-
Notifications
You must be signed in to change notification settings - Fork 16
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Include output from interactive cells in Foyle requests #1756
Conversation
This fixes a bug in the serializaiton of the notebook before sending it to Foyle that caused the output of interactive cells not to be included in the requests. The problem is that we need to call addExecInfo before converting the VSCode NotebookData representation to the proto. That handles copying the output of the interactive terminals into the NotebookData structure. This necessitated some code refactoring. In order to call addExecInfo we need an instance of the kernel. We create a new Converter class to keep track of the kernel and also provide reuse in the logic for converting notebook data to protos for Foyle. Since addExecInfo is async we need to change buildReq to return a promise and refactor some of the logic to be non blocking. * Fix jlewi/foyle#286
@sourishkrout this is ready when you are. |
stateful#1756 branch: jlewi/outputs commit d279e74 Author: Jeremy Lewi <jeremy@lewi.us> Date: Wed Oct 23 15:25:38 2024 -0700 Include output from interactive cells in Foyle requests This fixes a bug in the serializaiton of the notebook before sending it to Foyle that caused the output of interactive cells not to be included in the requests. The problem is that we need to call addExecInfo before converting the VSCode NotebookData representation to the proto. That handles copying the output of the interactive terminals into the NotebookData structure. This necessitated some code refactoring. In order to call addExecInfo we need an instance of the kernel. We create a new Converter class to keep track of the kernel and also provide reuse in the logic for converting notebook data to protos for Foyle. Since addExecInfo is async we need to change buildReq to return a promise and refactor some of the logic to be non blocking. * Fix jlewi/foyle#286
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
✅ LGTM
I will let you resolve the merge conflict.
src/extension/ai/stream.ts
Outdated
@@ -82,45 +82,45 @@ export class StreamCreator { | |||
} | |||
|
|||
log.info('handleEvent: building request') | |||
let req = this.handlers.buildRequest(event, firstRequest) | |||
this.handlers.buildRequest(event, firstRequest).then((req) => { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
You could likely make handleEvent = async (event: ...
and use await here instead of then
. Not an issue but more readable.
There are a few other instances of this in the code. Not a merge-stopper.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
You could likely make handleEvent = async (event: ... and use await here instead of then. Not an issue but more readable.
I'll take a look at making that change. I thought about making the function async and I wasn't sure how deep that would propogate if I had to make the callers async but I'll give it a shot.
src/extension/ai/ghost.ts
Outdated
), | ||
) | ||
// N.B. handlEvent is aysnc. So we need to use "then" to make sure the event gets processed | ||
this.streamCreator |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
@sourishkrout Is this the right way to call an async function from a non async function?
Since the return type is null we don't need to await the async function.
However, I think in the past you mentioned that if you invoke an async function but don't do anything with its return value it might not get scheduled.
Does using "then(()=> {})" solve this problem?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It does not solve the problem. At JS runtime, there isn't such a thing as an "async function". What makes it async is returning a Promise<>
type. Async/await uses a generator under the hood to pause execution to unravel promises; however, semantically, there's no difference to .then(...)
. It just makes you write async code that looks much more like sync code for readability.
That being said, if a promise type ("a future") is not being then'd or awaited upstream, you wind up with the same exact problem no matter how the promise is "returned."
The only reason this likely appears to work is that promises start running immediately (as opposed to when you await/then them), and VS Code is a long-running process, so the scopes live long enough to allow the promises to complete. I'm making assumptions here because I haven't thoroughly inspected the upstream code.
As a Golang programmer, this is the same as running three go func()
inside a function but not using a WaitGroup to synchronize them. Similarly, this might work if the "main thread" runs long enough for all three to complete and no downstream processing requires their completion. Otherwise, all concurrent functions get killed when the main thread dies.
In other words, my suggestion to use async
was only intended to be cosmetic. If there is a problem with async execution, both then
and await
will be prone to it.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The only reason this likely appears to work is that promises start running immediately (as opposed to when you await/then them), and VS Code is a long-running process, so the scopes live long enough to allow the promises to complete
That seems fine to me. Concretely, it seems fine to treat it as a fire and forget and assume
- vscode runs long enough for them to complete
or - If vscode shuts down early enough then the request might not have completed.
Here's the problem I'm trying to solve. handleEvent is an async function being called from a non async function handleOnDidChangeNotebookCell. handleOnDidChangeNotebookCell is the listener for events
vscode-runme/src/extension/ai/manager.ts
Line 60 in 165da75
vscode.workspace.onDidChangeTextDocument(eventGenerator.handleOnDidChangeNotebookCell), |
I originally thought I couldn't declare handleEvent as async and include an await function because I thought the listener couldn't be an async function. However, it looks like if I declare handleOnDidChangeNotebookCell to be async it still works fine; so I could update it to be async and then include an await function.
However, my suposition is that if the listener (handleOnDidChangeNotebookCell) is async its returning a promise that is not being awaited on. So we still wind up with a Promise that we aren't awaiting its just happening in a different part of the code.
Fundamentally, I think this is alright. The listener (handleOnDidChangeNotebookCell) is firing off a request to generate a completion and the response is handled asynchronously. If vscode exits before that async function can get scheduled and finish processing than we are dropping the completion generation logic on the floor rather than doing some graceful handling/shutdown. But dropping it on the floor seems fine; I'm not sure what graceful handling would actually look like.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
So I changed handleOnDidChangeNotebookCell and added some awaits. I made this change because I agree with your earlier comment that using await is cleaner.
@sourishkrout so I think this is good to go but let me know if you think otherwise.
@sourishkrout I updated the code to use async; I left a question for you in the comments can you please take a look? The other thing is it doesn't look like interactive cells are properly triggered if you have a long running command. I was hoping #1744 fixed this but it looks like #1744 only fixed it for non-interactive. I'll update jlewi/foyle#309 |
* Include output from interactive cells in Foyle requests This fixes a bug in the serializaiton of the notebook before sending it to Foyle that caused the output of interactive cells not to be included in the requests. The problem is that we need to call addExecInfo before converting the VSCode NotebookData representation to the proto. That handles copying the output of the interactive terminals into the NotebookData structure. This necessitated some code refactoring. In order to call addExecInfo we need an instance of the kernel. We create a new Converter class to keep track of the kernel and also provide reuse in the logic for converting notebook data to protos for Foyle. Since addExecInfo is async we need to change buildReq to return a promise and refactor some of the logic to be non blocking. * Fix jlewi/foyle#286 * Update to use await. * Add a comment. * Use await.
This fixes a bug in the serializaiton of the notebook before sending it to Foyle that caused the output of interactive cells not to be included in the requests.
The problem is that we need to call addExecInfo before converting the VSCode NotebookData representation to the proto. That handles copying the output of the interactive terminals into the NotebookData structure.
This necessitated some code refactoring. In order to call addExecInfo we need an instance of the kernel.
We create a new Converter class to keep track of the kernel and also provide reuse in the logic for converting notebook data to protos for Foyle.
Since addExecInfo is async we need to change buildReq to return a promise and refactor some of the logic to be non blocking.