peers/forgefed

+++ title = "Actor Factory" date = 2024-08-11 [extra] author = "Pere Lev" +++

Until now, publishing a new resource, whether it's a comment or a project, has been happening only locally, on the user's home server. It means that whoever wanted to implement a ForgeFed-federated forge, would need to host the implementations of all actor/resource/service types on one domain name, acting as a big all-in-one forge.

This is no longer the case! And that's what this post is about. Actor creation is obviously a part of the core of ForgeFed's API, so I decided it's time to take care of this upgrade :)

As always, my task board is available.

Federated Actions

Every request sent on the federated network is an interaction between two or more actors. Some of these requests can occur between actors of different servers, and some can't.

Examples of federation-supporting requests:

Follow an (possibly remote) actor
Open an issue (possibly on a remote issue tracker)
Add a repository to a project (both of which might be on different servers)

Examples of requests that don't support federation, or weren't until now:

Create a user account (this happens between your client/browser and your home instance of choice)
Editing admin settings of other instances
Creating a project

Until now, creating a new actor (project, ticket tracker, etc.) was limited to your home instance, i.e. the actor is launched on the same instance where your account is.

On one hand, it makes sense: Why would other servers allow you to create projects on them?

On the other hand:

It's an artificial limitation, i.e. this action could easily support federation
It blocks the possibility of implementing micro-services that live on different domains, where each service specializes in a specific type(s) of actor
It blocks the possibility of distributed teams to allow all their members to create resources on the team's main instance

Factory Actor

So, shall we make actor creation support federation? That's exactly what I did!

There's a new type of actor (which will find its way to the ForgeFed specification very soon as well), called Factory. The only job of a Factory is to create resource actors (everything except Person and Factory) on its home instance.

With that piece in place, Create activities are no longer sent into "thin air", they're sent to a Factory. And the Factory handles creating the actor. Person actors can no longer launch new projects by themselves.

Since a Factory it a resource actor, it can have its own collaborators and teams that have access to it, and it implements the Object Capability system, just like the other actor types. This means a Factory can have remote collaborators, and remote teams, thus allowing remote users to create new actors via the Factory.

But who factories the factory?

So, Factory is the only actor type that can be created in the "old" way, i.e. without requiring a Factory. The client/browser sends a client-to-server (C2S) Create activity to the Person's outbox, and the Person actor launches a new Factory. However, we wouldn't want any user to be able to do this. So, in the Vervis settings file, there's a list of users who are approved for creating new Factory actors.

Summary of actor creation:

Person: Created via the registration API (see file API.md in the Vervis source repo for details)
Factory: Created via a C2S request to a Person actor, and limited to users who are allowed to create factories (normally, the server admins)
Any other ForgeFed actor type: Created via a Factory

The Little Details

I needed to figure out certain details before the implementation:

How to set which actor types the Factory is able to create?
How to control who is allowed to create Factories, who is allowed to use them, who is allowed to change their settings and delete them?
How to make a Factory automatically serve every new user on another instance? e.g. so that people on the Person micro-service can create repos on the Repository micro-service

These details might get changes and updates, but right now here's how they work:

There's a new property availableActorTypes which specifies, for a given Factory, which types it's allowed to create
That property is specified when Creating the Factory, and when editing it
Editing uses a new activity type Patch where object is the object being edited and the other fields are object fields being set to a new value
The name Patch is already use to represent a code patch, but the term Patch works for both cases because they appear in different situations
I'd use Edit etc. but Patch might become standard name on the Fediverse, coming from the PATCH method of HTTP
Who can create Factories isn't controlled by the OCAP system, but instead using a setting (but it could also be in the DB) that lists admin usernames
To use a Factory to create actors, you need to have a valid Grant for it, with the role being at least write (a.k.a developer)
A factory can have both Person collaborators and Teams that have access to it, so if a Person micro-service creates a Team containing all users, and that Team is given access to the remote Repository-Factory, users can now creates repos on the Repository micro-service

Bonus: Synchronized Push

In web apps there's often a single big database, and concurrent access to it is managed using atomic isolated transactions. Actor programming brings a new approach: Each actor has its own database, its own data. Since actors handle messages one by one, there's no concurrent access, and transactions happen on the actor level rather than database level.

But that beautiful model works only when the actor has exclusive access to its data and database! And that's exactly a problem Vervis had until now:

When a human asks to merge a PR by sending an activity to the Repository actor's inbox, the Repository actor handles the merge
When a human pushes commits via SSH, the SSH server component of Vervis handles push by directly accessing the repo, without consulting the Repository actor

Thus, we have 2 components that possibly concurrently access the repo! There's no risk of corruption, thanks to Git locking concurrent access, but there's difficulty for Repository actor handlers to be atomic, because whenever they touch the actual Git repo, there's a chance a git-push will concurrently happen and modify the repo.

To fix that,

I made a much-needed upgrade to the actor system (which is going to be the same system that ForgeFed-v2 a.k.a OcapForge a.k.a Forgely a.k.a insert-better-name-here will be using!), allowing each actor type to define its own set of methods
I added a method to the Repository actor, that simply asks it to wait while the SSH server handles a git-push

So technically, the Repository actor doesn't have exclusive write access to the repo. But whenever a push happens, it now goes through the actor message queue, which means no concurrent access anymore, and Repository actor message handlers can now be truly atomic, and easier to implement, knowing no surprising changes will suddenly appear.

Implementation

Funding

I really want to thank NLnet for funding this work! The extended grant is allowing me to continue backend work, and allowing André to work on the Anvil frontend.

See It in Action

I recorded a little demo of all this! Watch it on my PeerTube instance.

If you want to play with things yourself, you can create account(s) on the demo instances - fig, grape, walnut - and try the things I've mentioned and done in the video.

If you encounter any bugs, let me know! Or open an issue

Comments

Come chat with us on Matrix!

And we have an account for ForgeFed on the Fediverse: https://floss.social/@forgefed

Right after publishing this post, I'll make a toot there to announce the post, and you can comment there :)

index.md 9.6 KB 永久連結 文件歷史 原始文件