enable termination of 'outdated' long callbacks when page changes

[JamesKunstle]

Is your feature request related to a problem? Please describe.
Long callbacks run for all Figures on all of my application's pages. When I switch between pages quickly the tasks in the backend Celery queue execute in-order regardless of old requests being outdated.

Ex: If I'm on page 1, and I navigate to page 2, then quickly to page 3 then page 4, the long callbacks for figures corresponding to pages 2, then 3, then 4 will run in order, slowing the delivery of Figures pertinent to page 4.

However, this isn't the case if I refresh a single page multiple times. If I'm on page 2 and I refresh the page 5 times, the long callbacks for the Figures associated with page 2 don't execute 5 times in order- the first 4 'render requests' are 'revoked' in the Celery queue, with the final request completing.

Describe the solution you'd like
I'd like a page switch to trigger the termination of newly 'outdated' long callbacks in the backend.

Describe alternatives you've considered
In the universe of Dash objects, I considered implementing a 'Sentry' that watches the URL and triggers a 'cancel' function that's bound to all long callbacks, cancelling them via the 'cancelable' interface of callbacks. In the Dash execution graph, however, I don't think I can guarantee that this callback would always precede the scheduling of new, up-to-date callbacks, so I don't think this is a reasonable solution.

Additional context
Here's an idea for a patch:

The dash_renderer frontend is aware of when it's going to try to execute a callback for a job that is still running. If a to-be-called callback's output list matches the output list of the job that the frontend is waiting for, it issues an 'oldJob' ID to the backend via the request headers. Source

In the backend, receiving these 'oldJob' ID's triggers that job's termination in the Celery backend. Source

It's evident that the frontend is doing job bookkeeping, tracking the 'output' list of jobs that have been scheduled in the backend but that haven't returned for the frontend. If the frontend could also track the page that the job is intended for, and compare that to the window.location.pathname when the job cleanup is already happening, the 'oldJob' param could be set and the backend could clean up any currently running long callbacks for the previous page.

A parameter could be set in the Dash python object to enable and disable this feature- it'd generally NOT be mutually exclusive of memoization because cancellations wouldn't always happen but it'd be more conservative, and memoization wouldn't always have an opportunity to happen.

[cdolfi]

Very interested in this patch as well

[BSd3v]

Hello @JamesKunstle,

Please note my response here:

https://community.plotly.com/t/cancelling-outdated-background-callbacks-when-switching-pages/76765

[BSd3v]

When trying to use this:

import time
import os

import dash
from dash import Dash, DiskcacheManager, CeleryManager, Input, Output, html, callback, dcc

if 'REDIS_URL' in os.environ:
    # Use Redis & Celery if REDIS_URL set as an env variable
    from celery import Celery
    celery_app = Celery(__name__, broker=os.environ['REDIS_URL'], backend=os.environ['REDIS_URL'])
    background_callback_manager = CeleryManager(celery_app)

else:
    # Diskcache for non-production apps when developing locally
    import diskcache
    cache = diskcache.Cache("./cache")
    background_callback_manager = DiskcacheManager(cache)

app = Dash(__name__, background_callback_manager=background_callback_manager, use_pages=True, pages_folder='')

app.layout = html.Div(
    [
        dcc.Link('page1', '/'), dcc.Link('page2', '/2'),
        dash.page_container
    ]
)

layout1 = [html.Div([html.P(id="paragraph_id", children=["Button not clicked"])]),
        html.Button(id="button_id", children="Run Job!"),
        html.Button(id="cancel_button_id", children="Cancel Running Job!")]

layout2 = [html.Div([html.P(id="paragraph_id2", children=["Button not clicked"])]),
        html.Button(id="button_id2", children="Run Job!"),
        html.Button(id="cancel_button_id2", children="Cancel Running Job!")]

dash.register_page('page1', path='/', layout=layout1)
dash.register_page('page1', path='/2', layout=layout2)

@callback(
    output=Output("paragraph_id", "children"),
    inputs=Input("button_id", "n_clicks"),
    background=True,
    running=[
        (Output("button_id", "disabled"), True, False),
        (Output("cancel_button_id", "disabled"), False, True),
    ],
    cancel=[Input("cancel_button_id", "n_clicks"), Input('_pages_location', 'href')],
)

@callback(
    output=Output("paragraph_id2", "children"),
    inputs=Input("button_id2", "n_clicks"),
    background=True,
    running=[
        (Output("button_id2", "disabled"), True, False),
        (Output("cancel_button_id2", "disabled"), False, True),
    ],
    cancel=[Input("cancel_button_id2", "n_clicks"), Input('_pages_location', 'href')],
)
def update_clicks(n_clicks):
    time.sleep(4.0)
    return [f"Clicked {n_clicks} times"]


if __name__ == "__main__":
    app.run(debug=True)

Ran into an issue with duplicate outputs on the ids, this is most likely due to the cancel being listed in a callback, should probably alter these to allow for duplicate outputs.

@alexcjohnson

[JamesKunstle]

Here's a direct reference to the code where this change could be made:
plotly/dash/blob/a7a12d180e16eac0a84b88e2b8dc8f7c7601cbaf/dash/_callback.py#L174

[alexcjohnson]

Spot on @JamesKunstle @BSd3v - I think (a) the cancel callback should always allow duplicates since it's not actually doing anything with that output, and (b) we should explore a dedicated feature that maybe even by default cancels background callbacks on page changes, since it seems like generally the output would be discarded in that case anyhow. It would need a way to opt out of course, or we make the feature opt-in if we think this would constitute a breaking change. @T4rk1n thoughts?

[T4rk1n]

We should add allow_duplicate to the outputs of the cancel callback, but there is probably more to do in the renderer to collect the cancelJobs since it should have already been working with only one callback.

[T4rk1n]

Found more issues when trying to fix:

• Adding allow_duplicate with a single same Input still throw a DuplicateOutput on the renderer
• the cancel inputs on the pages are not found on the renderer check.

[T4rk1n]

@alexcjohnson I tried combining all the cancel inputs into a single callback, there is a validation error:

A nonexistent object was used in an "Input" of a Dash callback. The id of this object is

That seems like a duplicate of another validation we have that takes the validation_layout and is configurable with suppress_callback_exception. The other validation runs on every callbacks seems to clash with pages if you have a callback that shares inputs it makes that impossible. Think we can remove that?

[alexcjohnson]

If there are bg callbacks on separate pages (or in whatever way not all in the DOM simultaneously) then we'd have some inputs present and others missing, right? That's not going to work, unless we make optional inputs, something we've talked about at various points in the past.

But I guess due to the DuplicateOutput issue we'll need to combine any bg callbacks with identical cancel triggers into a single callback.

[T4rk1n]

we'd have some inputs present and others missing, right? That's not going to work, unless we make optional inputs, something we've talked about at various points in the past.

Yes, that is what I was looking to do, turn out to be a pretty heavy refactor.

Maybe one last thing I can try is isolate all the callback inputs in single callbacks for each input, I think that might work.

[JamesKunstle]

@T4rk1n @BSd3v I saw that a PR was merged for this, allowing for the 'cancel' Input callback to be reused for multiple background callbacks. Wanted to check in on whether this issue should also be closed, or whether something else needs to be finished first?

Is it reasonable to expect this patch to be in the next major release?

[T4rk1n]

@JamesKunstle It was a partial fix, I think there is still value in adding auto cancellation of running tasks when leaving the page so we may leave this issue open. What do you think @alexcjohnson ?

Is it reasonable to expect this patch to be in the next major release?

Yes, it's going in next release.