def purge_files(min_size, purgable_files):
# we try to keep the cache as full as possible and defer deleting
# any files until we can fullfil an allocation request completely
# - since no request will start opportunisticly anyway. Also only
# delete enough to ensure min_size bytes.
bytes_freed = 0
bytes_freeable = sum(sf.size for sf in purgable_files)
requests_to_delete = set()
if min_size <= bytes_freeable:
for sf in purgable_files:
if bytes_freed >= min_size:
break
logger.debug('deleting %s' % sf)
# raises OSError on any error except if not existing:
if sf.staging_task:
util.unlink(sf.staging_task.path_out)
is_deleted = util.unlink(sf.path_staged())
if not is_deleted:
logger.warn('tried purging staged file %s - but it is already '
'gone' % sf)
sf.state = 'offline'
bytes_freed += sf.size
requests_to_delete.update(set(sf.requests))
return bytes_freed, requests_to_delete
def fail_request(r, msg):
r.state = 'failed'
for t in staging_tasks_owned_by(r):
logger.debug('revoking task %s belonging only to failed request %s' % \
(t, r))
t.cancel()
notify_user(r, msg)
mark_request_deletable.apply_async(args=[r.uuid],
countdown=config.REQUEST_PINNING_TIME)
def create_mars_request(verb, file_name, target=None):
sf = tasks.session.query(StagableFile).get(file_name)
logger.debug('creating MARS request from %s' % sf)
params = {}
for l in sf.params.split(',\n'):
k,v = l.split('=', 1)
params[k.strip()] = v.strip()
if target:
params['target'] = target
req = MarsRequest(verb, params)
logger.debug('MARS request is: %s' % req)
return req
def submit_sizing_tasks(r_uuid):
files_unknown_size = tasks.session.query(StagableFile).\
join(StagableFile.requests).\
filter(DownloadRequest.uuid == r_uuid).\
filter(StagableFile.size == None).\
filter(StagableFile.sizing_task == None)
logger.info('=> submitting sizing tasks for request %s: %s' % \
(r_uuid, ', '.join(sf.name for sf in files_unknown_size)))
for sf in files_unknown_size:
async_result = tasks.worker.estimate_size.delay(sf.name)
sf.sizing_task = Task(async_result.id)
logger.debug('submitted: %s' % sf.sizing_task)
tasks.session.commit()
def join_staging_task(self, task_id):
try:
staging_task = get_task(task_id)
logger.debug('staging task for %s is: %s' % (task_id, staging_task))
# BUG: FIXME: why is this a list and not a file, we do use uselist=False
sf = staging_task.stagable_file[0]
logger.info('staging task %s completed, file is %s' % (task_id, sf.name))
handle_staging_result(staging_task, sf)
tasks.session.commit()
except Exception, e:
logger.error('join_staging_task(task_id=%s) unexpectedly failed (task '
'id %s): %s, retrying in 60 s...' %
(task_id, join_staging_task.request.id, str(e)))
raise self.retry(exc=e, countdown=60)
def mark_request_deletable(self, r_uuid):
try:
r = tasks.session.query(DownloadRequest).get(r_uuid)
logger.info('marking %s request %s as deletable' % (r.state, r_uuid))
r.is_deletable = True
tasks.session.commit()
# since there might be pending requests that wait for space to
# be freed up:
logger.debug('=> invoking scheduler')
schedule_tasks.delay()
except Exception, e:
logger.error('mark_request_deletable(uuid=%s) unexpectedly failed '
'(task id %s): %s, retrying in 60 s...' %
(r_uuid, mark_request_deletable.request.id, str(e)))
raise self.retry(exc=e, countdown=60)
def estimate_size(self, file_name):
try:
logger.info('calling MARS to estimate size for %s' % file_name)
size = None
re_size = re.compile(r'^size=([\d]+);$')
mars_request = create_mars_request(verb='LIST', file_name=file_name)
mars_request.params['OUTPUT'] = 'COST'
logger.debug('mars_request: %s' % mars_request)
log_fn = logger.debug
for rc,fd,l in util.exec_proc([ 'mars' ], logger, stdin=str(mars_request)):
if fd == 2:
log_fn = logger.warn
if l is not None:
log_fn(l.strip())
if size is None and fd == 1 and l is not None:
m = re_size.match(l)
if m:
size = int(m.group(1).replace(",", ""))
assert rc is not None and fd is None and l is None
# don't trust size if mars returns non-zero
if rc != 0 or size is None:
logger.error('failed to compute size, rc = %d' % rc)
logger.error('MARS request was: %s' % mars_request)
raise TaskFailure('failed to compute size, rc = %d' % rc)
elif size == 0:
logger.error('size is %d' % size)
logger.error('MARS request was: %s' % mars_request)
raise HardTaskFailure('size is %d' % size)
est_size = size * config.FILE_SIZE_WEIGHT + config.FILE_SIZE_EXTRA
logger.debug('MARS reported size: %d bytes, after compensating: size * '
'%d + %d = %d' % \
(size, config.FILE_SIZE_WEIGHT, config.FILE_SIZE_EXTRA,
est_size))
logger.info('size is %d' % est_size)
return est_size
except HardTaskFailure:
raise
except Exception, e:
logger.error('estimate_size(file_name=%s) unexpectedly failed (task id '
'%s): %s, retrying in 60 s...' %
(file_name, estimate_size.request.id, str(e)))
raise self.retry(exc=e, countdown=60)
def register_request(self, uuid, openid, file_to_query):
try:
r = DownloadRequest(uuid, openid)
logger.debug('created new request %s' % r)
file_names = file_to_query.keys()
registered_files = tasks.session.query(StagableFile).\
filter(StagableFile.name.in_(file_names))
logger.debug('registered files: %s' % registered_files.all())
new_file_names = set(file_names) - set(x.name for x in registered_files)
logger.debug('new file names: %s' % new_file_names)
new_files = [ StagableFile(file_name, file_to_query[file_name]) for file_name in new_file_names ]
logger.debug('new files: %s' % new_files)
registered_files.update(
{ StagableFile.request_count : StagableFile.request_count + 1 },
synchronize_session=False)
r.files.extend(registered_files)
r.files.extend(new_files)
# This won't throw an integrityerror as long as this task is
# single threaded. If not, there's always a risk that one or more
# identical StagableFile may be inserted concurrently.
tasks.session.add(r)
tasks.session.commit()
logger.info('registered request %s openid=%s, file_names=%s '
'(unregistered=%s)' % \
(r.uuid, openid, file_names, new_file_names))
submit_sizing_tasks(r.uuid)
logger.debug('=> invoking scheduler')
schedule_tasks.delay()
except Exception, e:
# Note that some errors will never complete OK, for example if
# we try to register two or more requests with the same uuid
# (unlikely) - these cases are not handled and will need
# manual admin intervention (in this case removal of the
# redundant requests)
logger.error('register_request(uuid=%s, openid=%s, file_to_query=%s) '
'unexpectedly failed (task id %s): %s, retrying in 60 s...' %
(uuid, openid, file_to_query, register_request.request.id,
str(e)))
raise self.retry(exc=e, countdown=60)
def join_sizing_tasks(files):
with allow_join_result():
for sf in files:
try:
assert sf.size is None, sf
logger.debug('<= awaiting size estimation result for %s, '
'task is %s' % (sf, sf.sizing_task))
if sf.sizing_task is None:
# this means it failed before and the task got
# unregistered from sf
raise TaskFailure('size estimation has failed - a new task '
'will be created on the next request '
'containing this file')
size = sf.sizing_task.get()
logger.debug('size estimated to %d bytes' % size)
assert size >= 0, size
sf.size = size
except Exception, e:
logger.warning('size estimation task %s failed: %s' % \
(sf.sizing_task, e))
raise e
finally:
def handle_staging_result(staging_task, sf):
with allow_join_result():
logger.debug('getting staging result')
try:
staging_task.get() # returns None, but more importantly propagates any exception in the task
except Exception, e:
logger.warning('%s failed: %s' % (staging_task, e))
# consider resubmitting the task a limited amount since it
# should only fail in rare cases (if so then do this by
# raising a RetryException or similar here)
logger.debug('deregistering %s from %s' % (sf.staging_task, sf))
tasks.session.delete(staging_task)
sf.staging_task = None
assert sf.staging_task is None, sf
logger.debug('db commit')
tasks.session.commit()
for r in dispatched_requests(sf):
logger.info('request %s failed' % r.uuid)
fail_request(r, 'Staging of %s failed: %s' % (sf, str(e)))
return # we deliberately only handle exceptions from the staging task, any other exception will propagate
def stage_file(self, file_name, target_dir, path_out):
try:
end_target = os.path.join(target_dir, file_name)
tmp_target = os.path.join(target_dir, uuid.uuid4().get_hex())
logger.info('staging %s' % end_target)
logger.debug('tmp_target: %s' % tmp_target)
mars_request = create_mars_request(verb='RETRIEVE',
file_name=file_name,
target=tmp_target)
logger.debug('mars_request: %s' % mars_request)
with open(path_out, 'w') as f:
for rc,fd,l in util.exec_proc([ 'mars' ], logger, stdin=str(mars_request)):
if fd is not None and l is not None:
if fd == 1:
logger.debug('fd=%s, l=%s' % (fd, l.strip() if l else l))
else:
logger.warning('fd=%s, l=%s' % (fd, l.strip() if l else l))
f.write(l)
f.flush()
f.flush()
os.fdatasync(f.fileno())
f.close()
if rc != 0:
logger.error('failed to stage %s, rc = %d' % (file_name, rc))
logger.error('MARS request was: %s' % mars_request)
logger.debug('removing temp file %s' % tmp_target)
util.unlink(tmp_target) # FIXME: use try...finally
raise TaskFailure('mars returned %d' % rc)
logger.debug('moving temp file %s -> %s' % (tmp_target, end_target))
os.rename(tmp_target, end_target)
logger.info('%s is staged online' % end_target)
logger.debug('=> invoking scheduler')
tasks.scheduler.join_staging_task.delay(stage_file.request.id)
except Exception, e:
logger.error('stage_file(file_name=%s, target_dir=%s, path_out=%s) '
'unexpectedly failed (task id %s): %s, retrying in '
'60 s...' %
(file_name, target_dir, path_out, stage_file.request.id,
str(e)))
raise self.retry(exc=e, countdown=60)
def schedule_tasks(self):
try:
dispatchable_requests = get_dispatchable_requests()
logger.info('running scheduler - %d dispatchable request(s): %s' % \
( dispatchable_requests.count(),
', '.join(r.uuid for r in dispatchable_requests)))
reserved_files = get_reserved_files()
online_files = get_online_files()
reserved_space = sum(sf.size for sf in reserved_files)
used_space = sum(sf.size for sf in online_files)
available_space = config.STAGE_SPACE - used_space
purgable_files = get_purgable_files()
purgable_amount = sum(sf.size for sf in purgable_files)
assert len(reserved_files.all()) == len(set(reserved_files)), \
reserved_files.all()
assert used_space >= 0, used_space
assert reserved_space >= 0, (reserved_files, reserved_space)
logger.debug('reserved files: %s' % ', '.join(sf.name for sf in reserved_files))
logger.debug('online files: %s' % ', '.join(sf.name for sf in online_files))
logger.debug('purgable amount: %d bytes from %s' % \
(purgable_amount, ', '.join(sf.name for sf in purgable_files)))
logger.info('total staging space: %d bytes, used: %d bytes, reserved: %d '
'bytes. Max %d bytes available for new requests (%d bytes '
'purgable)' % \
(config.STAGE_SPACE, used_space, reserved_space,
config.STAGE_SPACE - reserved_space, purgable_amount))
dispatch_tasks = True
num_tasks_dispatched = 0
num_tasks_failed = 0
num_tasks_deferred = 0
consumed_space = 0
for rs in dispatchable_requests:
try:
files_offline_not_being_staged = get_files_offline_not_being_staged(rs)
logger.info('scheduling %s, available space %d bytes, offline files: %s' % \
(rs.uuid, available_space,
', '.join(sf.name for sf in files_offline_not_being_staged)))
logger.debug('offline files: %s' % \
files_offline_not_being_staged.all())
logger.info('waiting for all size estimation tasks...')
join_sizing_tasks(files_offline_not_being_staged.\
filter(StagableFile.size == None))
except Exception, e:
# Note: all requests will keep on failing until this task has been resubmitted
logger.warning('sizing estimation of %s failed: %s, request -> '
'failed' % (rs.uuid, e))
fail_request(rs, 'sizing estimation of %s failed: %s' % (rs.uuid, e))
num_tasks_failed += 1
continue
assert all(f.size is not None for f in files_offline_not_being_staged),\
files_offline_not_being_staged
offline_size = sum(f.size for f in files_offline_not_being_staged)
total_size = sum(f.size for f in rs.files)
logger.info('total request size: %d bytes, %d offline bytes' % \
(total_size, offline_size))
# if staging fails any later request will we must make sure
# either to fail any later requests for that file
# fastforward all dispatchable zero cost requests and any
# requests that is impossible to fulfill:
# created -> failed,
# created -> dispatched,
# created -> finished:
if total_size > config.STAGE_SPACE:
logger.info('fast forwarding %s -> failed - since there is no way '
'it can be fulfilled (needs %d of %d bytes '
'available)' % \
(rs.uuid, total_size, config.STAGE_SPACE))
fail_request(rs, '%s can not be fulfilled (needs %d of %d bytes '
'available)' % (rs.uuid, total_size, config.STAGE_SPACE))
num_tasks_failed += 1
elif not files_offline_not_being_staged.first():
if get_files_offline_being_staged(rs).first():
logger.info('fast forwarding %s -> dispatched - since it '
'fully overlaps with staging tasks in '
'progress' % rs.uuid)
rs.state = 'dispatched'
num_tasks_dispatched += 1
else:
# this should not fail unless db is being concurrently
# updated:
assert all(x.state == 'online' for x in rs.files), rs
logger.info('fast forwarding %s -> finished - since all '
'files are online' % rs.uuid)
finish_request(rs)
num_tasks_dispatched += 1
elif dispatch_tasks:
if offline_size > available_space:
logger.info('%s requires more space than what is available '
'(%d > %d), initiating cache purging' % \
( rs.uuid, offline_size, available_space ))
required_amount = offline_size - available_space
bytes_freed, requests_to_delete = purge_files(required_amount,
purgable_files)
available_space += bytes_freed
logger.info('freed up %d/%d bytes (%d%%) - available space %d '
'bytes' % \
( bytes_freed, required_amount,
int(round(bytes_freed / float(offline_size) * 100, 3)),
available_space ))
if requests_to_delete:
assert all(x.is_deletable for x in requests_to_delete), requests_to_delete
logger.info('deleting all deletable requests from db '
'belonging to files that have been '
'purged: %s' % \
', '.join(x.uuid for x in requests_to_delete))
tasks.session.query(DownloadRequest).\
filter(DownloadRequest.uuid.in_(x.uuid for x in requests_to_delete)).\
delete(synchronize_session='fetch')
if offline_size > available_space:
logger.info('%d bytes needed but only %d bytes available -'
' not considering any further requests for task'
' dispatching until the next request has '
'finished/failed' % \
(offline_size, available_space))
dispatch_tasks = False
# schedule_tasks will be triggered upon next
# request failure or deletion after finish
num_tasks_deferred += 1
# optimisation possibility: we can also dispatch as many
# tasks as there currently is room for even if the request
# can not be fully staged yet (but that would also
# complicate how we compute reserved/available_space)
if dispatch_tasks:
logger.info('staging offline files for request %s: %s' % \
(rs.uuid,
', '.join(sf.name for sf in files_offline_not_being_staged)))
for sf in files_offline_not_being_staged:
path_out = util.create_tempfile()
# we'd like to chain the join_staging_tasks here
# rather than calling it from the stage_file task,
# but then we cannot register the async result
# like this...
async_result = tasks.worker.stage_file.delay(sf.name,
sf.path,
path_out)
# note the critical window if worker dies here the
# task may have started but will not be registered
# with f (the same goes for submitting sizing
# tasks). We choose risking more than one
# identical staging task running (which is
# harmless and has a low probability, though
# wastes some resources) rather than risking
# having the task registered with the file but not
# actually running (which is harder to handle)
#
# (possibly create a celery task group consisting
# of staging -> checksumming):
sf.staging_task = Task(async_result.id, path_out)
# IMPORTANT: commit immediately since stage_file
# calls join_staging_tasks which looks up the task
# in the db:
tasks.session.commit()
logger.info('=> staging task for %s is %s' % \
(sf.name, sf.staging_task.uuid))
consumed_space += offline_size
available_space -= offline_size
rs.state = 'dispatched'
num_tasks_dispatched += 1
else:
logger.info('request %s is not eliglible for dispatching since it '
'can not be fast forwarded and at least one higher '
'priority request is still waiting to be dispatched' % \
rs.uuid)
num_tasks_deferred += 1
logger.info('scheduling iteration completed, %d tasks dispatched, %d '
'tasks failed, %d tasks deferred. Consumed space %d bytes, '
'total space available %d bytes' % \
(num_tasks_dispatched, num_tasks_failed, num_tasks_deferred,
consumed_space, available_space))
tasks.session.commit()
if sf.sizing_task is None:
# this means it failed before and the task got
# unregistered from sf
raise TaskFailure('size estimation has failed - a new task '
'will be created on the next request '
'containing this file')
size = sf.sizing_task.get()
logger.debug('size estimated to %d bytes' % size)
assert size >= 0, size
sf.size = size
except Exception, e:
logger.warning('size estimation task %s failed: %s' % \
(sf.sizing_task, e))
raise e
finally:
logger.debug('unregistering size estimation task %s' % \
(sf.sizing_task))
tasks.session.delete(sf.sizing_task)
sf.sizing_task = None
@tasks.cel.task(bind=True, acks_late=True, ignore_results=True)
def mark_request_deletable(self, r_uuid):
try:
r = tasks.session.query(DownloadRequest).get(r_uuid)
logger.info('marking %s request %s as deletable' % (r.state, r_uuid))
r.is_deletable = True
tasks.session.commit()
# since there might be pending requests that wait for space to
# be freed up:
logger.debug('=> invoking scheduler')
schedule_tasks.delay()
