Any recommended strategies for cache invalidation

[michaelbukachi]

This question might not be specific to this library but just a general question. Are there any recommended approaches for cache invalidation when relationships (related data) are involved?
E.g

class User(Base):
     id = Column(...)
    name = Column(...)

    roles = relationship('Role', ....)

class Role(Base)
     id = Column(....)
    name = Column(....)
    user_id = Column(....)

How would we go about invalidating a user when we update one of their roles? I'm able to invalidate all users when a role is updated, however, I don't think that is ideal. I'm using the official example as a reference.

Note: Assume no lazyloading is being used i.e only eagerloading

[zzzeek]

cc @jvanasco , IIRC he has a lot of opinions on this

[michaelbukachi]

@jvanasco please share your opinions 😁. Any videos, articles etc?

[jvanasco]

There's no universal solution to caching and invalidation. Getting it right on one project is wrong for another.

In very general terms, you want to have a good understanding of what your cached data is, how it is structured, and the access patterns. When it comes to User/Role situations like your example, I generally view that association as (i) Very High Read, (ii) Low Write, and (iii) Strongly Partitioned from the User and Role records in the sense that I don't need any of the record's other information - just the existence or lack-of the association. In the context of a cloud based web application, my goal is to load the user's role information on every page request or action. With a scenario like that, I would likely have a dogpile "region" named/prefixed user2role with the key as the UserID and the value as a list of the role ids. That would allow me to efficiently store and query the association data, and avoid the overhead of storing/transferring user profile or role information.

When it comes to invalidation, I like to create triggers and policies. If I edit a User's SQLAlchemy model (in the database), I will either catch that with a SQLAlchemy Event to automatically trigger a policy, or I will explicitly invoke a a policy. The policy is just a small routine that says "If the edited object is a user, invalidate these cache items; if they edited object is a user and a role, invalidate these cache items as well".

[michaelbukachi]

@jvanasco You are totally right. There is truly no universal solution 😱. I think I figured out an implementation that works though. I was struggling with coming with the right policies i.e when this obj changes, invalidate this obj and that obj as well. Thanks for the insights though, I had totally overlooked "partitioning".