))
self._portmap = portmap or {}
self._launch_time = launch_time
self._log_dir = log_dir or os.path.join(sandbox, '.logs')
self._pathspec = TaskPath(root=checkpoint_root, task_id=self._task_id, log_dir=self._log_dir)
try:
ThermosTaskValidator.assert_valid_task(task)
ThermosTaskValidator.assert_valid_ports(task, self._portmap)
except ThermosTaskValidator.InvalidTaskError as e:
raise self.InvalidTask('Invalid task: %s' % e)
context = ThermosContext(
task_id=self._task_id,
ports=self._portmap,
user=self._user)
self._task, uninterp = (task % Environment(thermos=context)).interpolate()
if len(uninterp) > 0:
raise self.InvalidTask('Failed to interpolate task, missing: %s' %
', '.join(str(ref) for ref in uninterp))
try:
ThermosTaskValidator.assert_same_task(self._pathspec, self._task)
except ThermosTaskValidator.InvalidTaskError as e:
raise self.InvalidTask('Invalid task: %s' % e)
self._plan = None # plan currently being executed (updated by Handlers)
self._regular_plan = planner_class(self._task, clock=clock,
process_filter=lambda proc: proc.final().get() == False)
self._finalizing_plan = planner_class(self._task, clock=clock,
process_filter=lambda proc: proc.final().get() == True)
self._chroot = chroot
self._sandbox = sandbox
self._terminal_state = None
self._ckpt = None
self._process_map = dict((p.name().get(), p) for p in self._task.processes())
self._task_processes = {}
self._stages = dict((state, stage(self)) for state, stage in self.STAGES.items())
self._finalization_start = None
self._preemption_deadline = None
self._watcher = ProcessMuxer(self._pathspec)
self._state = RunnerState(processes = {})
# create runner state
universal_handler = universal_handler or TaskRunnerUniversalHandler
self._dispatcher = CheckpointDispatcher()
self._dispatcher.register_handler(universal_handler(self))
self._dispatcher.register_handler(TaskRunnerProcessHandler(self))
self._dispatcher.register_handler(TaskRunnerTaskHandler(self))
# recover checkpointed runner state and update plan
self._recovery = True
self._replay_runner_ckpt()
@property
def task(self):
return self._task
@property
def task_id(self):
return self._task_id
@property
def state(self):
return self._state
@property
def processes(self):
return self._task_processes
def task_state(self):
return self._state.statuses[-1].state if self._state.statuses else TaskState.ACTIVE
def close_ckpt(self):
"""Force close the checkpoint stream. This is necessary for runners terminated through
exception propagation."""
log.debug('Closing the checkpoint stream.')
self._ckpt.close()
@contextmanager
def control(self, force=False):
"""
Bind to the checkpoint associated with this task, position to the end of the log if
it exists, or create it if it doesn't. Fails if we cannot get "leadership" i.e. a
file lock on the checkpoint stream.
"""
if self.is_terminal():
raise TaskRunner.StateError('Cannot take control of a task in terminal state.')
if self._sandbox:
safe_mkdir(self._sandbox)
ckpt_file = self._pathspec.getpath('runner_checkpoint')
try:
self._ckpt = TaskRunnerHelper.open_checkpoint(ckpt_file, force=force, state=self._state)
except TaskRunnerHelper.PermissionError:
raise TaskRunner.PermissionError('Unable to open checkpoint %s' % ckpt_file)
log.debug('Flipping recovery mode off.')
self._recovery = False
self._set_task_status(self.task_state())
self._resume_task()
try:
yield
except Exception as e:
log.error('Caught exception in self.control(): %s' % e)
log.error(' %s' % traceback.format_exc())
self._ckpt.close()
def _resume_task(self):
assert self._ckpt is not None
unapplied_updates = self._replay_process_ckpts()
if self.is_terminal():
raise self.StateError('Cannot resume terminal task.')
self._initialize_ckpt_header()
self._replay(unapplied_updates)
def _ckpt_write(self, record):
"""
Write to the checkpoint stream if we're not in recovery mode.
"""
if not self._recovery:
self._ckpt.write(record)
def _replay(self, checkpoints):
"""
Replay a sequence of RunnerCkpts.
"""
for checkpoint in checkpoints:
self._dispatcher.dispatch(self._state, checkpoint)
def _replay_runner_ckpt(self):
"""
Replay the checkpoint stream associated with this task.
"""
ckpt_file = self._pathspec.getpath('runner_checkpoint')
if os.path.exists(ckpt_file):
fp = open(ckpt_file, "r")
ckpt_recover = ThriftRecordReader(fp, RunnerCkpt)
for record in ckpt_recover:
log.debug('Replaying runner checkpoint record: %s' % record)
self._dispatcher.dispatch(self._state, record, recovery=True)
ckpt_recover.close()
def _replay_process_ckpts(self):
"""
Replay the unmutating process checkpoints. Return the unapplied process updates that
would mutate the runner checkpoint stream.
"""
process_updates = self._watcher.select()
unapplied_process_updates = []
for process_update in process_updates:
if self._dispatcher.would_update(self._state, process_update):
unapplied_process_updates.append(process_update)
else:
self._dispatcher.dispatch(self._state, process_update, recovery=True)
return unapplied_process_updates
def _initialize_ckpt_header(self):
"""
Initializes the RunnerHeader for this checkpoint stream if it has not already
been constructed.
"""
if self._state.header is None:
header = RunnerHeader(
task_id=self._task_id,
launch_time_ms=int(self._launch_time*1000),
sandbox=self._sandbox,
log_dir=self._log_dir,
hostname=socket.gethostname(),
user=self._user,
ports=self._portmap)
runner_ckpt = RunnerCkpt(runner_header=header)
self._dispatcher.dispatch(self._state, runner_ckpt)
def _set_task_status(self, state):
update = TaskStatus(state=state, timestamp_ms=int(self._clock.time() * 1000),
runner_pid=os.getpid(), runner_uid=os.getuid())
runner_ckpt = RunnerCkpt(task_status=update)
self._dispatcher.dispatch(self._state, runner_ckpt, self._recovery)
def _finalization_remaining(self):
# If a preemption deadline has been set, use that.
if self._preemption_deadline:
return max(0, self._preemption_deadline - self._clock.time())
# Otherwise, use the finalization wait provided in the configuration.
finalization_allocation = self.task.finalization_wait().get()
if self._finalization_start is None:
return sys.float_info.max
else:
waited = max(0, self._clock.time() - self._finalization_start)
return max(0, finalization_allocation - waited)
def _set_process_status(self, process_name, process_state, **kw):
if 'sequence_number' in kw:
sequence_number = kw.pop('sequence_number')
log.debug('_set_process_status(%s <= %s, seq=%s[force])' % (process_name,
ProcessState._VALUES_TO_NAMES.get(process_state), sequence_number))
else:
current_run = self._current_process_run(process_name)
if not current_run:
assert process_state == ProcessState.WAITING
sequence_number = 0
else:
sequence_number = current_run.seq + 1
log.debug('_set_process_status(%s <= %s, seq=%s[auto])' % (process_name,
ProcessState._VALUES_TO_NAMES.get(process_state), sequence_number))
runner_ckpt = RunnerCkpt(process_status=ProcessStatus(
process=process_name, state=process_state, seq=sequence_number, **kw))
self._dispatcher.dispatch(self._state, runner_ckpt, self._recovery)
def _task_process_from_process_name(self, process_name, sequence_number):
"""
Construct a Process() object from a process_name, populated with its
correct run number and fully interpolated commandline.
"""
run_number = len(self.state.processes[process_name]) - 1
pathspec = self._pathspec.given(process=process_name, run=run_number)
process = self._process_map.get(process_name)
if process is None:
raise self.InternalError('FATAL: Could not find process: %s' % process_name)
def close_ckpt_and_fork():
pid = os.fork()
if pid == 0 and self._ckpt is not None:
self._ckpt.close()
return pid
return Process(
process.name().get(),
process.cmdline().get(),
sequence_number,
pathspec,
self._sandbox,
self._user,
chroot=self._chroot,
fork=close_ckpt_and_fork)
def deadlocked(self, plan=None):
"""Check whether a plan is deadlocked, i.e. there are no running/runnable processes, and the
plan is not complete."""
plan = plan or self._regular_plan
now = self._clock.time()
running = list(plan.running)
runnable = list(plan.runnable_at(now))
waiting = list(plan.waiting_at(now))
log.debug('running:%d runnable:%d waiting:%d complete:%s' % (
len(running), len(runnable), len(waiting), plan.is_complete()))
return len(running + runnable + waiting) == 0 and not plan.is_complete()
def is_healthy(self):
"""Check whether the TaskRunner is healthy. A healthy TaskRunner is not deadlocked and has not
reached its max_failures count."""
max_failures = self._task.max_failures().get()
deadlocked = self.deadlocked()
under_failure_limit = max_failures == 0 or len(self._regular_plan.failed) < max_failures
log.debug('max_failures:%d failed:%d under_failure_limit:%s deadlocked:%s ==> health:%s' % (
max_failures, len(self._regular_plan.failed), under_failure_limit, deadlocked,
not deadlocked and under_failure_limit))
return not deadlocked and under_failure_limit
def _current_process_run(self, process_name):
if process_name not in self._state.processes or len(self._state.processes[process_name]) == 0:
return None
return self._state.processes[process_name][-1]
def is_process_lost(self, process_name):
"""Determine whether or not we should mark a task as LOST and do so if necessary."""
current_run = self._current_process_run(process_name)
if not current_run:
raise self.InternalError('No current_run for process %s!' % process_name)
def forked_but_never_came_up():
return current_run.state == ProcessState.FORKED and (
self._clock.time() - current_run.fork_time > TaskRunner.LOST_TIMEOUT.as_(Time.SECONDS))
def running_but_coordinator_died():
if current_run.state != ProcessState.RUNNING:
return False
coordinator_pid, _, _ = TaskRunnerHelper.scan_process(self.state, process_name)
if coordinator_pid is not None:
return False
elif self._watcher.has_data(process_name):
return False
return True
if forked_but_never_came_up() or running_but_coordinator_died():
log.info('Detected a LOST task: %s' % current_run)
log.debug(' forked_but_never_came_up: %s' % forked_but_never_came_up())
log.debug(' running_but_coordinator_died: %s' % running_but_coordinator_died())
return True
return False
def _run_plan(self, plan):
log.debug('Schedule pass:'******'running: %s' % ' '.join(plan.running))
log.debug('finished: %s' % ' '.join(plan.finished))
launched = []
for process_name in plan.running:
if self.is_process_lost(process_name):
self._set_process_status(process_name, ProcessState.LOST)
now = self._clock.time()
runnable = list(plan.runnable_at(now))
waiting = list(plan.waiting_at(now))
log.debug('runnable: %s' % ' '.join(runnable))
log.debug('waiting: %s' % ' '.join(
'%s[T-%.1fs]' % (process, plan.get_wait(process)) for process in waiting))
def pick_processes(process_list):
if self._task.max_concurrency().get() == 0:
return process_list
num_to_pick = max(self._task.max_concurrency().get() - len(running), 0)
return process_list[:num_to_pick]
for process_name in pick_processes(runnable):
tp = self._task_processes.get(process_name)
if tp:
current_run = self._current_process_run(process_name)
assert current_run.state == ProcessState.WAITING
else:
self._set_process_status(process_name, ProcessState.WAITING)
tp = self._task_processes[process_name]
log.info('Forking Process(%s)' % process_name)
tp.start()
launched.append(tp)
return len(launched) > 0
def _terminate_plan(self, plan):
for process in plan.running:
last_run = self._current_process_run(process)
if last_run and last_run.state in (ProcessState.FORKED, ProcessState.RUNNING):
TaskRunnerHelper.terminate_process(self.state, process)
def has_running_processes(self):
"""
Returns True if any processes associated with this task have active pids.
"""
process_tree = TaskRunnerHelper.scantree(self.state)
return any(any(process_set) for process_set in process_tree.values())
def has_active_processes(self):
"""
Returns True if any processes are in non-terminal states.
"""
return any(not TaskRunnerHelper.is_process_terminal(run.state) for run in
filter(None, (self._current_process_run(process) for process in self.state.processes)))
def collect_updates(self, timeout=None):
"""
Collects and applies updates from process checkpoint streams. Returns the number
of applied process checkpoints.
"""
if self.has_active_processes():
sleep_interval = self.COORDINATOR_INTERVAL_SLEEP.as_(Time.SECONDS)
total_time = 0.0
while True:
process_updates = self._watcher.select()
for process_update in process_updates:
self._dispatcher.dispatch(self._state, process_update, self._recovery)
if process_updates:
return len(process_updates)
if timeout and total_time >= timeout:
break
total_time += sleep_interval
self._clock.sleep(sleep_interval)
return 0
def is_terminal(self):
return TaskRunnerHelper.is_task_terminal(self.task_state())
def terminal_state(self):
if self._terminal_state:
log.debug('Forced terminal state: %s' %
TaskState._VALUES_TO_NAMES.get(self._terminal_state, 'UNKNOWN'))
return self._terminal_state
else:
return TaskState.SUCCESS if self.is_healthy() else TaskState.FAILED
def run(self, force=False):
"""
Entrypoint to runner. Assume control of checkpoint stream, and execute TaskRunnerStages
until runner is terminal.
"""
if self.is_terminal():
return
with self.control(force):
self._run()
def _run(self):
iteration_time = self.MAX_ITERATION_TIME.as_(Time.SECONDS)
while not self.is_terminal():
start = self._clock.time()
# step 1: execute stage corresponding to the state we're currently in
runner = self._stages[self.task_state()]
iteration_wait = runner.run()
if iteration_wait is None:
log.debug('Run loop: No more work to be done in state %s' %
TaskState._VALUES_TO_NAMES.get(self.task_state(), 'UNKNOWN'))
self._set_task_status(runner.transition_to())
continue
log.debug('Run loop: Work to be done within %.1fs' % iteration_wait)
# step 2: check child process checkpoint streams for updates
if not self.collect_updates(iteration_wait):
# If we don't collect any updates, at least 'touch' the checkpoint stream
# so as to prevent garbage collection.
elapsed = self._clock.time() - start
if elapsed < iteration_wait:
log.debug('Update collection only took %.1fs, idling %.1fs' % (
elapsed, iteration_wait - elapsed))
self._clock.sleep(iteration_wait - elapsed)
log.debug('Run loop: No updates collected, touching checkpoint.')
os.utime(self._pathspec.getpath('runner_checkpoint'), None)
# step 3: reap any zombie child processes
TaskRunnerHelper.reap_children()
def kill(self, force=False, terminal_status=TaskState.KILLED,
preemption_wait=Amount(1, Time.MINUTES)):
"""
Kill all processes associated with this task and set task/process states as terminal_status
(defaults to KILLED)
"""
log.debug('Runner issued kill: force:%s, preemption_wait:%s' % (
force, preemption_wait))
assert terminal_status in (TaskState.KILLED, TaskState.LOST)
self._preemption_deadline = self._clock.time() + preemption_wait.as_(Time.SECONDS)
with self.control(force):
if self.is_terminal():
log.warning('Task is not in ACTIVE state, cannot issue kill.')
return
self._terminal_state = terminal_status
if self.task_state() == TaskState.ACTIVE:
self._set_task_status(TaskState.CLEANING)
self._run()
def lose(self, force=False):
"""
Mark a task as LOST and kill any straggling processes.
"""
self.kill(force, preemption_wait=Amount(0, Time.SECONDS), terminal_status=TaskState.LOST)
def _kill(self):
processes = TaskRunnerHelper.scantree(self._state)
for process, pid_tuple in processes.items():
current_run = self._current_process_run(process)
coordinator_pid, pid, tree = pid_tuple
if TaskRunnerHelper.is_process_terminal(current_run.state):
if coordinator_pid or pid or tree:
log.warning('Terminal process (%s) still has running pids:' % process)
log.warning(' coordinator_pid: %s' % coordinator_pid)
log.warning(' pid: %s' % pid)
log.warning(' tree: %s' % tree)
TaskRunnerHelper.kill_process(self.state, process)
else:
if coordinator_pid or pid or tree:
log.info('Transitioning %s to KILLED' % process)
self._set_process_status(process, ProcessState.KILLED,
stop_time=self._clock.time(), return_code=-1)
else:
log.info('Transitioning %s to LOST' % process)
if current_run.state != ProcessState.WAITING:
self._set_process_status(process, ProcessState.LOST)
def inspect(self, task_id):
"""
Reconstructs the checkpoint stream and returns a CheckpointInspection.
"""
dispatcher = CheckpointDispatcher()
state = RunnerState(processes={})
muxer = ProcessMuxer(self._path.given(task_id=task_id))
runner_processes = []
coordinator_processes = set()
processes = set()
def consume_process_record(record):
if not record.process_status:
return
try:
user_uid = pwd.getpwnam(state.header.user).pw_uid
except KeyError:
log.error('Could not find user: %s' % state.header.user)
return
if record.process_status.state == ProcessState.FORKED:
coordinator_processes.add(
(record.process_status.coordinator_pid, user_uid,
record.process_status.fork_time))
elif record.process_status.state == ProcessState.RUNNING:
processes.add((record.process_status.pid, user_uid,
record.process_status.start_time))
# replay runner checkpoint
runner_pid = None
runner_latest_update = 0
try:
with open(
self._path.given(
task_id=task_id).getpath('runner_checkpoint')) as fp:
with closing(ThriftRecordReader(fp, RunnerCkpt)) as ckpt:
for record in ckpt:
dispatcher.dispatch(state, record)
runner_latest_update = max(
runner_latest_update,
self.get_timestamp(record.process_status))
# collect all bound runners
if record.task_status:
if record.task_status.runner_pid != runner_pid:
runner_processes.append(
(record.task_status.runner_pid,
record.task_status.runner_uid
or 0, record.task_status.timestamp_ms))
runner_pid = record.task_status.runner_pid
elif record.process_status:
consume_process_record(record)
except (IOError, OSError, RecordIO.Error) as err:
log.debug('Error inspecting task runner checkpoint: %s' % err)
return
# register existing processes in muxer
for process_name in state.processes:
muxer.register(process_name)
# read process checkpoints
process_latest_update = runner_latest_update
for record in muxer.select():
process_latest_update = max(
process_latest_update,
self.get_timestamp(record.process_status))
consume_process_record(record)
return CheckpointInspection(
runner_latest_update=runner_latest_update,
process_latest_update=process_latest_update,
runner_processes=runner_processes,
coordinator_processes=coordinator_processes,
processes=processes)