use std::fmt::Debug; use async_trait::async_trait; use futures_signals::signal::Mutable; use async_oneshot::Sender; use async_channel::Receiver; use crate::state::State; use crate::db::{ state::StateAccessor, }; /// A resource in BFFH has to contain several different parts; /// - Currently set state /// - Execution state of attached actors (⇒ BFFH's job) /// - Output of interal logic of a resource /// ⇒ Resource logic gets read access to set state and write access to output state. /// ⇒ state `update` happens via resource logic. This logic should do access control. If the update /// succeeds then BFFH stores those input parameters ("set" state) and results / output state. /// Storing input parameters is relevant so that BFFH can know that an "update" is a no-op /// without having to run the module code. /// ⇒ in fact actors only really care about the output state, and shouldn't (need to) see "set" /// state. /// ⇒ example reserving: /// - Claimant sends 'update' message with a new state /// - Doesn't set the state until `update` has returned Ok. /// - This runs the `update` function with that new state and the claimants user context returning /// either an Ok or an Error. /// - Error is returned to Claimant to show user, stop. /// - On ok: /// - Commit new "set" state, storing it and making it visible to all other claimants /// - Commit new output state, storing it and notifying all connected actors / Notify /// ⇒ BFFHs job in this whole ordeal is: /// - Message passing primitives so that update message are queued /// - As reliable as possible storage system for input and output state /// - Again message passing so that updates are broadcasted to all Notify and Actors. /// ⇒ Resource module's job is: /// - Validating updates semantically i.e. are the types correct /// - Check authorization of updates i.e. is this user allowed to do that #[async_trait] pub trait Resource: Debug { /// Run whatever internal logic this resource has for the given State update, and return the /// new output state that this update produces. async fn on_update(&mut self, input: &State) -> Result; async fn shutdown(&mut self); } #[derive(Debug)] pub struct Passthrough; #[async_trait] impl Resource for Passthrough { async fn on_update(&mut self, input: &State) -> Result { Ok(input.clone()) } async fn shutdown(&mut self) {} } /// Error type a resource implementation can produce #[derive(Debug)] pub enum Error { Internal(Box), Denied, } // TODO: more message context pub struct Update { pub state: State, pub errchan: Sender, } #[derive(Debug)] pub struct ResourceDriver { // putput res: Box, // input rx: Receiver, // output db: StateAccessor, signal: Mutable, } impl ResourceDriver { pub async fn drive_to_end(&mut self) { while let Ok(update) = self.rx.recv().await { let state = update.state; let mut errchan = update.errchan; match self.res.on_update(&state).await { Ok(outstate) => { // FIXME: Send any error here to some global error collector. A failed write to // the DB is not necessarily fatal, but it means that BFFH is now in an // inconsistent state until a future update succeeds with writing to the DB. // Not applying the new state isn't correct either since we don't know what the // internal logic of the resource has done to make this happen. // Another half right solution is to unwrap and recreate everything. // "Best" solution would be to tell the resource to rollback their interal // changes on a fatal failure and then notify the Claimant, while simply trying // again for temporary failures. let _ = self.db.set(&state, &outstate); self.signal.set(outstate); }, Err(e) => { let _ = errchan.send(e); } } } } } #[cfg(test)] mod tests { use std::pin::Pin; use std::task::Poll; use std::future::Future; use super::*; #[futures_test::test] async fn test_passthrough_is_id() { let inp = crate::state::tests::gen_random(); let mut res = Passthrough; let out = res.on_update(&inp).await.unwrap(); assert_eq!(inp, out); } #[test] fn test_passthrough_is_always_ready() { let inp = State::build().finish(); let mut res = Passthrough; let mut cx = futures_test::task::panic_context(); if let Poll::Ready(_) = Pin::new(&mut res.on_update(&inp)).poll(&mut cx) { return; } panic!("Passthrough returned Poll::Pending") } }