CODEWITHSUNDEEP
pythongoogle-app-engineeventual-consistency
Peter Leontev
Peter Leontev

Reputation: 107

AppEngine achieving strong consistency

I am trying to achieve strong consistency. Let's call my model PVPPlayer:

class PVPPlayer(ndb.Model):
    points = ndb.IntegerProperty()

Every key for the model is created like this:

pvp_player = PVPPlayer(key=ndb.Key(Profile, "test_id", PVPPlayer, "test_id"))

where Profile is parent model:

class Profile(ndb.Model):
    def build_key(cls, some_id):
        return ndb.Key(cls, some_id)

I have 2 REST api url:

1) update_points
2) get_points

In 1) I do :

# I use transaction because I have to update all the models in single batch 
@ndb.transactional(xg=True, retries=3)
def some_func(points):
    pvp_player = ndb.Key(Profile, "test_id", PVPPlayer, "test_id").get()
    pvp_player.points += points 
    pvp_player.put()
    # update other models here`

In 2) I do:

pvp_player = ndb.Key(Profile, "test_id", PVPPlayer, "test_id").get()
return pvp_player.points`

My flow looks like this:

1) update_points()
2) get_points()
3) update_points()
4) get_points()`
...

Problem:

Using get() guarantees strong consistency so what I don't understand is why sometimes as the result of get_points() I get stale data like points were not updated at all.

Example:

POST get_points -> 0
POST sleep 1-3 sec
POST update_points -> 15
POST sleep 1-3 sec
POST get_points -> 15
POST sleep 1-3 sec
POST update_points -> 20
POST sleep 1-3 sec
POST get_points -> 15 !!!`

Upvotes: 0

Views: 129

Answers (2)

glmvrml
glmvrml

Reputation: 1632

First check your logs, one of the updates must be failing with error, because your logic is basically correct.

Also double-check all your updates are wrapped in transactions to avoid races. Cloud Datastore: ways to avoid race conditions

This case is likely not about consistency issues, but stomping updates, checkout this links for some interesting cases:

http://engineering.khanacademy.org/posts/transaction-safety.htm http://engineering.khanacademy.org/posts/user-write-lock.htm

Upvotes: 2

Konstantinos G
Konstantinos G

Reputation: 46

Is there a case you exceed the write limit per entity group, that is one update per second? I think this could break the strong consistency of the entity group as mentioned in the documentation.

Upvotes: 1

Related Questions