def _bulk_cleanup(self, jobs: Sequence["Job"],
ex: Executor) -> Iterable["Job"]:
jobs = self.bulk_sync_status(jobs)
# safety check
for job in jobs:
assert job.driver == self.__class__
if job.status in (Job.Status.SUBMITTED, Job.Status.RUNNING):
raise InvalidJobStatus(
f"Job {job} might be running, please kill first")
logger.debug("Cleaning up %d jobs", len(jobs))
def run(job: Job) -> Job:
for d in ["log_dir", "output_dir"]:
try:
path = job.data[d]
if os.path.exists(path):
logger.debug("Path %s exists, attempting to delete",
path)
rmtree(path)
except Exception:
logger.error("Unable to remove directory %s", d)
return job
futures = [ex.submit(run, j) for j in jobs]
for f in as_completed(futures):
yield f.result()
def run(job: Job) -> Job:
for d in ["log_dir", "output_dir"]:
try:
path = job.data[d]
if os.path.exists(path):
logger.debug("Path %s exists, attempting to delete",
path)
rmtree(path)
except Exception:
logger.error("Unable to remove directory %s", d)
return job
def bulk_remove(self,
jobs: Sequence["Job"],
do_cleanup: bool = True) -> None:
logger.debug("Removing %d jobs", len(jobs))
if do_cleanup:
jobs = cast(Sequence[Job], self.bulk_cleanup(jobs))
ids = [j.job_id for j in jobs]
with database.atomic():
for chunk in chunks(ids, self.select_batch_size):
Job.delete().where( # type: ignore
Job.job_id << chunk # type: ignore
).execute()
def cleanup(self, job: "Job") -> "Job":
job = self.sync_status(job)
if job.status in (Job.Status.SUBMITTED, Job.Status.RUNNING):
raise InvalidJobStatus(
f"Job {job} might be running, please kill first")
logger.debug("Cleanup job %s", job)
for d in ["log_dir", "output_dir"]:
path = job.data[d]
if os.path.exists(path):
logger.debug("Path %s exists, attempting to delete", path)
rmtree(path)
return job
def clean(job: Job) -> Job:
# need to make sure the output artifacts are gone, since we're reusing the same job dir
for name in ["stdout"]:
path = job.data[name]
if os.path.exists(path):
logger.debug("Removing %s", path)
os.remove(path)
assert not os.path.exists(path)
for d in ["output_dir"]:
path = job.data[d]
if os.path.exists(path):
logger.debug("Removing %s", path)
rmtree(path)
os.makedirs(path)
return job
def wait_gen(
self,
job: Union["Job", List["Job"]],
poll_interval: Optional[int] = None,
timeout: Optional[int] = None,
) -> Iterable[List[Job]]:
start = datetime.datetime.now()
poll_interval = poll_interval or 30
jobs: List[Job]
if isinstance(job, Job):
jobs = [job]
elif isinstance(job, list):
jobs = job
else:
raise TypeError("Argument is neither job nor list of jobs")
# pre-check for status
for job in jobs:
if job.status == Job.Status.CREATED:
raise ValueError(f"Job is in status {job.status}, cannot wait")
logger.debug("Begin waiting for %d jobs", len(jobs))
while True:
now = datetime.datetime.now()
delta: datetime.timedelta = now - start
if timeout is not None:
if delta.total_seconds() > timeout:
raise TimeoutError()
logger.debug("Refreshing %d", len(jobs))
jobs = list(self.bulk_sync_status(jobs)) # overwrite with updated
# filter out all that are considered waitable
remaining_jobs = [
j for j in jobs
if j.status not in (Job.Status.COMPLETED, Job.Status.FAILED,
Job.Status.UNKNOWN)
]
if len(remaining_jobs) == 0:
logger.debug("Waiting completed")
break
yield jobs
logger.debug(
"Waiting. Elapsed time: %s, %d jobs remaining",
delta,
len(remaining_jobs),
)
time.sleep(poll_interval)
def resubmit(self, job: "Job") -> "Job":
logger.debug("Resubmit job %s", job)
job = self.sync_status(job)
if job.status not in (
Job.Status.FAILED,
Job.Status.COMPLETED,
Job.Status.UNKNOWN,
):
raise InvalidJobStatus(
f"Job {job} not in valid status for resubmit")
try:
self.kill(job) # attempt to kill
except Exception:
pass
# need to make sure the output artifacts are gone, since we're reusing the same job dir
for name in ["stdout"]:
path = job.data[name]
if os.path.exists(path):
logger.debug("Removing %s", path)
os.remove(path)
assert not os.path.exists(path)
for d in ["output_dir"]:
path = job.data[d]
if os.path.exists(path):
logger.debug("Removing %s", path)
rmtree(path)
os.makedirs(path)
# reset to created
job.status = Job.Status.CREATED
job.save()
self.submit(job) # this will reset the status
return job
def remove(self, job: "Job") -> None:
logger.debug("Removing job %s", job)
job = self.cleanup(job)
job.delete_instance()
def bulk_resubmit(self,
jobs: Iterable["Job"],
do_submit: bool = True) -> Iterable["Job"]:
logger.debug("Resubmitting jobs")
jobs = self.bulk_sync_status(list(jobs))
# check status is ok
for job in jobs:
if job.status not in (
Job.Status.FAILED,
Job.Status.COMPLETED,
Job.Status.UNKNOWN,
):
raise InvalidJobStatus(
f"Job {job} not in valid status for resubmit")
try:
jobs = self.bulk_kill(jobs) # attempt to kill
except Exception:
pass
def clean(job: Job) -> Job:
# need to make sure the output artifacts are gone, since we're reusing the same job dir
for name in ["stdout"]:
path = job.data[name]
if os.path.exists(path):
logger.debug("Removing %s", path)
os.remove(path)
assert not os.path.exists(path)
for d in ["output_dir"]:
path = job.data[d]
if os.path.exists(path):
logger.debug("Removing %s", path)
rmtree(path)
os.makedirs(path)
return job
nthreads = 40
logger.debug("Cleaning up on %d threads", nthreads)
with ThreadPoolExecutor(nthreads) as ex:
jobs = list(ex.map(clean, jobs))
# update status
with database.atomic():
with database.atomic():
now = datetime.datetime.utcnow()
def jobit() -> Iterable[Job]:
for job in jobs:
job.status = Job.Status.CREATED
job.updated_at = now
yield job
Job.bulk_update(
jobit(),
fields=[Job.status, Job.updated_at],
batch_size=self.batch_size,
)
jobs = Job.bulk_select(Job.job_id, [j.job_id for j in jobs],
batch_size=self.select_batch_size)
if do_submit:
self.bulk_submit(jobs)
return jobs
