Google Cloud Platform Blog: Efficient Use of Asynchronous Operations in Google App Engine

Google Cloud Platform Blog

Product updates, customer stories, and tips and tricks on Google Cloud Platform

Efficient Use of Asynchronous Operations in Google App Engine

Thursday, October 15, 2015

Today's guest post comes from Jason Vertrees, Ph.D., CTO of RealMassive, and David Lippa, Principal Software Engineer. RealMassive provides real-time commercial real estate information and is powered by Google App Engine. taskletdeferred taskyieldduring¹Only 5 transactional tasks before the transaction ends, and not necessarily immediately

All entities are tagged with a NDB key to the User who performed an operation on that entity

A Space represents a suite in a Building. It’s located on a floor of the Building, and has a size in square feet. It may also have attachments, like images and videos

A Building represents a physical Building, with its own attributes (such as total square footage, owner, etc.) and a street address.

Spaces and Buildings have mutual references to one another.

from google.appengine.ext import ndb, deferred

class User(ndb.Model):
username = ndb.StringProperty()
email = ndb.StringProperty()

class BaseModel(ndb.Model):
created_by = ndb.KeyProperty(User)
edited_by = ndb.KeyProperty(User, repeated=True)

class Media(BaseModel):
url = ndb.StringProperty()
edited_by = ndb.KeyProperty(User, repeated=True)

class Address(ndb.StructuredProperty):
street = ndb.StringProperty()
city = ndb.StringProperty()
state = ndb.StringProperty()
zipcode = ndb.StringProperty()

class Building(BaseModel):
address = ndb.StructuredProperty(Address)
size = ndb.IntegerProperty()
# prevent cyclical dependencies
space_keys = ndb.KeyProperty(kind=’Space’, repeated=True)
owner = ndb.KeyProperty(User)

def remove(self):
for s in space_keys:
s.get().remove()
self.key.delete()

class Space(BaseModel):
building = ndb.KeyProperty(Building)
sf_available = ndb.IntegerProperty()
floor_number = ndb.IntegerProperty()
suite_number = ndb.IntegerProperty()
attachments = ndb.KeyProperty(Media, repeated=True)

def remove(self):
for a in attachments:
a.delete()

For a Building with many Spaces, the removal process could cause a DeadlineExceededError to occur, since the processes are performed in the same thread that is executing the HTTP request.

The removals are not transactional: if a single operation fails, there can be orphans.

Media can be orphaned as well, since there’s no reference counting with the Media

remove()ndb.transactionalclass Building(BaseModel):
address = ndb.StructuredProperty(Address)
size = ndb.IntegerProperty()
# prevent cyclical dependencies
space_keys = ndb.KeyProperty(kind=’Space’, repeated=True)
owner = ndb.KeyProperty(User)

@ndb.transactional(xg=True)
def remove(self):
for s in space_keys:
deferred.defer(s.get().remove)
self.key.delete()

class Space(BaseModel):
building = ndb.KeyProperty(Building)
sf_available = ndb.IntegerProperty()
floor_number = ndb.IntegerProperty()
suite_number = ndb.IntegerProperty()
attachments = ndb.KeyProperty(Media, repeated=True)

@ndb.transactional(xg=True)
def remove(self):
for a in attachments:
deferred.defer(a.delete)

If a queue supports simultaneous operations, there’s no way to guarantee deletion order unless the queue rate is set to 1 per second.

Building.remove() could exceed deadlines since it calls get() before delegating to Space.remove().

The semantics of this code is very confusing, since it appears that everything is done within the same transaction.

It won’t scale. If a Building has more than 5 Spaces, this could exceed the limit on the number of deferred tasks within a transaction. Also, cross-group transactions are limited to 30 seconds of clock time.

class Building(BaseModel):
address = ndb.StructuredProperty(Address)
size = ndb.IntegerProperty()
# prevent cyclical dependencies
space_keys = ndb.KeyProperty(kind=’Space’, repeated=True)
owner = ndb.KeyProperty(User)

@ndb.transactional(xg=True)
@ndb.tasklet
def remove(self):
futures = []
for s in space_keys:
sp = yield s.get_async()
futures.extend(sp.remove_async())
ndb.wait_all(futures)
self.key.delete()

class Space(BaseModel):
building = ndb.KeyProperty(Building)
sf_available = ndb.IntegerProperty()
floor_number = ndb.IntegerProperty()
suite_number = ndb.IntegerProperty()
attachments = ndb.KeyProperty(Media, repeated=True)

@ndb.transactional(xg=True)
def remove_async(self):
return ndb.delete_multi_async(self.attachments)

def remove(self):
ndb.Future.wait_all(self.remove_async())
1 We are not going to dive into the differences or details of task queues. For more information, please see the citation.