ipykernel icon indicating copy to clipboard operation
ipykernel copied to clipboard

Debugger does not conform to DAP specification

Open lukejriddle opened this issue 1 year ago • 1 comments

Per Microsoft's documentation for the Debug Adapter Protocol, a DAP client should send the launch/attach request in parallel with the various configuration requests, such as setBreakpoints and setExceptionBreakpoints. Only once those configuration requests have received responses should the client send configurationDone, at which time the adapter will return a response to the original launch/attach request. This allows the adapter to fully configure itself before the client considers itself attached.

ipykernel does not currently conform to this specification. It includes a _handle_init_sequence method which manually sends the configurationDone request to the adapter rather than use the original one from the client. It does this in response to the original attach request, too, well before the client would have sent that message itself.

This is problematic because:

  1. This method assumes DAP to work synchronously, whereas by definition the initialization sequence is meant to be conducted in parallel
  2. The configurationDone request may make it to the adapter before the actual configuration requests (setBreakpoints...). This may cause issues (which we have actually seen) where the DAP client receives a response to attach before the adapter is done configuring. This leads to breakpoints not being set in time and ultimately being missed.

The root of this problem is the synchronous handling of DAP requests by ipykernel; the kernel cannot send Request B until Request A receives a response, so this parallel nature of initialization is currently not possible to conform to.

The solution to this problem would be to change how ipykernel handles DAP requests. One solution would be to use a queue to which requests are pushed, similar to how responses are handled. We can implement a working solution and submit a PR, if the maintainers are okay with it.

lukejriddle avatar Dec 10 '24 18:12 lukejriddle

What are the user-facing consequences of the status quo? Would changing it improve the debugger UX somehow?

krassowski avatar Dec 16 '24 11:12 krassowski