A proposed alternative to "suppress_callback_exception"

[eddy-ojb]

I use the "suppress_callback_exceptions" option a lot but think it makes sense to have an option where we can add e.g. a decorator to callbacks we care about and see why they aren't firing (.e.g "cannot see 'x' div, or 'cannot parse children of div' or w/e. I can see this being super helpful, rather than completely suppressing the exceptions and leaving the developer to guess everywhere.

[chriddyp]

In principle I like this but I'm not sure technically how it would (i.e. how would we know that a callback isn't firing).

Could you walk through a couple of simple examples that demonstrate cases where callbacks wouldn't fire but we could provide better error messages?

[rmarren1]

dash-renderer recently got a new feature which might help in situations where suppress_callback_exceptions covers up an error: https://github.com/plotly/dash-renderer/blob/master/CHANGELOG.md#0131---2018-07-18.

[eddy-ojb]

For me, the point that @rmarren1 makes in issue 57 hits the nail on the head: "[suppressing] puts you in debugging hell for large applications".

Apologies for missing this link. I will try update my renderer and see if this helps first.

If this doesn't help, happy to make another example of where it can go wrong and what could be useful in its place. I agree with you guys where you suggest in another thread that better exception reporting is a cheap but effective solution here, rather than channeling info through javascript back to the console.

[eddy-ojb]

@chriddyp shall i close the issue and move to another thread?

[rmarren1]

That error information will be more visible in the future after plotly/dash-renderer#64 is worked through, which should functionally give you the same information as if exceptions were not suppressed.

[eddy-ojb]

This actually excites me 😃. I updated my renderer to 13.1 - still some issues but can see theres a lot of work that you're done already to address this .

I think adding other exceptions to callbacks checks would be helpful too. For example, if you consider creating the following callback for a dynamic dropdown:

@app.expanded_callback(
    dash.dependencies.Output('dropdown_holder1', 'children'), 
    [dash.dependencies.Input('some-div', 'children')])
def create_dropdown1(div_value):
    
    # check if some condition met. if so create dropdown:
    dropdown  = dcc.Dropdown(
                            id='my-dropdown',
                            options=[
                                    {'label': 'New York City', 'value': 'NYC'},
                                    {'label': 'Montreal', 'value': 'MTL'},
                                    {'label': 'San Francisco', 'value': 'SF'}
                                    ],
                            value='NYC',
                            className='col-md-12',
                            ),

    return dropdown

the dropdown is a dcc, yet you have to return it as a div with children, without it being in a list.

This baffled me for ages as I was trying to return it as:
return [dropdown]

This is because it is inline with outputting a dcc in this example:

@app.expanded_callback(
    dash.dependencies.Output('graph-div', 'children'), 
    [dash.dependencies.Input('my-dropdown', 'value')])
def create_dropdown1(div_value):

    # create graph
    some_graph = dcc.Graph(..) 
    return [some_graph] `

but:
return some_graph does not work.

Is it possible to check the type of the content returned by a callback and check that it is compatible with the output type specified in the callback when in debug mode?
Alternatively, I have seen the @chriddyp solution here, which I haven;t tried yet but will give a go soon.

[rmarren1]

Is it possible to check the type of the content returned by a callback and check that it is compatible with the output type specified in the callback when in debug mode?

We are actually working on a big PR #340 which does just that. I'm pushing changes to it currently and will have a pre-release soon, so you can subscribe to that and be one of the first to test-drive it 🙂

Also I would expect those examples above to work, do they not? If so it may be a bug we can open.

[eddy-ojb]

Yes it doesnt work with my current setup. Before this is opened i will update my packages and build some toy examples tomorrow to confirm its still relevant.

[chriddyp]

Also I would expect those examples above to work, do they not? If so it may be a bug we can open.

Yeah, if a callback updates children, any of these return objects should work:

• string
• single dash Component
• number
• None
• Or a list of those

[eddy-ojb]

This is a bug. I've created a Gist to illustrate 2 problems. I've posted version info below to replicate the issue.

1. If deleting the only component in the layout, no dashboard is rendered and no error message is displayed to inform the developer - this stumped me for a while.

2. At the moment, returning a dropdown, radio button or input dcc as e.g. [dropdown] in create_dropdown1 breaks the callback process. This is inconsistent with the Graph dcc, which now returns as [graph] or graph. I tried only with these 4 components.

@rmarren1 have subscribed to #340 and happy to test drive the new updates.

The Dash versions to replicate the above are currently inline with official installation versions:

pip install dash==0.26.2  # The core dash backend
pip install dash-renderer==0.13.1  # The dash front-end
pip install dash-html-components==0.11.0  # HTML components
pip install dash-core-components==0.28.0  # Supercharged components
pip install plotly==3.1.1
pip install numpy

[rmarren1]

Check line 51 of the gist, you end have dcc.Dropdown(...),. When you end with a comma like that, python casts this to the tuple (dcc.Dropdown(...),) which is not valid.

This is a tough error to debug though, as it silently fails and just says component.type is undefined as a front-end exception, and is a good example of a case where #340 can save a lot of headaches.

[eddy-ojb]

Good spot! and thanks!

Actually, to hammer home the importance of your work on #340, I did spot it was a tuple but forgot about it trying to solve error 1.