def test_scope_queue_name(self):
self.assertEqual(utils.scope_queue_name('my-q'), '/my-q')
self.assertEqual(utils.scope_queue_name('my-q', None), '/my-q')
self.assertEqual(utils.scope_queue_name('my-q', '123'), '123/my-q')
self.assertEqual(utils.scope_queue_name(None), '/')
self.assertEqual(utils.scope_queue_name(None, '123'), '123/')
def test_scope_queue_name(self):
self.assertEqual(utils.scope_queue_name('my-q'), '/my-q')
self.assertEqual(utils.scope_queue_name('my-q', None), '/my-q')
self.assertEqual(utils.scope_queue_name('my-q', '123'), '123/my-q')
self.assertEqual(utils.scope_queue_name(None), '/')
self.assertEqual(utils.scope_queue_name(None, '123'), '123/')
def create(self, name, project=None):
# NOTE(flaper87): If the connection fails after it was called
# and we retry to insert the queue, we could end up returning
# `False` because of the `DuplicatedKeyError` although the
# queue was indeed created by this API call.
#
# TODO(kgriffs): Commented out `retries_on_autoreconnect` for
# now due to the above issue, since creating a queue is less
# important to make super HA.
try:
# NOTE(kgriffs): Start counting at 1, and assume the first
# message ever posted will succeed and set t to a UNIX
# "modified at" timestamp.
counter = {'v': 1, 't': 0}
scoped_name = utils.scope_queue_name(name, project)
self._collection.insert({
'p_q': scoped_name,
'm': {},
'c': counter
})
except pymongo.errors.DuplicateKeyError:
return False
else:
return True
def _count(self, queue_name, project=None, include_claimed=False):
"""Return total number of messages in a queue.
This method is designed to very quickly count the number
of messages in a given queue. Expired messages are not
counted, of course. If the queue does not exist, the
count will always be 0.
Note: Some expired messages may be included in the count if
they haven't been GC'd yet. This is done for performance.
"""
query = {
# Messages must belong to this queue and project.
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
# NOTE(kgriffs): Messages must be finalized (i.e., must not
# be part of an unfinalized transaction).
#
# See also the note wrt 'tx' within the definition
# of ACTIVE_INDEX_FIELDS.
'tx': None,
}
if not include_claimed:
# Exclude messages that are claimed
query['c.e'] = {'$lte': timeutils.utcnow_ts()}
collection = self._collection(queue_name, project)
return collection.find(query).hint(COUNTING_INDEX_FIELDS).count()
def bulk_get(self, queue_name, message_ids, project=None):
message_ids = [mid for mid in map(utils.to_oid, message_ids) if mid]
if not message_ids:
return iter([])
now = timeutils.utcnow_ts()
# Base query, always check expire time
query = {
'_id': {
'$in': message_ids
},
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
}
collection = self._collection(queue_name, project)
# NOTE(flaper87): Should this query
# be sorted?
messages = collection.find(query).hint(ID_INDEX_FIELDS)
def denormalizer(msg):
return _basic_message(msg, now)
return utils.HookedCursor(messages, denormalizer)
def _count(self, queue_name, project=None, include_claimed=False):
"""Return total number of messages in a queue.
This method is designed to very quickly count the number
of messages in a given queue. Expired messages are not
counted, of course. If the queue does not exist, the
count will always be 0.
Note: Some expired messages may be included in the count if
they haven't been GC'd yet. This is done for performance.
"""
query = {
# Messages must belong to this queue and project.
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
# NOTE(kgriffs): Messages must be finalized (i.e., must not
# be part of an unfinalized transaction).
#
# See also the note wrt 'tx' within the definition
# of ACTIVE_INDEX_FIELDS.
'tx': None,
}
if not include_claimed:
# Exclude messages that are claimed
query['c.e'] = {'$lte': timeutils.utcnow_ts()}
collection = self._collection(queue_name, project)
return collection.find(query).hint(COUNTING_INDEX_FIELDS).count()
def _claimed(self, queue_name, claim_id,
expires=None, limit=None, project=None):
if claim_id is None:
claim_id = {'$ne': None}
query = {
'p_q': utils.scope_queue_name(queue_name, project),
'c.id': claim_id,
'c.e': {'$gt': expires or timeutils.utcnow_ts()},
}
# NOTE(kgriffs): Claimed messages bust be queried from
# the primary to avoid a race condition caused by the
# multi-phased "create claim" algorithm.
preference = pymongo.read_preferences.ReadPreference.PRIMARY
collection = self._collection(queue_name, project)
msgs = collection.find(query, sort=[('k', 1)],
read_preference=preference)
if limit is not None:
msgs = msgs.limit(limit)
now = timeutils.utcnow_ts()
def denormalizer(msg):
doc = _basic_message(msg, now)
doc['claim'] = msg['c']
return doc
return utils.HookedCursor(msgs, denormalizer)
def _unclaim(self, queue_name, claim_id, project=None):
cid = utils.to_oid(claim_id)
# NOTE(cpp-cabrera): early abort - avoid a DB query if we're handling
# an invalid ID
if cid is None:
return
# NOTE(cpp-cabrera): unclaim by setting the claim ID to None
# and the claim expiration time to now
now = timeutils.utcnow_ts()
scope = utils.scope_queue_name(queue_name, project)
collection = self._collection(queue_name, project)
collection.update({
PROJ_QUEUE: scope,
'c.id': cid
}, {'$set': {
'c': {
'id': None,
'e': now
}
}},
upsert=False,
multi=True)
def _claimed(self, queue_name, claim_id,
expires=None, limit=None, project=None):
if claim_id is None:
claim_id = {'$ne': None}
query = {
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
'c.id': claim_id,
'c.e': {'$gt': expires or timeutils.utcnow_ts()},
}
# NOTE(kgriffs): Claimed messages bust be queried from
# the primary to avoid a race condition caused by the
# multi-phased "create claim" algorithm.
preference = pymongo.read_preferences.ReadPreference.PRIMARY
collection = self._collection(queue_name, project)
msgs = collection.find(query, sort=[('k', 1)],
read_preference=preference).hint(
CLAIMED_INDEX_FIELDS
)
if limit is not None:
msgs = msgs.limit(limit)
now = timeutils.utcnow_ts()
def denormalizer(msg):
doc = _basic_message(msg, now)
doc['claim'] = msg['c']
return doc
return utils.HookedCursor(msgs, denormalizer)
def bulk_delete(self, queue_name, message_ids, project=None):
message_ids = [mid for mid in map(utils.to_oid, message_ids) if mid]
query = {
'_id': {'$in': message_ids},
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
}
collection = self._collection(queue_name, project)
collection.remove(query, w=0)
def bulk_delete(self, queue_name, message_ids, project=None):
message_ids = [mid for mid in map(utils.to_oid, message_ids) if mid]
query = {
'_id': {'$in': message_ids},
'p_q': utils.scope_queue_name(queue_name, project),
}
collection = self._collection(queue_name, project)
collection.remove(query, w=0)
def get(self, project, queue):
fields = {'_id': 0}
key = utils.scope_queue_name(queue, project)
entry = self._col.find_one({PRIMARY_KEY: key}, fields=fields)
if entry is None:
raise errors.QueueNotMapped(project, queue)
return _normalize(entry)
def update(self, queue, claim_id, metadata, project=None):
cid = utils.to_oid(claim_id)
if cid is None:
raise exceptions.ClaimDoesNotExist(claim_id, queue, project)
now = timeutils.utcnow_ts()
ttl = int(metadata.get('ttl', 60))
expires = now + ttl
msg_ctrl = self.driver.message_controller
claimed = msg_ctrl._claimed(queue,
cid,
expires=now,
limit=1,
project=project)
try:
next(claimed)
except StopIteration:
raise exceptions.ClaimDoesNotExist(claim_id, queue, project)
meta = {
'id': cid,
't': ttl,
'e': expires,
}
# TODO(kgriffs): Create methods for these so we don't interact
# with the messages collection directly (loose coupling)
scope = utils.scope_queue_name(queue, project)
collection = msg_ctrl._collection(queue, project)
collection.update({
'p_q': scope,
'c.id': cid
}, {'$set': {
'c': meta
}},
upsert=False,
multi=True)
# NOTE(flaper87): Dirty hack!
# This sets the expiration time to
# `expires` on messages that would
# expire before claim.
collection.update({
'p_q': scope,
'e': {
'$lt': expires
},
'c.id': cid
}, {'$set': {
'e': expires,
't': ttl
}},
upsert=False,
multi=True)
def get(self, project, queue):
fields = {'_id': 0}
key = utils.scope_queue_name(queue, project)
entry = self._col.find_one({PRIMARY_KEY: key},
fields=fields)
if entry is None:
raise errors.QueueNotMapped(queue, project)
return _normalize(entry)
def _purge_queue(self, queue_name, project=None):
"""Removes all messages from the queue.
Warning: Only use this when deleting the queue; otherwise
you can cause a side-effect of reseting the marker counter
which can cause clients to miss tons of messages.
If the queue does not exist, this method fails silently.
:param queue_name: name of the queue to purge
:param project: ID of the project to which the queue belongs
"""
scope = utils.scope_queue_name(queue_name, project)
collection = self._collection(queue_name, project)
collection.remove({PROJ_QUEUE: scope}, w=0)
def _purge_queue(self, queue_name, project=None):
"""Removes all messages from the queue.
Warning: Only use this when deleting the queue; otherwise
you can cause a side-effect of reseting the marker counter
which can cause clients to miss tons of messages.
If the queue does not exist, this method fails silently.
:param queue_name: name of the queue to purge
:param project: ID of the project to which the queue belongs
"""
scope = utils.scope_queue_name(queue_name, project)
collection = self._collection(queue_name, project)
collection.remove({PROJ_QUEUE: scope}, w=0)
def create(self, name, project=None):
try:
# NOTE(kgriffs): Start counting at 1, and assume the first
# message ever posted will succeed and set t to a UNIX
# "modified at" timestamp.
counter = {'v': 1, 't': 0}
scoped_name = utils.scope_queue_name(name, project)
self._collection.insert({'p_q': scoped_name, 'm': {},
'c': counter})
except pymongo.errors.DuplicateKeyError:
return False
else:
return True
def delete(self, queue_name, message_id, project=None, claim=None):
# NOTE(cpp-cabrera): return early - this is an invalid message
# id so we won't be able to find it any way
mid = utils.to_oid(message_id)
if mid is None:
return
collection = self._collection(queue_name, project)
query = {
'_id': mid,
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
}
# NOTE(cpp-cabrera): return early - the user gaves us an
# invalid claim id and that renders the rest of this
# request moot
cid = utils.to_oid(claim)
if cid is None:
return
now = timeutils.utcnow_ts()
cursor = collection.find(query).hint(ID_INDEX_FIELDS)
try:
message = next(cursor)
except StopIteration:
return
is_claimed = (message['c']['id'] is not None
and message['c']['e'] > now)
if claim is None:
if is_claimed:
raise errors.MessageIsClaimed(message_id)
else:
if message['c']['id'] != cid:
# NOTE(kgriffs): Read from primary in case the message
# was just barely claimed, and claim hasn't made it to
# the secondary.
pref = pymongo.read_preferences.ReadPreference.PRIMARY
message = collection.find_one(query, read_preference=pref)
if message['c']['id'] != cid:
raise errors.MessageIsClaimedBy(message_id, claim)
collection.remove(query['_id'], w=0)
def delete(self, queue_name, message_id, project=None, claim=None):
# NOTE(cpp-cabrera): return early - this is an invalid message
# id so we won't be able to find it any way
mid = utils.to_oid(message_id)
if mid is None:
return
collection = self._collection(queue_name, project)
query = {
'_id': mid,
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
}
# NOTE(cpp-cabrera): return early - the user gaves us an
# invalid claim id and that renders the rest of this
# request moot
cid = utils.to_oid(claim)
if cid is None:
return
now = timeutils.utcnow_ts()
cursor = collection.find(query).hint(ID_INDEX_FIELDS)
try:
message = next(cursor)
except StopIteration:
return
is_claimed = (message['c']['id'] is not None and
message['c']['e'] > now)
if claim is None:
if is_claimed:
raise errors.MessageIsClaimed(message_id)
else:
if message['c']['id'] != cid:
# NOTE(kgriffs): Read from primary in case the message
# was just barely claimed, and claim hasn't made it to
# the secondary.
pref = pymongo.read_preferences.ReadPreference.PRIMARY
message = collection.find_one(query, read_preference=pref)
if message['c']['id'] != cid:
raise errors.MessageIsClaimedBy(message_id, claim)
collection.remove(query['_id'], w=0)
def get(self, queue_name, message_id, project=None):
mid = utils.to_oid(message_id)
if mid is None:
raise errors.MessageDoesNotExist(message_id, queue_name, project)
now = timeutils.utcnow_ts()
query = {"_id": mid, PROJ_QUEUE: utils.scope_queue_name(queue_name, project)}
collection = self._collection(queue_name, project)
message = list(collection.find(query).limit(1).hint(ID_INDEX_FIELDS))
if not message:
raise errors.MessageDoesNotExist(message_id, queue_name, project)
return _basic_message(message[0], now)
def _unclaim(self, queue_name, claim_id, project=None):
cid = utils.to_oid(claim_id)
# NOTE(cpp-cabrera): early abort - avoid a DB query if we're handling
# an invalid ID
if cid is None:
return
# NOTE(cpp-cabrera): unclaim by setting the claim ID to None
# and the claim expiration time to now
now = timeutils.utcnow_ts()
scope = utils.scope_queue_name(queue_name, project)
collection = self._collection(queue_name, project)
collection.update({PROJ_QUEUE: scope, 'c.id': cid},
{'$set': {'c': {'id': None, 'e': now}}},
upsert=False, multi=True)
def pop(self, queue_name, limit, project=None):
query = {PROJ_QUEUE: utils.scope_queue_name(queue_name, project)}
# Only include messages that are not part of
# any claim, or are part of an expired claim.
now = timeutils.utcnow_ts()
query["c.e"] = {"$lte": now}
collection = self._collection(queue_name, project)
fields = {"_id": 1, "t": 1, "b": 1}
messages = (collection.find_and_modify(query, fields=fields, remove=True) for _ in range(limit))
messages = itertools.ifilter(None, messages)
final_messages = [_basic_message(message, now) for message in messages]
return final_messages
def create(self, name, project=None):
try:
# NOTE(kgriffs): Start counting at 1, and assume the first
# message ever posted will succeed and set t to a UNIX
# "modified at" timestamp.
counter = {'v': 1, 't': 0}
scoped_name = utils.scope_queue_name(name, project)
self._collection.insert({
'p_q': scoped_name,
'm': {},
'c': counter
})
except pymongo.errors.DuplicateKeyError:
return False
else:
return True
def get(self, queue_name, message_id, project=None):
mid = utils.to_oid(message_id)
if mid is None:
raise errors.MessageDoesNotExist(message_id, queue_name, project)
now = timeutils.utcnow_ts()
query = {
'_id': mid,
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
}
collection = self._collection(queue_name, project)
message = list(collection.find(query).limit(1).hint(ID_INDEX_FIELDS))
if not message:
raise errors.MessageDoesNotExist(message_id, queue_name, project)
return _basic_message(message[0], now)
def bulk_get(self, queue_name, message_ids, project=None):
message_ids = [mid for mid in map(utils.to_oid, message_ids) if mid]
if not message_ids:
return iter([])
now = timeutils.utcnow_ts()
# Base query, always check expire time
query = {"_id": {"$in": message_ids}, PROJ_QUEUE: utils.scope_queue_name(queue_name, project)}
collection = self._collection(queue_name, project)
# NOTE(flaper87): Should this query
# be sorted?
messages = collection.find(query).hint(ID_INDEX_FIELDS)
def denormalizer(msg):
return _basic_message(msg, now)
return utils.HookedCursor(messages, denormalizer)
def update(self, queue, claim_id, metadata, project=None):
cid = utils.to_oid(claim_id)
if cid is None:
raise errors.ClaimDoesNotExist(claim_id, queue, project)
now = timeutils.utcnow_ts()
ttl = int(metadata.get('ttl', 60))
expires = now + ttl
msg_ctrl = self.driver.message_controller
claimed = msg_ctrl._claimed(queue, cid, expires=now,
limit=1, project=project)
try:
next(claimed)
except StopIteration:
raise errors.ClaimDoesNotExist(claim_id, queue, project)
meta = {
'id': cid,
't': ttl,
'e': expires,
}
# TODO(kgriffs): Create methods for these so we don't interact
# with the messages collection directly (loose coupling)
scope = utils.scope_queue_name(queue, project)
collection = msg_ctrl._collection(queue, project)
collection.update({'p_q': scope, 'c.id': cid},
{'$set': {'c': meta}},
upsert=False, multi=True)
# NOTE(flaper87): Dirty hack!
# This sets the expiration time to
# `expires` on messages that would
# expire before claim.
collection.update({'p_q': scope,
'e': {'$lt': expires},
'c.id': cid},
{'$set': {'e': expires, 't': ttl}},
upsert=False, multi=True)
def _count(self, queue_name, project=None, include_claimed=False):
"""Return total number of messages in a queue.
This method is designed to very quickly count the number
of messages in a given queue. Expired messages are not
counted, of course. If the queue does not exist, the
count will always be 0.
Note: Some expired messages may be included in the count if
they haven't been GC'd yet. This is done for performance.
"""
query = {
# Messages must belong to this queue
PROJ_QUEUE: utils.scope_queue_name(queue_name, project)
}
if not include_claimed:
# Exclude messages that are claimed
query["c.e"] = {"$lte": timeutils.utcnow_ts()}
collection = self._collection(queue_name, project)
return collection.find(query).hint(COUNTING_INDEX_FIELDS).count()
def delete(self, queue_name, message_id, project=None, claim=None):
# NOTE(cpp-cabrera): return early - this is an invalid message
# id so we won't be able to find it any way
mid = utils.to_oid(message_id)
if mid is None:
return
collection = self._collection(queue_name, project)
query = {
'_id': mid,
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
}
# NOTE(cpp-cabrera): return early - the user gaves us an
# invalid claim id and that renders the rest of this
# request moot
cid = utils.to_oid(claim)
if cid is None:
return
now = timeutils.utcnow_ts()
message = collection.find_one(query)
if message is None:
return
is_claimed = (message['c']['id'] is not None and
message['c']['e'] > now)
if claim is None:
if is_claimed:
raise errors.MessageIsClaimed(message_id)
else:
if message['c']['id'] != cid:
raise errors.MessageIsClaimedBy(message_id, claim)
collection.remove(query['_id'], w=0)
def delete(self, queue_name, message_id, project=None, claim=None):
# NOTE(cpp-cabrera): return early - this is an invalid message
# id so we won't be able to find it any way
mid = utils.to_oid(message_id)
if mid is None:
return
collection = self._collection(queue_name, project)
query = {
'_id': mid,
'p_q': utils.scope_queue_name(queue_name, project),
}
# NOTE(cpp-cabrera): return early - the user gaves us an
# invalid claim id and that renders the rest of this
# request moot
cid = utils.to_oid(claim)
if cid is None:
return
now = timeutils.utcnow_ts()
message = collection.find_one(query)
if message is None:
return
is_claimed = (message['c']['id'] is not None and
message['c']['e'] > now)
if claim is None:
if is_claimed:
raise exceptions.MessageIsClaimed(message_id)
else:
if message['c']['id'] != cid:
raise exceptions.MessageIsClaimedBy(message_id, claim)
collection.remove(query['_id'], w=0)
def _count(self, queue_name, project=None, include_claimed=False):
"""Return total number of messages in a queue.
This method is designed to very quickly count the number
of messages in a given queue. Expired messages are not
counted, of course. If the queue does not exist, the
count will always be 0.
Note: Some expired messages may be included in the count if
they haven't been GC'd yet. This is done for performance.
"""
query = {
# Messages must belong to this queue
'p_q': utils.scope_queue_name(queue_name, project),
}
if not include_claimed:
# Exclude messages that are claimed
query['c.e'] = {'$lte': timeutils.utcnow_ts()}
collection = self._collection(queue_name, project)
return collection.find(query).hint(COUNTING_INDEX_FIELDS).count()
def pop(self, queue_name, limit, project=None):
query = {
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
}
# Only include messages that are not part of
# any claim, or are part of an expired claim.
now = timeutils.utcnow_ts()
query['c.e'] = {'$lte': now}
collection = self._collection(queue_name, project)
fields = {'_id': 1, 't': 1, 'b': 1}
messages = (collection.find_and_modify(query,
fields=fields,
remove=True)
for _ in range(limit))
messages = itertools.ifilter(None, messages)
final_messages = [_basic_message(message, now) for message in messages]
return final_messages
def list(self, project=None, marker=None,
limit=None, detailed=False):
if limit is None:
limit = self.driver.limits_conf.default_queue_paging
query = {}
scoped_name = utils.scope_queue_name(marker, project)
if not scoped_name.startswith('/'):
# NOTE(kgriffs): scoped queue, e.g., 'project-id/queue-name'
project_prefix = '^' + project + '/'
query['p_q'] = {'$regex': project_prefix, '$gt': scoped_name}
elif scoped_name == '/':
# NOTE(kgriffs): list global queues, but exclude scoped ones
query['p_q'] = {'$regex': '^/'}
else:
# NOTE(kgriffs): unscoped queue, e.g., '/my-global-queue'
query['p_q'] = {'$regex': '^/', '$gt': scoped_name}
fields = {'p_q': 1, '_id': 0}
if detailed:
fields['m'] = 1
cursor = self._collection.find(query, fields=fields)
cursor = cursor.limit(limit).sort('p_q')
marker_name = {}
def normalizer(record):
queue = {'name': utils.descope_queue_name(record['p_q'])}
marker_name['next'] = queue['name']
if detailed:
queue['metadata'] = record['m']
return queue
yield utils.HookedCursor(cursor, normalizer)
yield marker_name and marker_name['next']
def create(self, name, project=None):
# NOTE(flaper87): If the connection fails after it was called
# and we retry to insert the queue, we could end up returning
# `False` because of the `DuplicatedKeyError` although the
# queue was indeed created by this API call.
#
# TODO(kgriffs): Commented out `retries_on_autoreconnect` for
# now due to the above issue, since creating a queue is less
# important to make super HA.
try:
# NOTE(kgriffs): Start counting at 1, and assume the first
# message ever posted will succeed and set t to a UNIX
# "modified at" timestamp.
counter = {'v': 1, 't': 0}
scoped_name = utils.scope_queue_name(name, project)
self._collection.insert({'p_q': scoped_name, 'm': {},
'c': counter})
except pymongo.errors.DuplicateKeyError:
return False
else:
return True
def list(self, project=None, marker=None, limit=None, detailed=False):
if limit is None:
limit = self.driver.limits_conf.default_queue_paging
query = {}
scoped_name = utils.scope_queue_name(marker, project)
if not scoped_name.startswith('/'):
# NOTE(kgriffs): scoped queue, e.g., 'project-id/queue-name'
project_prefix = '^' + project + '/'
query['p_q'] = {'$regex': project_prefix, '$gt': scoped_name}
elif scoped_name == '/':
# NOTE(kgriffs): list global queues, but exclude scoped ones
query['p_q'] = {'$regex': '^/'}
else:
# NOTE(kgriffs): unscoped queue, e.g., '/my-global-queue'
query['p_q'] = {'$regex': '^/', '$gt': scoped_name}
fields = {'p_q': 1, '_id': 0}
if detailed:
fields['m'] = 1
cursor = self._collection.find(query, fields=fields)
cursor = cursor.limit(limit).sort('p_q')
marker_name = {}
def normalizer(record):
queue = {'name': utils.descope_queue_name(record['p_q'])}
marker_name['next'] = queue['name']
if detailed:
queue['metadata'] = record['m']
return queue
yield utils.HookedCursor(cursor, normalizer)
yield marker_name and marker_name['next']
def post(self, queue_name, messages, client_uuid, project=None):
if not self._queue_ctrl.exists(queue_name, project):
raise errors.QueueDoesNotExist(queue_name, project)
now = timeutils.utcnow_ts()
now_dt = datetime.datetime.utcfromtimestamp(now)
collection = self._collection(queue_name, project)
# Set the next basis marker for the first attempt.
next_marker = self._queue_ctrl._get_counter(queue_name, project)
# Unique transaction ID to facilitate atomic batch inserts
transaction = objectid.ObjectId()
prepared_messages = [{
PROJ_QUEUE:
utils.scope_queue_name(queue_name, project),
't':
message['ttl'],
'e':
now_dt + datetime.timedelta(seconds=message['ttl']),
'u':
client_uuid,
'c': {
'id': None,
'e': now
},
'b':
message['body'] if 'body' in message else {},
'k':
next_marker + index,
'tx':
transaction,
} for index, message in enumerate(messages)]
# NOTE(kgriffs): Don't take the time to do a 2-phase insert
# if there is no way for it to partially succeed.
if len(prepared_messages) == 1:
transaction = None
prepared_messages[0]['tx'] = None
# Use a retry range for sanity, although we expect
# to rarely, if ever, reach the maximum number of
# retries.
#
# NOTE(kgriffs): With the default configuration (100 ms
# max sleep, 1000 max attempts), the max stall time
# before the operation is abandoned is 49.95 seconds.
for attempt in self._retry_range:
try:
ids = collection.insert(prepared_messages)
# Log a message if we retried, for debugging perf issues
if attempt != 0:
msgtmpl = _(u'%(attempts)d attempt(s) required to post '
u'%(num_messages)d messages to queue '
u'"%(queue)s" under project %(project)s')
LOG.debug(
msgtmpl,
dict(queue=queue_name,
attempts=attempt + 1,
num_messages=len(ids),
project=project))
# Update the counter in preparation for the next batch
#
# NOTE(kgriffs): Due to the unique index on the messages
# collection, competing inserts will fail as a whole,
# and keep retrying until the counter is incremented
# such that the competing marker's will start at a
# unique number, 1 past the max of the messages just
# inserted above.
self._queue_ctrl._inc_counter(queue_name,
project,
amount=len(ids))
# NOTE(kgriffs): Finalize the insert once we can say that
# all the messages made it. This makes bulk inserts
# atomic, assuming queries filter out any non-finalized
# messages.
if transaction is not None:
collection.update({'tx': transaction},
{'$set': {
'tx': None
}},
upsert=False,
multi=True)
return map(str, ids)
except pymongo.errors.DuplicateKeyError as ex:
# TODO(kgriffs): Record stats of how often retries happen,
# and how many attempts, on average, are required to insert
# messages.
# NOTE(kgriffs): This can be used in conjunction with the
# log line, above, that is emitted after all messages have
# been posted, to gauge how long it is taking for messages
# to be posted to a given queue, or overall.
#
# TODO(kgriffs): Add transaction ID to help match up loglines
if attempt == 0:
msgtmpl = _(u'First attempt failed while '
u'adding messages to queue '
u'"%(queue)s" under project %(project)s')
LOG.debug(msgtmpl, dict(queue=queue_name, project=project))
# NOTE(kgriffs): Never retry past the point that competing
# messages expire and are GC'd, since once they are gone,
# the unique index no longer protects us from getting out
# of order, which could cause an observer to miss this
# message. The code below provides a sanity-check to ensure
# this situation can not happen.
elapsed = timeutils.utcnow_ts() - now
if elapsed > MAX_RETRY_POST_DURATION:
msgtmpl = _(u'Exceeded maximum retry duration for queue '
u'"%(queue)s" under project %(project)s')
LOG.warning(msgtmpl, dict(queue=queue_name,
project=project))
break
# Chill out for a moment to mitigate thrashing/thundering
self._backoff_sleep(attempt)
# NOTE(kgriffs): Perhaps we failed because a worker crashed
# after inserting messages, but before incrementing the
# counter; that would cause all future requests to stall,
# since they would keep getting the same base marker that is
# conflicting with existing messages, until the messages that
# "won" expire, at which time we would end up reusing markers,
# and that could make some messages invisible to an observer
# that is querying with a marker that is large than the ones
# being reused.
#
# To mitigate this, we apply a heuristic to determine whether
# a counter has stalled. We attempt to increment the counter,
# but only if it hasn't been updated for a few seconds, which
# should mean that nobody is left to update it!
#
# Note that we increment one at a time until the logjam is
# broken, since we don't know how many messages were posted
# by the worker before it crashed.
next_marker = self._queue_ctrl._inc_counter(
queue_name, project, window=COUNTER_STALL_WINDOW)
# Retry the entire batch with a new sequence of markers.
#
# NOTE(kgriffs): Due to the unique index, and how
# MongoDB works with batch requests, we will never
# end up with a partially-successful update. The first
# document in the batch will fail to insert, and the
# remainder of the documents will not be attempted.
if next_marker is None:
# NOTE(kgriffs): Usually we will end up here, since
# it should be rare that a counter becomes stalled.
next_marker = self._queue_ctrl._get_counter(
queue_name, project)
else:
msgtmpl = (u'Detected a stalled message counter for '
u'queue "%(queue)s" under project %(project)s. '
u'The counter was incremented to %(value)d.')
LOG.warning(
msgtmpl,
dict(queue=queue_name,
project=project,
value=next_marker))
for index, message in enumerate(prepared_messages):
message['k'] = next_marker + index
except Exception as ex:
LOG.exception(ex)
raise
msgtmpl = _(u'Hit maximum number of attempts (%(max)s) for queue '
u'"%(queue)s" under project %(project)s')
LOG.warning(
msgtmpl,
dict(max=self.driver.mongodb_conf.max_attempts,
queue=queue_name,
project=project))
succeeded_ids = []
raise errors.MessageConflict(queue_name, project, succeeded_ids)
def delete(self, project, queue):
self._col.remove({PRIMARY_KEY: utils.scope_queue_name(queue, project)},
w=0)
def _get_scoped_query(name, project):
return {'p_q': utils.scope_queue_name(name, project)}
def exists(self, project, queue):
key = utils.scope_queue_name(queue, project)
return self._col.find_one({PRIMARY_KEY: key}) is not None
def _list(self, queue_name, project=None, marker=None,
echo=False, client_uuid=None, fields=None,
include_claimed=False, sort=1, limit=None):
"""Message document listing helper.
:param queue_name: Name of the queue to list
:param project: (Default None) Project `queue_name` belongs to. If
not specified, queries the "global" namespace/project.
:param marker: (Default None) Message marker from which to start
iterating. If not specified, starts with the first message
available in the queue.
:param echo: (Default False) Whether to return messages that match
client_uuid
:param client_uuid: (Default None) UUID for the client that
originated this request
:param fields: (Default None) Fields to include in emitted
documents
:param include_claimed: (Default False) Whether to include
claimed messages, not just active ones
:param sort: (Default 1) Sort order for the listing. Pass 1 for
ascending (oldest message first), or -1 for descending (newest
message first).
:param limit: (Default None) The maximum number of messages
to list. The results may include fewer messages than the
requested `limit` if not enough are available. If limit is
not specified
:returns: Generator yielding up to `limit` messages.
"""
if sort not in (1, -1):
raise ValueError(u'sort must be either 1 (ascending) '
u'or -1 (descending)')
now = timeutils.utcnow_ts()
query = {
# Messages must belong to this queue and project.
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
# NOTE(kgriffs): Messages must be finalized (i.e., must not
# be part of an unfinalized transaction).
#
# See also the note wrt 'tx' within the definition
# of ACTIVE_INDEX_FIELDS.
'tx': None,
}
if not echo:
query['u'] = {'$ne': client_uuid}
if marker is not None:
query['k'] = {'$gt': marker}
collection = self._collection(queue_name, project)
if not include_claimed:
# Only include messages that are not part of
# any claim, or are part of an expired claim.
query['c.e'] = {'$lte': now}
# Construct the request
cursor = collection.find(query, fields=fields, sort=[('k', sort)])
if limit is not None:
cursor.limit(limit)
# NOTE(flaper87): Suggest the index to use for this query to
# ensure the most performant one is chosen.
return cursor.hint(ACTIVE_INDEX_FIELDS)
def _list(self,
queue_name,
project=None,
marker=None,
echo=False,
client_uuid=None,
fields=None,
include_claimed=False,
sort=1,
limit=None):
"""Message document listing helper.
:param queue_name: Name of the queue to list
:param project: (Default None) Project `queue_name` belongs to. If
not specified, queries the "global" namespace/project.
:param marker: (Default None) Message marker from which to start
iterating. If not specified, starts with the first message
available in the queue.
:param echo: (Default False) Whether to return messages that match
client_uuid
:param client_uuid: (Default None) UUID for the client that
originated this request
:param fields: (Default None) Fields to include in emitted
documents
:param include_claimed: (Default False) Whether to include
claimed messages, not just active ones
:param sort: (Default 1) Sort order for the listing. Pass 1 for
ascending (oldest message first), or -1 for descending (newest
message first).
:param limit: (Default None) The maximum number of messages
to list. The results may include fewer messages than the
requested `limit` if not enough are available. If limit is
not specified
:returns: Generator yielding up to `limit` messages.
"""
if sort not in (1, -1):
raise ValueError(u'sort must be either 1 (ascending) '
u'or -1 (descending)')
now = timeutils.utcnow_ts()
query = {
# Messages must belong to this queue and project.
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
# NOTE(kgriffs): Messages must be finalized (i.e., must not
# be part of an unfinalized transaction).
#
# See also the note wrt 'tx' within the definition
# of ACTIVE_INDEX_FIELDS.
'tx': None,
}
if not echo:
query['u'] = {'$ne': client_uuid}
if marker is not None:
query['k'] = {'$gt': marker}
collection = self._collection(queue_name, project)
if not include_claimed:
# Only include messages that are not part of
# any claim, or are part of an expired claim.
query['c.e'] = {'$lte': now}
# Construct the request
cursor = collection.find(query,
fields=fields,
sort=[('k', sort)],
limit=limit)
# NOTE(flaper87): Suggest the index to use for this query to
# ensure the most performant one is chosen.
return cursor.hint(ACTIVE_INDEX_FIELDS)
def _insert(self, project, queue, shard, upsert):
key = utils.scope_queue_name(queue, project)
return self._col.update({PRIMARY_KEY: key}, {'$set': {
's': shard
}},
upsert=upsert)
def _get_scoped_query(name, project):
return {'p_q': utils.scope_queue_name(name, project)}
def _insert(self, project, queue, pool, upsert):
key = utils.scope_queue_name(queue, project)
return self._col.update({PRIMARY_KEY: key},
{'$set': {'s': pool}}, upsert=upsert)
def create(self, queue, metadata, project=None,
limit=storage.DEFAULT_MESSAGES_PER_CLAIM):
"""Creates a claim.
This implementation was done in a best-effort fashion.
In order to create a claim we need to get a list
of messages that can be claimed. Once we have that
list we execute a query filtering by the ids returned
by the previous query.
Since there's a lot of space for race conditions here,
we'll check if the number of updated records is equal to
the max number of messages to claim. If the number of updated
messages is lower than limit we'll try to claim the remaining
number of messages.
This 2 queries are required because there's no way, as for the
time being, to execute an update on a limited number of records.
"""
msg_ctrl = self.driver.message_controller
ttl = metadata['ttl']
grace = metadata['grace']
oid = objectid.ObjectId()
now = timeutils.utcnow_ts()
claim_expires = now + ttl
message_ttl = ttl + grace
message_expiration = datetime.datetime.utcfromtimestamp(
claim_expires + grace)
meta = {
'id': oid,
't': ttl,
'e': claim_expires,
}
# Get a list of active, not claimed nor expired
# messages that could be claimed.
msgs = msg_ctrl._active(queue, fields={'_id': 1}, project=project,
limit=limit)
messages = iter([])
ids = [msg['_id'] for msg in msgs]
if len(ids) == 0:
return (None, messages)
now = timeutils.utcnow_ts()
# NOTE(kgriffs): Set the claim field for
# the active message batch, while also
# filtering out any messages that happened
# to get claimed just now by one or more
# parallel requests.
#
# Filtering by just 'c.e' works because
# new messages have that field initialized
# to the current time when the message is
# posted. There is no need to check whether
# 'c' exists or 'c.id' is None.
collection = msg_ctrl._collection(queue, project)
updated = collection.update({'_id': {'$in': ids},
'c.e': {'$lte': now}},
{'$set': {'c': meta}},
upsert=False,
multi=True)['n']
# NOTE(flaper87): Dirty hack!
# This sets the expiration time to
# `expires` on messages that would
# expire before claim.
new_values = {'e': message_expiration, 't': message_ttl}
collection.update({'p_q': utils.scope_queue_name(queue, project),
'e': {'$lt': message_expiration},
'c.id': oid},
{'$set': new_values},
upsert=False, multi=True)
if updated != 0:
# NOTE(kgriffs): This extra step is necessary because
# in between having gotten a list of active messages
# and updating them, some of them may have been
# claimed by a parallel request. Therefore, we need
# to find out which messages were actually tagged
# with the claim ID successfully.
claim, messages = self.get(queue, oid, project=project)
return (str(oid), messages)
def post(self, queue_name, messages, client_uuid, project=None):
if not self._queue_ctrl.exists(queue_name, project):
raise errors.QueueDoesNotExist(queue_name, project)
now = timeutils.utcnow_ts()
now_dt = datetime.datetime.utcfromtimestamp(now)
collection = self._collection(queue_name, project)
# Set the next basis marker for the first attempt.
next_marker = self._queue_ctrl._get_counter(queue_name, project)
prepared_messages = [
{
"t": message["ttl"],
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
"e": now_dt + datetime.timedelta(seconds=message["ttl"]),
"u": client_uuid,
"c": {"id": None, "e": now},
"b": message["body"] if "body" in message else {},
"k": next_marker + index,
}
for index, message in enumerate(messages)
]
# Use a retry range for sanity, although we expect
# to rarely, if ever, reach the maximum number of
# retries.
#
# NOTE(kgriffs): With the default configuration (100 ms
# max sleep, 1000 max attempts), the max stall time
# before the operation is abandoned is 49.95 seconds.
for attempt in self._retry_range:
try:
ids = collection.insert(prepared_messages)
# Log a message if we retried, for debugging perf issues
if attempt != 0:
msgtmpl = _(
u"%(attempts)d attempt(s) required to post "
u"%(num_messages)d messages to queue "
u'"%(queue)s" under project %(project)s'
)
LOG.debug(
msgtmpl, dict(queue=queue_name, attempts=attempt + 1, num_messages=len(ids), project=project)
)
# Update the counter in preparation for the next batch
#
# NOTE(kgriffs): Due to the unique index on the messages
# collection, competing inserts will fail as a whole,
# and keep retrying until the counter is incremented
# such that the competing marker's will start at a
# unique number, 1 past the max of the messages just
# inserted above.
self._queue_ctrl._inc_counter(queue_name, project, amount=len(ids))
return map(str, ids)
except pymongo.errors.DuplicateKeyError as ex:
# Try again with the remaining messages
# TODO(kgriffs): Record stats of how often retries happen,
# and how many attempts, on average, are required to insert
# messages.
# NOTE(kgriffs): This can be used in conjunction with the
# log line, above, that is emitted after all messages have
# been posted, to guage how long it is taking for messages
# to be posted to a given queue, or overall.
#
# TODO(kgriffs): Add transaction ID to help match up loglines
if attempt == 0:
msgtmpl = _(
u"First attempt failed while "
u"adding messages to queue "
u'"%(queue)s" under project %(project)s'
)
LOG.debug(msgtmpl, dict(queue=queue_name, project=project))
# NOTE(kgriffs): Never retry past the point that competing
# messages expire and are GC'd, since once they are gone,
# the unique index no longer protects us from getting out
# of order, which could cause an observer to miss this
# message. The code below provides a sanity-check to ensure
# this situation can not happen.
elapsed = timeutils.utcnow_ts() - now
if elapsed > MAX_RETRY_POST_DURATION:
msgtmpl = _(u"Exceeded maximum retry duration for queue " u'"%(queue)s" under project %(project)s')
LOG.warning(msgtmpl, dict(queue=queue_name, project=project))
break
# Chill out for a moment to mitigate thrashing/thundering
self._backoff_sleep(attempt)
# NOTE(kgriffs): Perhaps we failed because a worker crashed
# after inserting messages, but before incrementing the
# counter; that would cause all future requests to stall,
# since they would keep getting the same base marker that is
# conflicting with existing messages, until the messages that
# "won" expire, at which time we would end up reusing markers,
# and that could make some messages invisible to an observer
# that is querying with a marker that is large than the ones
# being reused.
#
# To mitigate this, we apply a heuristic to determine whether
# a counter has stalled. We attempt to increment the counter,
# but only if it hasn't been updated for a few seconds, which
# should mean that nobody is left to update it!
#
# Note that we increment one at a time until the logjam is
# broken, since we don't know how many messages were posted
# by the worker before it crashed.
next_marker = self._queue_ctrl._inc_counter(queue_name, project, window=COUNTER_STALL_WINDOW)
# Retry the entire batch with a new sequence of markers.
#
# NOTE(kgriffs): Due to the unique index, and how
# MongoDB works with batch requests, we will never
# end up with a partially-successful update. The first
# document in the batch will fail to insert, and the
# remainder of the documents will not be attempted.
if next_marker is None:
# NOTE(kgriffs): Usually we will end up here, since
# it should be rare that a counter becomes stalled.
next_marker = self._queue_ctrl._get_counter(queue_name, project)
else:
msgtmpl = (
u"Detected a stalled message counter for "
u'queue "%(queue)s" under project %(project)s. '
u"The counter was incremented to %(value)d."
)
LOG.warning(msgtmpl, dict(queue=queue_name, project=project, value=next_marker))
for index, message in enumerate(prepared_messages):
message["k"] = next_marker + index
except Exception as ex:
# TODO(kgriffs): Query the DB to get the last marker that
# made it, and extrapolate from there to figure out what
# needs to be retried.
LOG.exception(ex)
raise
msgtmpl = _(u"Hit maximum number of attempts (%(max)s) for queue " u'"%(queue)s" under project %(project)s')
LOG.warning(msgtmpl, dict(max=self.driver.mongodb_conf.max_attempts, queue=queue_name, project=project))
succeeded_ids = []
raise errors.MessageConflict(queue_name, project, succeeded_ids)
def exists(self, project, queue):
key = utils.scope_queue_name(queue, project)
return self._col.find_one({PRIMARY_KEY: key}) is not None
def delete(self, project, queue):
self._col.remove({PRIMARY_KEY: utils.scope_queue_name(queue, project)},
w=0)
def post(self, queue_name, messages, client_uuid, project=None):
# NOTE(flaper87): This method should be safe to retry on
# autoreconnect, since we've a 2-step insert for messages.
# The worst-case scenario is that we'll increase the counter
# several times and we'd end up with some non-active messages.
if not self._queue_ctrl.exists(queue_name, project):
raise errors.QueueDoesNotExist(queue_name, project)
now = timeutils.utcnow_ts()
now_dt = datetime.datetime.utcfromtimestamp(now)
collection = self._collection(queue_name, project)
# Set the next basis marker for the first attempt.
next_marker = self._queue_ctrl._get_counter(queue_name, project)
# Unique transaction ID to facilitate atomic batch inserts
transaction = objectid.ObjectId()
prepared_messages = [
{
PROJ_QUEUE: utils.scope_queue_name(queue_name, project),
't': message['ttl'],
'e': now_dt + datetime.timedelta(seconds=message['ttl']),
'u': client_uuid,
'c': {'id': None, 'e': now},
'b': message['body'] if 'body' in message else {},
'k': next_marker + index,
'tx': transaction,
}
for index, message in enumerate(messages)
]
# NOTE(kgriffs): Don't take the time to do a 2-phase insert
# if there is no way for it to partially succeed.
if len(prepared_messages) == 1:
transaction = None
prepared_messages[0]['tx'] = None
# Use a retry range for sanity, although we expect
# to rarely, if ever, reach the maximum number of
# retries.
#
# NOTE(kgriffs): With the default configuration (100 ms
# max sleep, 1000 max attempts), the max stall time
# before the operation is abandoned is 49.95 seconds.
for attempt in self._retry_range:
try:
ids = collection.insert(prepared_messages)
# Log a message if we retried, for debugging perf issues
if attempt != 0:
msgtmpl = _(u'%(attempts)d attempt(s) required to post '
u'%(num_messages)d messages to queue '
u'"%(queue)s" under project %(project)s')
LOG.debug(msgtmpl,
dict(queue=queue_name,
attempts=attempt + 1,
num_messages=len(ids),
project=project))
# Update the counter in preparation for the next batch
#
# NOTE(kgriffs): Due to the unique index on the messages
# collection, competing inserts will fail as a whole,
# and keep retrying until the counter is incremented
# such that the competing marker's will start at a
# unique number, 1 past the max of the messages just
# inserted above.
self._queue_ctrl._inc_counter(queue_name, project,
amount=len(ids))
# NOTE(kgriffs): Finalize the insert once we can say that
# all the messages made it. This makes bulk inserts
# atomic, assuming queries filter out any non-finalized
# messages.
if transaction is not None:
collection.update({'tx': transaction},
{'$set': {'tx': None}},
upsert=False, multi=True)
return map(str, ids)
except pymongo.errors.DuplicateKeyError as ex:
# TODO(kgriffs): Record stats of how often retries happen,
# and how many attempts, on average, are required to insert
# messages.
# NOTE(kgriffs): This can be used in conjunction with the
# log line, above, that is emitted after all messages have
# been posted, to gauge how long it is taking for messages
# to be posted to a given queue, or overall.
#
# TODO(kgriffs): Add transaction ID to help match up loglines
if attempt == 0:
msgtmpl = _(u'First attempt failed while '
u'adding messages to queue '
u'"%(queue)s" under project %(project)s')
LOG.debug(msgtmpl, dict(queue=queue_name, project=project))
# NOTE(kgriffs): Never retry past the point that competing
# messages expire and are GC'd, since once they are gone,
# the unique index no longer protects us from getting out
# of order, which could cause an observer to miss this
# message. The code below provides a sanity-check to ensure
# this situation can not happen.
elapsed = timeutils.utcnow_ts() - now
if elapsed > MAX_RETRY_POST_DURATION:
msgtmpl = _(u'Exceeded maximum retry duration for queue '
u'"%(queue)s" under project %(project)s')
LOG.warning(msgtmpl,
dict(queue=queue_name, project=project))
break
# Chill out for a moment to mitigate thrashing/thundering
self._backoff_sleep(attempt)
# NOTE(kgriffs): Perhaps we failed because a worker crashed
# after inserting messages, but before incrementing the
# counter; that would cause all future requests to stall,
# since they would keep getting the same base marker that is
# conflicting with existing messages, until the messages that
# "won" expire, at which time we would end up reusing markers,
# and that could make some messages invisible to an observer
# that is querying with a marker that is large than the ones
# being reused.
#
# To mitigate this, we apply a heuristic to determine whether
# a counter has stalled. We attempt to increment the counter,
# but only if it hasn't been updated for a few seconds, which
# should mean that nobody is left to update it!
#
# Note that we increment one at a time until the logjam is
# broken, since we don't know how many messages were posted
# by the worker before it crashed.
next_marker = self._queue_ctrl._inc_counter(
queue_name, project, window=COUNTER_STALL_WINDOW)
# Retry the entire batch with a new sequence of markers.
#
# NOTE(kgriffs): Due to the unique index, and how
# MongoDB works with batch requests, we will never
# end up with a partially-successful update. The first
# document in the batch will fail to insert, and the
# remainder of the documents will not be attempted.
if next_marker is None:
# NOTE(kgriffs): Usually we will end up here, since
# it should be rare that a counter becomes stalled.
next_marker = self._queue_ctrl._get_counter(
queue_name, project)
else:
msgtmpl = (u'Detected a stalled message counter for '
u'queue "%(queue)s" under project %(project)s. '
u'The counter was incremented to %(value)d.')
LOG.warning(msgtmpl,
dict(queue=queue_name,
project=project,
value=next_marker))
for index, message in enumerate(prepared_messages):
message['k'] = next_marker + index
except Exception as ex:
LOG.exception(ex)
raise
msgtmpl = _(u'Hit maximum number of attempts (%(max)s) for queue '
u'"%(queue)s" under project %(project)s')
LOG.warning(msgtmpl,
dict(max=self.driver.mongodb_conf.max_attempts,
queue=queue_name,
project=project))
succeeded_ids = []
raise errors.MessageConflict(queue_name, project,
succeeded_ids)