Reintroduce debug as a valid global key for Pipeline's params

[ZanSara]

Problem
Passing the debug parameter to a pipeline as a global arg (like p.run(query="Why?", params={"debug": True}) ) did not work, as debug was not included in the set of valid global parameters which was populated from node signatures.

Solution
Add debug in the valid global parameters set in Pipeline.run()

[julian-risch]

Please note that pipeline.run() has a debug parameter:

haystack/haystack/pipelines/base.py

Line 568 in fd46a42

debug: Optional[bool] = None,

@ZanSara Maybe there is some confusion regarding that?

[ZanSara]

AFAIK we allow both options by design. We actually have several ways to enable debug output in a pipeline:

1. By setting it to one component explicitly: es_retriever.debug = True

2. By providing a debug as a parameter to one component at run: result = p.run(query="Why?", params={"ESRetriever": {"debug": True}})

3. By providing it as a global parameter of the pipeline: result = p.run(query="Why?", params={"debug": True})

4. By setting it explicitly to the pipeline: result = p.run(query="Why?", debug=True)

We can argue they're too many ways, but as long as they're documented I think they should all work. In addition, the code looks like all the four options worked well at some point, and this is a regression (just check how small is the change I needed to make to re-enable option 3).

However, deprecating some of these is also an option. @brandenchan what do you think?

[brandenchan]

So I see that option 2. is a needed feature so we should definitely keep that in code and document it.

We also need a global debug method and I don't mind whether we have 3. / 4. / both. Option 4. will have a nice docstring accompanying it but Option 3. is perhaps a little closer to how they are used in option 2. Either way, I think its better in our docs if we only recommend one of the two.

Option 1. should be a byproduct of having the global debug so I imagine it will always work, but I also think that we don't have to document this.

[ZanSara]

Ok, I agree option 2 is necessary as it is and option 1 is a byproduct of 2. Between 3 and 4, what do you prefer? I have no preference to be honest. Keeping both is not an issue code-wise, but I will disable either of them for clarity if we decide that it's better that way 🙂

[julian-risch]

However, deprecating some of these is also an option. @brandenchan what do you think?

If we decide to deprecate one of the options then I'd argue for deprecating option 4 (result = p.run(query="Why?", debug=True)). It was added in this PR: #1558 but we can achieve the same with option 3: result = p.run(query="Why?", params={"debug": True})

[julian-risch]

That way, debug is handled in the same way as any other parameter that can be passed to all nodes.

[brandenchan]

If we decide to deprecate one of the options then I'd argue for deprecating option 4 (result = p.run(query="Why?", debug=True)). It was added in this PR: #1558 but we can achieve the same with option 3: result = p.run(query="Why?", params={"debug": True})

Sure I can get behind this though currently I don't feel a strong need to deprecate one.

[brandenchan]

LGTM