Warning
This document is for an in-development version of Galaxy. You can alternatively view this page in the latest release if it exists or view the top of the latest release's documentation.
Source code for galaxy.jobs.handler
"""
Galaxy job handler, prepares, runs, tracks, and finishes Galaxy jobs
"""
import datetime
import os
import time
from collections import defaultdict
from six.moves.queue import (
Empty,
Queue
)
from sqlalchemy.exc import OperationalError
from sqlalchemy.sql.expression import (
and_,
func,
null,
or_,
select,
true
)
from galaxy import model
from galaxy.jobs import (
JobDestination,
JobWrapper,
TaskWrapper
)
from galaxy.jobs.mapper import JobNotReadyException
from galaxy.util import unicodify
from galaxy.util.custom_logging import get_logger
from galaxy.util.monitors import Monitors
from galaxy.web_stack.handlers import HANDLER_ASSIGNMENT_METHODS
from galaxy.web_stack.message import JobHandlerMessage
log = get_logger(__name__)
# States for running a job. These are NOT the same as data states
JOB_WAIT, JOB_ERROR, JOB_INPUT_ERROR, JOB_INPUT_DELETED, JOB_READY, JOB_DELETED, JOB_ADMIN_DELETED, JOB_USER_OVER_QUOTA, JOB_USER_OVER_TOTAL_WALLTIME = 'wait', 'error', 'input_error', 'input_deleted', 'ready', 'deleted', 'admin_deleted', 'user_over_quota', 'user_over_total_walltime'
DEFAULT_JOB_PUT_FAILURE_MESSAGE = 'Unable to run job due to a misconfiguration of the Galaxy job running system. Please contact a site administrator.'
[docs]class JobHandler(object):
"""
Handle the preparation, running, tracking, and finishing of jobs
"""
[docs] def __init__(self, app):
self.app = app
# The dispatcher launches the underlying job runners
self.dispatcher = DefaultJobDispatcher(app)
# Queues for starting and stopping jobs
self.job_queue = JobHandlerQueue(app, self.dispatcher)
self.job_stop_queue = JobHandlerStopQueue(app, self.dispatcher)
[docs]class JobHandlerQueue(Monitors):
"""
Job Handler's Internal Queue, this is what actually implements waiting for
jobs to be runnable and dispatching to a JobRunner.
"""
STOP_SIGNAL = object()
[docs] def __init__(self, app, dispatcher):
"""Initializes the Job Handler Queue, creates (unstarted) monitoring thread"""
self.app = app
self.dispatcher = dispatcher
self.sa_session = app.model.context
self.track_jobs_in_database = self.app.config.track_jobs_in_database
# Initialize structures for handling job limits
self.__clear_job_count()
# Keep track of the pid that started the job manager, only it
# has valid threads
self.parent_pid = os.getpid()
# Contains new jobs. Note this is not used if track_jobs_in_database is True
self.queue = Queue()
# Contains jobs that are waiting (only use from monitor thread)
self.waiting_jobs = []
# Contains wrappers of jobs that are limited or ready (so they aren't created unnecessarily/multiple times)
self.job_wrappers = {}
name = "JobHandlerQueue.monitor_thread"
self._init_monitor_thread(name, target=self.__monitor, config=app.config)
self.__grab_query = None
self.__grab_conn_opts = {'autocommit': False}
self.__initialize_job_grabbing()
def __initialize_job_grabbing(self):
grabbable_methods = {
HANDLER_ASSIGNMENT_METHODS.DB_TRANSACTION_ISOLATION,
HANDLER_ASSIGNMENT_METHODS.DB_SKIP_LOCKED,
}
try:
method = [m for m in self.app.job_config.handler_assignment_methods if m in grabbable_methods][0]
except IndexError:
return
subq = select([model.Job.id]) \
.where(and_(
model.Job.table.c.handler.in_(self.app.job_config.self_handler_tags),
model.Job.table.c.state == model.Job.states.NEW)) \
.order_by(model.Job.table.c.id)
if self.app.job_config.handler_max_grab:
subq = subq.limit(self.app.job_config.handler_max_grab)
if method == HANDLER_ASSIGNMENT_METHODS.DB_SKIP_LOCKED:
subq = subq.with_for_update(skip_locked=True)
self.__grab_query = model.Job.table.update() \
.returning(model.Job.table.c.id) \
.where(model.Job.table.c.id.in_(subq)) \
.values(handler=self.app.config.server_name)
if method == HANDLER_ASSIGNMENT_METHODS.DB_TRANSACTION_ISOLATION:
self.__grab_conn_opts['isolation_level'] = 'SERIALIZABLE'
log.info(
"Handler job grabber initialized with '%s' assignment method for handler '%s', tag(s): %s", method,
self.app.config.server_name, ', '.join(str(x) for x in self.app.job_config.handler_tags)
)
[docs] def start(self):
"""
Starts the JobHandler's thread after checking for any unhandled jobs.
"""
log.debug('Handler queue starting for jobs assigned to handler: %s', self.app.config.server_name)
# Recover jobs at startup
self.__check_jobs_at_startup()
# Start the queue
self.monitor_thread.start()
# The stack code is initialized in the application
JobHandlerMessage().bind_default_handler(self, '_handle_message')
self.app.application_stack.register_message_handler(self._handle_message, name=JobHandlerMessage.target)
log.info("job handler queue started")
[docs] def job_wrapper(self, job, use_persisted_destination=False):
return JobWrapper(job, self, use_persisted_destination=use_persisted_destination)
[docs] def job_pair_for_id(self, id):
job = self.sa_session.query(model.Job).get(id)
return job, self.job_wrapper(job, use_persisted_destination=True)
def __check_jobs_at_startup(self):
"""
Checks all jobs that are in the 'new', 'queued' or 'running' state in
the database and requeues or cleans up as necessary. Only run as the
job handler starts.
In case the activation is enforced it will filter out the jobs of inactive users.
"""
jobs_at_startup = []
if self.track_jobs_in_database:
in_list = (model.Job.states.QUEUED,
model.Job.states.RUNNING)
else:
in_list = (model.Job.states.NEW,
model.Job.states.QUEUED,
model.Job.states.RUNNING)
if self.app.config.user_activation_on:
jobs_at_startup = self.sa_session.query(model.Job).enable_eagerloads(False) \
.outerjoin(model.User) \
.filter(model.Job.state.in_(in_list) &
(model.Job.handler == self.app.config.server_name) &
or_((model.Job.user_id == null()), (model.User.active == true()))).all()
else:
jobs_at_startup = self.sa_session.query(model.Job).enable_eagerloads(False) \
.filter(model.Job.state.in_(in_list) &
(model.Job.handler == self.app.config.server_name)).all()
for job in jobs_at_startup:
if not self.app.toolbox.has_tool(job.tool_id, job.tool_version, exact=True):
log.warning("(%s) Tool '%s' removed from tool config, unable to recover job" % (job.id, job.tool_id))
self.job_wrapper(job).fail('This tool was disabled before the job completed. Please contact your Galaxy administrator.')
elif job.job_runner_name is not None and job.job_runner_external_id is None:
# This could happen during certain revisions of Galaxy where a runner URL was persisted before the job was dispatched to a runner.
log.debug("(%s) Job runner assigned but no external ID recorded, adding to the job handler queue" % job.id)
job.job_runner_name = None
if self.track_jobs_in_database:
job.set_state(model.Job.states.NEW)
else:
self.queue.put((job.id, job.tool_id))
elif job.job_runner_name is not None and job.job_runner_external_id is not None and job.destination_id is None:
# This is the first start after upgrading from URLs to destinations, convert the URL to a destination and persist
job_wrapper = self.job_wrapper(job)
job_destination = self.dispatcher.url_to_destination(job.job_runner_name)
if job_destination.id is None:
job_destination.id = 'legacy_url'
job_wrapper.set_job_destination(job_destination, job.job_runner_external_id)
self.dispatcher.recover(job, job_wrapper)
log.info('(%s) Converted job from a URL to a destination and recovered' % (job.id))
elif job.job_runner_name is None:
# Never (fully) dispatched
log.debug("(%s) No job runner assigned and job still in '%s' state, adding to the job handler queue" % (job.id, job.state))
if self.track_jobs_in_database:
job.set_state(model.Job.states.NEW)
else:
self.queue.put((job.id, job.tool_id))
else:
# Already dispatched and running
job_wrapper = self.__recover_job_wrapper(job)
self.dispatcher.recover(job, job_wrapper)
if self.sa_session.dirty:
self.sa_session.flush()
def __recover_job_wrapper(self, job):
# Already dispatched and running
job_wrapper = self.job_wrapper(job)
# Use the persisted destination as its params may differ from
# what's in the job_conf xml
job_destination = JobDestination(id=job.destination_id, runner=job.job_runner_name, params=job.destination_params)
# resubmits are not persisted (it's a good thing) so they
# should be added back to the in-memory destination on startup
try:
config_job_destination = self.app.job_config.get_destination(job.destination_id)
job_destination.resubmit = config_job_destination.resubmit
except KeyError:
log.debug('(%s) Recovered destination id (%s) does not exist in job config (but this may be normal in the case of a dynamically generated destination)', job.id, job.destination_id)
job_wrapper.job_runner_mapper.cached_job_destination = job_destination
return job_wrapper
def __monitor(self):
"""
Continually iterate the waiting jobs, checking is each is ready to
run and dispatching if so.
"""
while self.monitor_running:
try:
# If jobs are locked, there's nothing to monitor and we skip
# to the sleep.
if not self.app.job_manager.job_lock:
self.__monitor_step()
except Exception:
log.exception("Exception in monitor_step")
# With sqlite backends we can run into locked databases occasionally
# To avoid that the monitor step locks again we backoff a little longer.
self._monitor_sleep(5)
self._monitor_sleep(1)
def __monitor_step(self):
"""
Called repeatedly by `monitor` to process waiting jobs.
"""
monitor_step_timer = self.app.execution_timer_factory.get_timer(
'internal.galaxy.jobs.handlers.monitor_step',
'Job handler monitor step complete.'
)
if self.__grab_query is not None:
self.__grab_unhandled_jobs()
self.__handle_waiting_jobs()
log.trace(monitor_step_timer.to_str())
def __grab_unhandled_jobs(self):
"""
Attempts to "grab" jobs (assign unassigned jobs to itself) using DB serialization methods, if enabled. This
simply sets `Job.handler` to the current server name, which causes the job to be picked up by
`__handle_waiting_jobs()`.
"""
# an excellent discussion on PostgreSQL concurrency safety:
# https://blog.2ndquadrant.com/what-is-select-skip-locked-for-in-postgresql-9-5/
self.sa_session.expunge_all()
conn = self.sa_session.connection(execution_options=self.__grab_conn_opts)
with conn.begin() as trans:
try:
rows = conn.execute(self.__grab_query).fetchall()
if rows:
log.debug('Grabbed job(s): %s', ', '.join(str(row[0]) for row in rows))
trans.commit()
else:
trans.rollback()
except OperationalError as e:
# If this is a serialization failure on PostgreSQL, then e.orig is a psycopg2 TransactionRollbackError
# and should have attribute `code`. Other engines should just report the message and move on.
if int(getattr(e.orig, 'pgcode', -1)) != 40001:
log.debug('Grabbing job failed (serialization failures are ok): %s', unicodify(e))
trans.rollback()
def __handle_waiting_jobs(self):
"""
Gets any new jobs (either from the database or from its own queue), then iterates over all new and waiting jobs
to check the state of the jobs each depends on. If the job has dependencies that have not finished, it goes to
the waiting queue. If the job has dependencies with errors, it is marked as having errors and removed from the
queue. If the job belongs to an inactive user it is ignored. Otherwise, the job is dispatched.
"""
# Pull all new jobs from the queue at once
jobs_to_check = []
resubmit_jobs = []
if self.track_jobs_in_database:
# Clear the session so we get fresh states for job and all datasets
self.sa_session.expunge_all()
# Fetch all new jobs
hda_not_ready = self.sa_session.query(model.Job.id).enable_eagerloads(False) \
.join(model.JobToInputDatasetAssociation) \
.join(model.HistoryDatasetAssociation) \
.join(model.Dataset) \
.filter(and_(model.Job.state == model.Job.states.NEW,
model.Dataset.state.in_(model.Dataset.non_ready_states))).subquery()
ldda_not_ready = self.sa_session.query(model.Job.id).enable_eagerloads(False) \
.join(model.JobToInputLibraryDatasetAssociation) \
.join(model.LibraryDatasetDatasetAssociation) \
.join(model.Dataset) \
.filter(and_(model.Job.state == model.Job.states.NEW,
model.Dataset.state.in_(model.Dataset.non_ready_states))).subquery()
if self.app.config.user_activation_on:
jobs_to_check = self.sa_session.query(model.Job).enable_eagerloads(False) \
.outerjoin(model.User) \
.filter(and_((model.Job.state == model.Job.states.NEW),
or_((model.Job.user_id == null()), (model.User.active == true())),
(model.Job.handler == self.app.config.server_name),
~model.Job.table.c.id.in_(hda_not_ready),
~model.Job.table.c.id.in_(ldda_not_ready))) \
.order_by(model.Job.id).all()
else:
jobs_to_check = self.sa_session.query(model.Job).enable_eagerloads(False) \
.filter(and_((model.Job.state == model.Job.states.NEW),
(model.Job.handler == self.app.config.server_name),
~model.Job.table.c.id.in_(hda_not_ready),
~model.Job.table.c.id.in_(ldda_not_ready))) \
.order_by(model.Job.id).all()
# Filter jobs with invalid input states
jobs_to_check = self.__filter_jobs_with_invalid_input_states(jobs_to_check)
# Fetch all "resubmit" jobs
resubmit_jobs = self.sa_session.query(model.Job).enable_eagerloads(False) \
.filter(and_((model.Job.state == model.Job.states.RESUBMITTED),
(model.Job.handler == self.app.config.server_name))) \
.order_by(model.Job.id).all()
else:
# Get job objects and append to watch queue for any which were
# previously waiting
for job_id in self.waiting_jobs:
jobs_to_check.append(self.sa_session.query(model.Job).get(job_id))
try:
while 1:
message = self.queue.get_nowait()
if message is self.STOP_SIGNAL:
return
# Unpack the message
job_id, tool_id = message
# Get the job object and append to watch queue
jobs_to_check.append(self.sa_session.query(model.Job).get(job_id))
except Empty:
pass
# Ensure that we get new job counts on each iteration
self.__clear_job_count()
# Check resubmit jobs first so that limits of new jobs will still be enforced
for job in resubmit_jobs:
log.debug('(%s) Job was resubmitted and is being dispatched immediately', job.id)
# Reassemble resubmit job destination from persisted value
jw = self.__recover_job_wrapper(job)
if jw.is_ready_for_resubmission(job):
self.increase_running_job_count(job.user_id, jw.job_destination.id)
self.dispatcher.put(jw)
# Iterate over new and waiting jobs and look for any that are
# ready to run
new_waiting_jobs = []
for job in jobs_to_check:
try:
# Check the job's dependencies, requeue if they're not done.
# Some of these states will only happen when using the in-memory job queue
if job.copied_from_job_id:
copied_from_job = self.sa_session.query(model.Job).get(job.copied_from_job_id)
job.numeric_metrics = copied_from_job.numeric_metrics
job.text_metrics = copied_from_job.text_metrics
job.dependencies = copied_from_job.dependencies
job.state = copied_from_job.state
job.job_stderr = copied_from_job.job_stderr
job.job_stdout = copied_from_job.job_stdout
job.tool_stderr = copied_from_job.tool_stderr
job.tool_stdout = copied_from_job.tool_stdout
job.command_line = copied_from_job.command_line
job.traceback = copied_from_job.traceback
job.tool_version = copied_from_job.tool_version
job.exit_code = copied_from_job.exit_code
job.job_runner_name = copied_from_job.job_runner_name
job.job_runner_external_id = copied_from_job.job_runner_external_id
continue
job_state = self.__check_job_state(job)
if job_state == JOB_WAIT:
new_waiting_jobs.append(job.id)
elif job_state == JOB_INPUT_ERROR:
log.info("(%d) Job unable to run: one or more inputs in error state" % job.id)
elif job_state == JOB_INPUT_DELETED:
log.info("(%d) Job unable to run: one or more inputs deleted" % job.id)
elif job_state == JOB_READY:
self.dispatcher.put(self.job_wrappers.pop(job.id))
log.info("(%d) Job dispatched" % job.id)
elif job_state == JOB_DELETED:
log.info("(%d) Job deleted by user while still queued" % job.id)
elif job_state == JOB_ADMIN_DELETED:
log.info("(%d) Job deleted by admin while still queued" % job.id)
elif job_state in (JOB_USER_OVER_QUOTA,
JOB_USER_OVER_TOTAL_WALLTIME):
if job_state == JOB_USER_OVER_QUOTA:
log.info("(%d) User (%s) is over quota: job paused" % (job.id, job.user_id))
else:
log.info("(%d) User (%s) is over total walltime limit: job paused" % (job.id, job.user_id))
job.set_state(model.Job.states.PAUSED)
for dataset_assoc in job.output_datasets + job.output_library_datasets:
dataset_assoc.dataset.dataset.state = model.Dataset.states.PAUSED
dataset_assoc.dataset.info = "Execution of this dataset's job is paused because you were over your disk quota at the time it was ready to run"
self.sa_session.add(dataset_assoc.dataset.dataset)
self.sa_session.add(job)
elif job_state == JOB_ERROR:
log.error("(%d) Error checking job readiness" % job.id)
else:
log.error("(%d) Job in unknown state '%s'" % (job.id, job_state))
new_waiting_jobs.append(job.id)
except Exception:
log.exception("failure running job %d", job.id)
# Update the waiting list
if not self.track_jobs_in_database:
self.waiting_jobs = new_waiting_jobs
# Remove cached wrappers for any jobs that are no longer being tracked
for id in list(self.job_wrappers.keys()):
if id not in new_waiting_jobs:
del self.job_wrappers[id]
# Flush, if we updated the state
self.sa_session.flush()
# Done with the session
self.sa_session.remove()
def __filter_jobs_with_invalid_input_states(self, jobs):
"""
Takes list of jobs and filters out jobs whose input datasets are in invalid state and
set these jobs and their dependent jobs to paused.
"""
job_ids_to_check = [j.id for j in jobs]
queries = []
for job_to_input, input_association in [(model.JobToInputDatasetAssociation, model.HistoryDatasetAssociation),
(model.JobToInputLibraryDatasetAssociation, model.LibraryDatasetDatasetAssociation)]:
q = self.sa_session.query(
model.Job.id,
input_association.deleted,
input_association._state,
input_association.name,
model.Dataset.deleted,
model.Dataset.purged,
model.Dataset.state,
).join(job_to_input.job) \
.join(input_association) \
.join(model.Dataset) \
.filter(model.Job.id.in_(job_ids_to_check)) \
.filter(or_(model.Dataset.deleted == true(),
model.Dataset.state != model.Dataset.states.OK,
input_association.deleted == true(),
input_association._state == input_association.states.FAILED_METADATA
)).all()
queries.extend(q)
jobs_to_pause = defaultdict(list)
jobs_to_fail = defaultdict(list)
jobs_to_ignore = defaultdict(list)
for (job_id, hda_deleted, hda_state, hda_name, dataset_deleted, dataset_purged, dataset_state) in queries:
if hda_deleted or dataset_deleted:
if dataset_purged:
# If the dataset has been purged we can't resume the job by undeleting the input
jobs_to_fail[job_id].append("Input dataset '%s' was deleted before the job started" % hda_name)
else:
jobs_to_pause[job_id].append("Input dataset '%s' was deleted before the job started" % hda_name)
elif hda_state == model.HistoryDatasetAssociation.states.FAILED_METADATA:
jobs_to_pause[job_id].append("Input dataset '%s' failed to properly set metadata" % hda_name)
elif dataset_state == model.Dataset.states.PAUSED:
jobs_to_pause[job_id].append("Input dataset '%s' was paused before the job started" % hda_name)
elif dataset_state == model.Dataset.states.ERROR:
jobs_to_pause[job_id].append("Input dataset '%s' is in error state" % hda_name)
elif dataset_state != model.Dataset.states.OK:
jobs_to_ignore[job_id].append("Input dataset '%s' is in %s state" % (hda_name, dataset_state))
for job_id in sorted(jobs_to_pause):
pause_message = ", ".join(jobs_to_pause[job_id])
pause_message = "%s. To resume this job fix the input dataset(s)." % pause_message
job, job_wrapper = self.job_pair_for_id(job_id)
try:
job_wrapper.pause(job=job, message=pause_message)
except Exception:
log.exception("(%s) Caught exception while attempting to pause job.", job_id)
for job_id in sorted(jobs_to_fail):
fail_message = ", ".join(jobs_to_fail[job_id])
job, job_wrapper = self.job_pair_for_id(job_id)
try:
job_wrapper.fail(fail_message)
except Exception:
log.exception("(%s) Caught exception while attempting to fail job.", job_id)
jobs_to_ignore.update(jobs_to_pause)
jobs_to_ignore.update(jobs_to_fail)
return [j for j in jobs if j.id not in jobs_to_ignore]
def __check_job_state(self, job):
"""
Check if a job is ready to run by verifying that each of its input
datasets is ready (specifically in the OK state). If any input dataset
has an error, fail the job and return JOB_INPUT_ERROR. If any input
dataset is deleted, fail the job and return JOB_INPUT_DELETED. If all
input datasets are in OK state, return JOB_READY indicating that the
job can be dispatched. Otherwise, return JOB_WAIT indicating that input
datasets are still being prepared.
"""
if not self.track_jobs_in_database:
in_memory_not_ready_state = self.__verify_in_memory_job_inputs(job)
if in_memory_not_ready_state:
return in_memory_not_ready_state
# Else, if tracking in the database, job.state is guaranteed to be NEW and
# the inputs are guaranteed to be OK.
# Create the job wrapper so that the destination can be set
job_id = job.id
job_wrapper = self.job_wrappers.get(job_id, None)
if not job_wrapper:
job_wrapper = self.job_wrapper(job)
self.job_wrappers[job_id] = job_wrapper
# If state == JOB_READY, assume job_destination also set - otherwise
# in case of various error or cancelled states do not assume
# destination has been set.
state, job_destination = self.__verify_job_ready(job, job_wrapper)
if state == JOB_READY:
# PASS. increase usage by one job (if caching) so that multiple jobs aren't dispatched on this queue iteration
self.increase_running_job_count(job.user_id, job_destination.id)
for job_to_input_dataset_association in job.input_datasets:
# We record the input dataset version, now that we know the inputs are ready
if job_to_input_dataset_association.dataset:
job_to_input_dataset_association.dataset_version = job_to_input_dataset_association.dataset.version
return state
def __verify_job_ready(self, job, job_wrapper):
""" Compute job destination and verify job is ready at that
destination by checking job limits and quota. If this method
return a job state of JOB_READY - it MUST also return a job
destination.
"""
job_destination = None
try:
assert job_wrapper.tool is not None, 'This tool was disabled before the job completed. Please contact your Galaxy administrator.'
# Cause the job_destination to be set and cached by the mapper
job_destination = job_wrapper.job_destination
except AssertionError as e:
log.warning("(%s) Tool '%s' removed from tool config, unable to run job" % (job.id, job.tool_id))
job_wrapper.fail(e)
return JOB_ERROR, job_destination
except JobNotReadyException as e:
job_state = e.job_state or JOB_WAIT
return job_state, None
except Exception as e:
failure_message = getattr(e, 'failure_message', DEFAULT_JOB_PUT_FAILURE_MESSAGE)
if failure_message == DEFAULT_JOB_PUT_FAILURE_MESSAGE:
log.exception('Failed to generate job destination')
else:
log.debug("Intentionally failing job with message (%s)" % failure_message)
job_wrapper.fail(failure_message)
return JOB_ERROR, job_destination
# job is ready to run, check limits
# TODO: these checks should be refactored to minimize duplication and made more modular/pluggable
state = self.__check_destination_jobs(job, job_wrapper)
if state == JOB_READY:
state = self.__check_user_jobs(job, job_wrapper)
if state == JOB_READY and self.app.config.enable_quotas:
quota = self.app.quota_agent.get_quota(job.user)
if quota is not None:
try:
usage = self.app.quota_agent.get_usage(user=job.user, history=job.history)
if usage > quota:
return JOB_USER_OVER_QUOTA, job_destination
except AssertionError:
pass # No history, should not happen with an anon user
# Check total walltime limits
if (state == JOB_READY and
"delta" in self.app.job_config.limits.total_walltime):
jobs_to_check = self.sa_session.query(model.Job).filter(
model.Job.user_id == job.user.id,
model.Job.update_time >= datetime.datetime.now() -
datetime.timedelta(
self.app.job_config.limits.total_walltime["window"]
),
model.Job.state == 'ok'
).all()
time_spent = datetime.timedelta(0)
for job in jobs_to_check:
# History is job.state_history
started = None
finished = None
for history in sorted(
job.state_history,
key=lambda history: history.update_time):
if history.state == "running":
started = history.create_time
elif history.state == "ok":
finished = history.create_time
time_spent += finished - started
if time_spent > self.app.job_config.limits.total_walltime["delta"]:
return JOB_USER_OVER_TOTAL_WALLTIME, job_destination
return state, job_destination
def __verify_in_memory_job_inputs(self, job):
""" Perform the same checks that happen via SQL for in-memory managed
jobs.
"""
if job.state == model.Job.states.DELETED:
return JOB_DELETED
elif job.state == model.Job.states.ERROR:
return JOB_ADMIN_DELETED
for dataset_assoc in job.input_datasets + job.input_library_datasets:
idata = dataset_assoc.dataset
if not idata:
continue
# don't run jobs for which the input dataset was deleted
if idata.deleted:
self.job_wrappers.pop(job.id, self.job_wrapper(job)).fail("input data %s (file: %s) was deleted before the job started" % (idata.hid, idata.file_name))
return JOB_INPUT_DELETED
# an error in the input data causes us to bail immediately
elif idata.state == idata.states.ERROR:
self.job_wrappers.pop(job.id, self.job_wrapper(job)).fail("input data %s is in error state" % (idata.hid))
return JOB_INPUT_ERROR
elif idata.state == idata.states.FAILED_METADATA:
self.job_wrappers.pop(job.id, self.job_wrapper(job)).fail("input data %s failed to properly set metadata" % (idata.hid))
return JOB_INPUT_ERROR
elif idata.state != idata.states.OK and not (idata.state == idata.states.SETTING_METADATA and job.tool_id is not None and job.tool_id == self.app.datatypes_registry.set_external_metadata_tool.id):
# need to requeue
return JOB_WAIT
# All inputs ready to go.
return None
def __clear_job_count(self):
self.user_job_count = None
self.user_job_count_per_destination = None
self.total_job_count_per_destination = None
[docs] def get_user_job_count(self, user_id):
self.__cache_user_job_count()
# This could have been incremented by a previous job dispatched on this iteration, even if we're not caching
rval = self.user_job_count.get(user_id, 0)
if not self.app.config.cache_user_job_count:
result = self.sa_session.execute(select([func.count(model.Job.table.c.id)])
.where(and_(model.Job.table.c.state.in_((model.Job.states.QUEUED,
model.Job.states.RUNNING,
model.Job.states.RESUBMITTED)),
(model.Job.table.c.user_id == user_id))))
for row in result:
# there should only be one row
rval += row[0]
return rval
def __cache_user_job_count(self):
# Cache the job count if necessary
if self.user_job_count is None and self.app.config.cache_user_job_count:
self.user_job_count = {}
query = self.sa_session.execute(select([model.Job.table.c.user_id, func.count(model.Job.table.c.user_id)])
.where(and_(model.Job.table.c.state.in_((model.Job.states.QUEUED,
model.Job.states.RUNNING,
model.Job.states.RESUBMITTED)),
(model.Job.table.c.user_id != null())))
.group_by(model.Job.table.c.user_id))
for row in query:
self.user_job_count[row[0]] = row[1]
elif self.user_job_count is None:
self.user_job_count = {}
[docs] def get_user_job_count_per_destination(self, user_id):
self.__cache_user_job_count_per_destination()
cached = self.user_job_count_per_destination.get(user_id, {})
if self.app.config.cache_user_job_count:
rval = cached
else:
# The cached count is still used even when we're not caching, it is
# incremented when a job is run by this handler to ensure that
# multiple jobs can't get past the limits in one iteration of the
# queue.
rval = {}
rval.update(cached)
result = self.sa_session.execute(select([model.Job.table.c.destination_id, func.count(model.Job.table.c.destination_id).label('job_count')])
.where(and_(model.Job.table.c.state.in_((model.Job.states.QUEUED, model.Job.states.RUNNING)), (model.Job.table.c.user_id == user_id)))
.group_by(model.Job.table.c.destination_id))
for row in result:
# Add the count from the database to the cached count
rval[row['destination_id']] = rval.get(row['destination_id'], 0) + row['job_count']
return rval
def __cache_user_job_count_per_destination(self):
# Cache the job count if necessary
if self.user_job_count_per_destination is None and self.app.config.cache_user_job_count:
self.user_job_count_per_destination = {}
result = self.sa_session.execute(select([model.Job.table.c.user_id, model.Job.table.c.destination_id, func.count(model.Job.table.c.user_id).label('job_count')])
.where(and_(model.Job.table.c.state.in_((model.Job.states.QUEUED, model.Job.states.RUNNING))))
.group_by(model.Job.table.c.user_id, model.Job.table.c.destination_id))
for row in result:
if row['user_id'] not in self.user_job_count_per_destination:
self.user_job_count_per_destination[row['user_id']] = {}
self.user_job_count_per_destination[row['user_id']][row['destination_id']] = row['job_count']
elif self.user_job_count_per_destination is None:
self.user_job_count_per_destination = {}
[docs] def increase_running_job_count(self, user_id, destination_id):
if self.app.job_config.limits.registered_user_concurrent_jobs or \
self.app.job_config.limits.anonymous_user_concurrent_jobs or \
self.app.job_config.limits.destination_user_concurrent_jobs:
if self.user_job_count is None:
self.user_job_count = {}
if self.user_job_count_per_destination is None:
self.user_job_count_per_destination = {}
self.user_job_count[user_id] = self.user_job_count.get(user_id, 0) + 1
if user_id not in self.user_job_count_per_destination:
self.user_job_count_per_destination[user_id] = {}
self.user_job_count_per_destination[user_id][destination_id] = self.user_job_count_per_destination[user_id].get(destination_id, 0) + 1
if self.app.job_config.limits.destination_total_concurrent_jobs:
if self.total_job_count_per_destination is None:
self.total_job_count_per_destination = {}
self.total_job_count_per_destination[destination_id] = self.total_job_count_per_destination.get(destination_id, 0) + 1
def __check_user_jobs(self, job, job_wrapper):
# TODO: Update output datasets' _state = LIMITED or some such new
# state, so the UI can reflect what jobs are waiting due to concurrency
# limits
if job.user:
# Check the hard limit first
if self.app.job_config.limits.registered_user_concurrent_jobs:
count = self.get_user_job_count(job.user_id)
# Check the user's number of dispatched jobs against the overall limit
if count >= self.app.job_config.limits.registered_user_concurrent_jobs:
return JOB_WAIT
# If we pass the hard limit, also check the per-destination count
id = job_wrapper.job_destination.id
count_per_id = self.get_user_job_count_per_destination(job.user_id)
if id in self.app.job_config.limits.destination_user_concurrent_jobs:
count = count_per_id.get(id, 0)
# Check the user's number of dispatched jobs in the assigned destination id against the limit for that id
if count >= self.app.job_config.limits.destination_user_concurrent_jobs[id]:
return JOB_WAIT
# If we pass the destination limit (if there is one), also check limits on any tags (if any)
if job_wrapper.job_destination.tags:
for tag in job_wrapper.job_destination.tags:
# Check each tag for this job's destination
if tag in self.app.job_config.limits.destination_user_concurrent_jobs:
# Only if there's a limit defined for this tag
count = 0
for id in [d.id for d in self.app.job_config.get_destinations(tag)]:
# Add up the aggregate job total for this tag
count += count_per_id.get(id, 0)
if count >= self.app.job_config.limits.destination_user_concurrent_jobs[tag]:
return JOB_WAIT
elif job.galaxy_session:
# Anonymous users only get the hard limit
if self.app.job_config.limits.anonymous_user_concurrent_jobs:
count = self.sa_session.query(model.Job).enable_eagerloads(False) \
.filter(and_(model.Job.session_id == job.galaxy_session.id,
or_(model.Job.state == model.Job.states.RUNNING,
model.Job.state == model.Job.states.QUEUED))).count()
if count >= self.app.job_config.limits.anonymous_user_concurrent_jobs:
return JOB_WAIT
else:
log.warning('Job %s is not associated with a user or session so job concurrency limit cannot be checked.' % job.id)
return JOB_READY
def __cache_total_job_count_per_destination(self):
# Cache the job count if necessary
if self.total_job_count_per_destination is None:
self.total_job_count_per_destination = {}
result = self.sa_session.execute(select([model.Job.table.c.destination_id, func.count(model.Job.table.c.destination_id).label('job_count')])
.where(and_(model.Job.table.c.state.in_((model.Job.states.QUEUED, model.Job.states.RUNNING))))
.group_by(model.Job.table.c.destination_id))
for row in result:
self.total_job_count_per_destination[row['destination_id']] = row['job_count']
[docs] def get_total_job_count_per_destination(self):
self.__cache_total_job_count_per_destination()
# Always use caching (at worst a job will have to wait one iteration,
# and this would be more fair anyway as it ensures FIFO scheduling,
# insofar as FIFO would be fair...)
return self.total_job_count_per_destination
def __check_destination_jobs(self, job, job_wrapper):
if self.app.job_config.limits.destination_total_concurrent_jobs:
id = job_wrapper.job_destination.id
count_per_id = self.get_total_job_count_per_destination()
if id in self.app.job_config.limits.destination_total_concurrent_jobs:
count = count_per_id.get(id, 0)
# Check the number of dispatched jobs in the assigned destination id against the limit for that id
if count >= self.app.job_config.limits.destination_total_concurrent_jobs[id]:
return JOB_WAIT
# If we pass the destination limit (if there is one), also check limits on any tags (if any)
if job_wrapper.job_destination.tags:
for tag in job_wrapper.job_destination.tags:
# Check each tag for this job's destination
if tag in self.app.job_config.limits.destination_total_concurrent_jobs:
# Only if there's a limit defined for this tag
count = 0
for id in [d.id for d in self.app.job_config.get_destinations(tag)]:
# Add up the aggregate job total for this tag
count += count_per_id.get(id, 0)
if count >= self.app.job_config.limits.destination_total_concurrent_jobs[tag]:
return JOB_WAIT
return JOB_READY
def _handle_setup_msg(self, job_id=None):
job = self.sa_session.query(model.Job).get(job_id)
if job.handler is None:
job.handler = self.app.config.server_name
self.sa_session.add(job)
self.sa_session.flush()
# If not tracking jobs in the database
self.put(job.id, job.tool_id)
else:
log.warning("(%s) Handler '%s' received setup message but handler '%s' is already assigned, ignoring", job.id, self.app.config.server_name, job.handler)
[docs] def put(self, job_id, tool_id):
"""Add a job to the queue (by job identifier)"""
if not self.track_jobs_in_database:
self.queue.put((job_id, tool_id))
self.sleeper.wake()
[docs] def shutdown(self):
"""Attempts to gracefully shut down the worker thread"""
if self.parent_pid != os.getpid():
# We're not the real job queue, do nothing
return
else:
log.info("sending stop signal to worker thread")
self.stop_monitoring()
if not self.app.config.track_jobs_in_database:
self.queue.put(self.STOP_SIGNAL)
# A message could still be received while shutting down, should be ok since they will be picked up on next startup.
self.app.application_stack.deregister_message_handler(name=JobHandlerMessage.target)
self.sleeper.wake()
self.shutdown_monitor()
log.info("job handler queue stopped")
self.dispatcher.shutdown()
[docs]class JobHandlerStopQueue(Monitors):
"""
A queue for jobs which need to be terminated prematurely.
"""
STOP_SIGNAL = object()
[docs] def __init__(self, app, dispatcher):
self.app = app
self.dispatcher = dispatcher
self.sa_session = app.model.context
# Keep track of the pid that started the job manager, only it
# has valid threads
self.parent_pid = os.getpid()
# Contains new jobs. Note this is not used if track_jobs_in_database is True
self.queue = Queue()
# Contains jobs that are waiting (only use from monitor thread)
self.waiting = []
name = "JobHandlerStopQueue.monitor_thread"
self._init_monitor_thread(name, config=app.config)
log.info("job handler stop queue started")
[docs] def start(self):
# Start the queue
self.monitor_thread.start()
log.info("job handler stop queue started")
[docs] def monitor(self):
"""
Continually iterate the waiting jobs, stop any that are found.
"""
# HACK: Delay until after forking, we need a way to do post fork notification!!!
time.sleep(10)
while self.monitor_running:
try:
self.monitor_step()
except Exception:
log.exception("Exception in monitor_step")
# Sleep
self._monitor_sleep(1)
[docs] def monitor_step(self):
"""
Called repeatedly by `monitor` to stop jobs.
"""
# Pull all new jobs from the queue at once
jobs_to_check = []
if self.app.config.track_jobs_in_database:
# Clear the session so we get fresh states for job and all datasets
self.sa_session.expunge_all()
# Fetch all new jobs
newly_deleted_jobs = self.sa_session.query(model.Job).enable_eagerloads(False) \
.filter((model.Job.state == model.Job.states.DELETED_NEW) &
(model.Job.handler == self.app.config.server_name)).all()
for job in newly_deleted_jobs:
# job.stderr is always a string (job.job_stderr + job.tool_stderr, possibly `''`),
# while any `not None` message returned in self.queue.get_nowait() is interpreted
# as an error, so here we use None if job.stderr is false-y
jobs_to_check.append((job, job.stderr or None))
# Also pull from the queue (in the case of Administrative stopped jobs)
try:
while 1:
message = self.queue.get_nowait()
if message is self.STOP_SIGNAL:
return
# Unpack the message
job_id, error_msg = message
# Get the job object and append to watch queue
jobs_to_check.append((self.sa_session.query(model.Job).get(job_id), error_msg))
except Empty:
pass
for job, error_msg in jobs_to_check:
if (job.state not in
(job.states.DELETED_NEW,
job.states.DELETED) and
job.finished):
# terminated before it got here
log.debug('Job %s already finished, not deleting or stopping', job.id)
continue
final_state = job.states.DELETED
if error_msg is not None:
final_state = job.states.ERROR
job.info = error_msg
job.set_final_state(final_state)
self.sa_session.add(job)
self.sa_session.flush()
if job.job_runner_name is not None:
# tell the dispatcher to stop the job
job_wrapper = JobWrapper(job, self, use_persisted_destination=True)
self.dispatcher.stop(job, job_wrapper)
[docs] def put(self, job_id, error_msg=None):
if not self.app.config.track_jobs_in_database:
self.queue.put((job_id, error_msg))
[docs] def shutdown(self):
"""Attempts to gracefully shut down the worker thread"""
if self.parent_pid != os.getpid():
# We're not the real job queue, do nothing
return
else:
log.info("sending stop signal to worker thread")
self.stop_monitoring()
if not self.app.config.track_jobs_in_database:
self.queue.put(self.STOP_SIGNAL)
self.shutdown_monitor()
log.info("job handler stop queue stopped")
[docs]class DefaultJobDispatcher(object):
[docs] def __init__(self, app):
self.app = app
self.job_runners = self.app.job_config.get_job_runner_plugins(self.app.config.server_name)
# Once plugins are loaded, all job destinations that were created from
# URLs can have their URL params converted to the destination's param
# dict by the plugin.
self.app.job_config.convert_legacy_destinations(self.job_runners)
log.debug("Loaded job runners plugins: " + ':'.join(self.job_runners.keys()))
def __get_runner_name(self, job_wrapper):
if job_wrapper.can_split():
runner_name = "tasks"
else:
runner_name = job_wrapper.job_destination.runner
return runner_name
[docs] def url_to_destination(self, url):
"""This is used by the runner mapper (a.k.a. dynamic runner) and
recovery methods to have runners convert URLs to destinations.
New-style runner plugin IDs must match the URL's scheme for this to work.
"""
runner_name = url.split(':', 1)[0]
try:
return self.job_runners[runner_name].url_to_destination(url)
except Exception:
log.exception("Unable to convert legacy job runner URL '%s' to job destination, destination will be the '%s' runner with no params", url, runner_name)
return JobDestination(runner=runner_name)
[docs] def put(self, job_wrapper):
runner_name = self.__get_runner_name(job_wrapper)
try:
if isinstance(job_wrapper, TaskWrapper):
# DBTODO Refactor
log.debug("(%s) Dispatching task %s to %s runner" % (job_wrapper.job_id, job_wrapper.task_id, runner_name))
else:
log.debug("(%s) Dispatching to %s runner" % (job_wrapper.job_id, runner_name))
self.job_runners[runner_name].put(job_wrapper)
except KeyError:
log.error('put(): (%s) Invalid job runner: %s' % (job_wrapper.job_id, runner_name))
job_wrapper.fail(DEFAULT_JOB_PUT_FAILURE_MESSAGE)
[docs] def stop(self, job, job_wrapper):
"""
Stop the given job. The input variable job may be either a Job or a Task.
"""
# The Job and Task classes have been modified so that their accessors
# will return the appropriate value.
# Note that Jobs and Tasks have runner_names, which are distinct from
# the job_runner_name and task_runner_name.
# The runner name is not set until the job has started.
# If we're stopping a task, then the runner_name may be
# None, in which case it hasn't been scheduled.
job_runner_name = job.get_job_runner_name()
if job_runner_name is not None:
runner_name = job_runner_name.split(":", 1)[0]
log.debug("Stopping job %s in %s runner" % (job_wrapper.get_id_tag(), runner_name))
try:
self.job_runners[runner_name].stop_job(job_wrapper)
except KeyError:
log.error('stop(): (%s) Invalid job runner: %s' % (job_wrapper.get_id_tag(), runner_name))
# Job and output dataset states have already been updated, so nothing is done here.
[docs] def recover(self, job, job_wrapper):
runner_name = (job.job_runner_name.split(":", 1))[0]
log.debug("recovering job %d in %s runner" % (job.id, runner_name))
try:
self.job_runners[runner_name].recover(job, job_wrapper)
except KeyError:
log.error('recover(): (%s) Invalid job runner: %s' % (job_wrapper.job_id, runner_name))
job_wrapper.fail(DEFAULT_JOB_PUT_FAILURE_MESSAGE)
[docs] def shutdown(self):
failures = []
for name, runner in self.job_runners.items():
try:
runner.shutdown()
except Exception:
failures.append(name)
log.exception("Failed to shutdown runner %s", name)
if failures:
raise Exception("Failed to shutdown runners: %s" % ', '.join(failures))