Updates (UpdateActor in particular) are not atomic

[thockin]

The logic today is:

1. RPC layer UpdateActor(newActor)
2. Get current value from DB
3. Check UID precondition
4. Use current version as expected version if not specified by user
5. Apply changes to in-memory state
6. Call storage layer UpdateActor(newActor)
7. Open txn
8. Get current value from DB
9. Check expected version == current version
10. Commit

This is ...pretty broken?

Between steps 3 and 7 the DB may have been updated and we are now operating on a totally different object.

We do Get twice.

I think we need to let the storage layer be atomic. To my eyes that looks something like:

1. RPC layer UpdateActor(newActor)
2. Create a callback
3. Call storage layer UpdateActor(newActor, callback)
4. Open txn
5. Get current value from DB
6. Call callback(new, old)
   6a) Check UID precondition IFF it was specified
   6b) Check version IFF it was specified
   6c) Apply changes to in-memory state
7. Commit

Am I misunderstanding something?

Julian Gutierrez Oschmann (@juli4n) Luiz Oliveira (@laoj2) Eitan Yarmush (@EItanya)

I am trying to do declarative validation and demonstrate what a "canonically correct" update operation looks like, and this set off all my alarms.

[juli4n]

Yee, this is tracked in #763, see #742 (comment)