The idea of a Sync engine is getting popular in recent frontend architectures. Why and how does it relate to declarative programming?
State: something (idk)
Event handler: updates state and has effects.
Event handler should not be async: fetching is an effect, and if its async, could lead to multiple events being handled at the same time, leading to a race. Consider an async transactional:
async deleteFile(filePath: string)
User presses button, triggering delete file. If its dispatched async, user could press it twice, invoking it twice. Which could be bad news.
Either the user has to wait out the async action before the UI is responsive again, (bad)
or we have to make the command synchronous somehow.
We do this via declarative effects. We store a piece of state whether we're currently trying to delete a file. In the first call, we set that state to true. A Service subscribes to this state, and when its set, it launches the async function. Meanwhile, the second time the user invokes the command, we'll see that the state has already been set, so we cancel the transaction. After the deletion is completed, the effect dispatches a transactional
resolveDeletedFile to set the flag back to false (and do whatever else after the file's deletion is confirmed).
Btw, this means services should be very very thin. Essentially a subscription, and a callback that fires an async call, awaits it, then dispatches a transactional to the database.
Move all the logic into synchronous transactionals. There should be hardly any logic anywhere else.
State, transactionals, View (render), and services.
Wrapping all the business logic into transactionals makes your system very very easy to test, without resolving to an e2e test.
Business logic is just initial state, a set of events, and an ending state. If all the logic is in transactionals, all you have to do is initiate your database, send events, and assert the ending state is right!
Mocking is a code smell because it means that the programmer has to simulate an external service to manage interactions with the database, instead of simply sending events themselves in the test.
Going back to the
deleteFile example: if in the asynchronous example, we just called
await deleteFile(...) we would have no choice but to create a mock service that will receive and respond to the
deleteFile request to test the logic. And then we need to make sure the mock filesystem responded correctly at all. So we write tests for that, creating more Mock services...
If instead, we used declarative effects, we respond to the database events directly in the test.
* Initialize a database
* Make sure the database is in a consistent state.
Its so much simpler!
Synchronous transactionals and declarative effects have such great implications for UX and DX. For users, it means everything happens locally, quickly. You get to pretend everything is local.
For developers, it wipes out a whole swath of race conditions you get from making event handlers asynchronous. Everything is simple, local, composition of transactinals. Also makes it easy to build new sycned state (although you have to worry about migrations.)
For a sync engine between peers, there are other things to consider, like consistency and causality. I really like (Replicache's)[https://replicache.dev] model here. It's simple and straightforward although you do need to assign a server or main node. I'm curious if there are more equal ways of handling this like via Actor model.
There is definitely an ergonomics issue here (at a much larger scale). Managing effects declaratively blows up the amount of independent state our app has when using declarative effects: there's a stateful flag for every single possible asynchronous call our app could make.
It also breaks apart the scope. If running
deleteFile asynchronously and awaiting, we had all this scope and context. If we do it as a transactional, we lose all of that context when we receive the
resolveDeletedFile, and it has to be encoded in the database state if we need it in that resolution transactional.
Can we fix this?
I'm not sure how but I think a semantic layer has to be introduced between the view's usage of state and that state's representation in the database. Something along the lines of objects?