Strange the article proposes itself for "Enterprise" yet has no mention of Google's Zanzibar and how it compares to the other approaches. AFAIK it doesn't use pre-computed values but just queries really fast (using Spanner so there's that)
Google's Zanzibar actually does both: for the vast majority of queries, it uses significant levels of caching and a permitted amount of staleness [1], allowing Spanner to return a (somewhat stale) copy of the relationship data from local nodes, rather than having to wait or coordinate with the other nodes.
However, some deeply recursive or wide relations can still be slow, so Zanzibar also has a pre-computation cache called Leopard that is used for a very specific subset of these relations [2]. For SpiceDB, we called our version of this cache Materialize and it is designed expressly for handling "Enterprise" levels of scale in a similar fashion, as sometimes it is simply too slow to walk these deep graphs in real-time.
In SpiceDB, this is known as the LookupResources [1] API, which returns all resources (of a particular type) that a particular subject (user in this case) has a particular permission on.
We have a guide on doing ACL-aware filtering and listing [2] with this API and describing other approaches for larger Enterprise scales
Disclaimer: I'm the co-founder and CTO of AuthZed, we develop SpiceDB, and I wrote our most recent implementation of LookupResources
We actually have users that synchronize their resources from various sources (AWS, Kubernetes, etc) into SpiceDB, explicitly so they can perform these kinds of queries!
One of the major benefits of a centralized authorization system is allowing for permissions queries across resources and subjects from multiple different services/sources (of course, with the need to synchronize the data in)
Happy to expand on how some users do so, if you're curious.
>We added a point of failure, as the permissions table can get out of sync with the actual data.
>The main risk with pre-computed permissions is data getting out of sync.
It would make sense to have permissions be a first class concept for databases and to ensure such a desync could never happen. Data being only read or written from specific users is a very common thing for data so it would be worth having first class support for it.
I'm struggling to understand what the issue that the author is getting at. The point of a database is that it's ACID compliant, wrap insets/updates/deletes in a transaction and no such drift would occur. What am I missing?
I don't think you are missing anything. I think he is just pointing out that technically nothing is enforcing this synchronization, so if someone forgets to wrap things in a transaction, it could get out of sync.
Depending on your DBMS and isolation level, using a transaction might not fix things. That being said I don’t think (at least for Postgres) most people are using an isolation level that could cause this.
Much more likely I think is that you can’t use the db to prevent invalid states here (unique constraint, etc) and you’re dependent on other areas of the code correctly implementing concurrency controls. Race condition in resource A causes problems in your permissions table now.
And just from a general engineering perspective, you should assume things are going to fail and assess what your path forward looks like when they do. Recovery script sounds like a good idea for a critical area.
I just want to point out you have to take care about that, yes you can have a trigger or a transaction to make sure it happens but it isn't there out of the box
Why is it a useful property that everything is always "in sync"? I propose this is not possible anyway. These systems are always asynchronous, and the time of check is always before the time of use, and it is always possible that a revocation occurs between them, and this problem cannot be eliminated.
Another approach to complex requirements without spending a lot of time querying databases is to use bitmaps. A set of permissions can be expressed through a bitmap and all you need to do in code is to "decode" that to what you actually let the user do.
The downside to this approach is that it requires some planning and to maintain in code what mask retrieves what permission(s).
I only did a quick read of permit.io offering but iirc they don't focus on hierarchical data. If having access to a resource cannot grant access to unbounded number of other independent resources (eg sharing a folder) then almost all issues of the article disappear
However, some deeply recursive or wide relations can still be slow, so Zanzibar also has a pre-computation cache called Leopard that is used for a very specific subset of these relations [2]. For SpiceDB, we called our version of this cache Materialize and it is designed expressly for handling "Enterprise" levels of scale in a similar fashion, as sometimes it is simply too slow to walk these deep graphs in real-time.
[1]: https://zanzibar.tech/24uQOiQnVi:1T:4S [2]: https://zanzibar.tech/21tieegnDR:0.H1AowI3SG:2O
We have a guide on doing ACL-aware filtering and listing [2] with this API and describing other approaches for larger Enterprise scales
Disclaimer: I'm the co-founder and CTO of AuthZed, we develop SpiceDB, and I wrote our most recent implementation of LookupResources
[1]: https://buf.build/authzed/api/docs/main:authzed.api.v1#authz... [2]: https://authzed.com/docs/spicedb/modeling/protecting-a-list-...
One of the major benefits of a centralized authorization system is allowing for permissions queries across resources and subjects from multiple different services/sources (of course, with the need to synchronize the data in)
Happy to expand on how some users do so, if you're curious.
Fine-grained authorization as an incremental computation problem
>The main risk with pre-computed permissions is data getting out of sync.
It would make sense to have permissions be a first class concept for databases and to ensure such a desync could never happen. Data being only read or written from specific users is a very common thing for data so it would be worth having first class support for it.
Much more likely I think is that you can’t use the db to prevent invalid states here (unique constraint, etc) and you’re dependent on other areas of the code correctly implementing concurrency controls. Race condition in resource A causes problems in your permissions table now.
And just from a general engineering perspective, you should assume things are going to fail and assess what your path forward looks like when they do. Recovery script sounds like a good idea for a critical area.
The downside to this approach is that it requires some planning and to maintain in code what mask retrieves what permission(s).
Scales both on the tech, and on the human side - e.g. your product manager can add roles (with CI approval) without requiring engineering involvement.
(I'm biased but still true)