Only swallow domain_errors in various algorithms

[WardBrian]

Submission Checklist

• Run unit tests: ./runTests.py src/test/unit
• Run cpplint: make cpplint
• Declare copyright holder and open-source license: see below

Summary

Closes #3258. This changes any place which caught std::exception from the model to only catch std::domain_error, matching what is done during initialization.

This prevents odd behavior like an indexing exception that only occurs randomly being swallowed if initialization happens to pass, and allows things like an exit() function in the language.

Intended Effect

Cause certain exceptions to terminate the inference algorithms.

How to Verify

This model will usually initialize but quickly fail now:

parameters {
  real mu;
}
model {
  mu ~ normal(0, 1);

  if (mu > 1) {
    array[1] int sigma;
    print(sigma[2]);
  }
}

Side Effects

I've installed handlers in the service functions, but code that calls the internals of these algorithms (e.g run_adaptive_sampler) not through the services will need to be ready to catch exceptions

Additionally, if any Math functions raise an exception which is intended to be recoverable but is not a domain_error, they will need updating. Because this is the current behavior in initialization, I suspect we have probably already smoked any of these out.

Documentation

Copyright and Licensing

Please list the copyright holder for the work you are submitting (this will be you or your assignee, such as a university or company):

Simons Foundation

By submitting this pull request, the copyright holder is agreeing to license the submitted work under the following licenses:

• Code: BSD 3-clause (https://opensource.org/licenses/BSD-3-Clause)
• Documentation: CC-BY 4.0 (https://creativecommons.org/licenses/by/4.0/)

[WardBrian]

@SteveBronder mind giving this a look?

[SteveBronder]

I think this is fine, one thing I would like clarified is what happens to things like std::out_of_range errors.

[SteveBronder]

Can we move this to the outer part of the while loop?

[SteveBronder]

Same here. Generally for any while loops I'd like to just have the try on the outside

[WardBrian]

Moved both

[SteveBronder]

Why? One pathfinder can fail while the others succeed?

[WardBrian]

If one path does something unrecoverable like indexes into a vector out of bounds, I think the right thing to do is stop and say "something is clearly wrong with your model".

Besides that, a primary motivation behind this PR is also to implement an exit() function into the language, which requires the ability to terminate the entire algorithm it is in. If one chain or one path calls exit, the entire thing still needs to halt.

[SteveBronder]

If one path does something unrecoverable like indexes into a vector out of bounds, I think the right thing to do is stop and say "something is clearly wrong with your model".

Are only unrecoverable errors able to be thrown here? Like if the call to log_prob throws it could be from something recoverable like one of the ode solvers failing for a given set of parameters

[WardBrian]

pathfinder_impl catches std::domain_errors, which is what the ODE solvers throw (and is generally what the math library considers 'recoverable'), so the only exceptions which reach this will be unrecoverable ones from the model or issues in our algorithm

[WardBrian]

std::out_of_range would be considered fatal by this and lead to the algorithms terminating. Is there something else you're looking for clarification on about them?

[SteveBronder]

out of range can make sense, but what about std::invalid_argument? Like quad_form_sym would throw that for symmetric matrices. Would we just fail on that error and end the program?

[WardBrian]

Those have been considered unrecoverable (at least during initialization) for a while. invalid_argument is thrown for things like check_square, which won't change if you just 'try again'.

I think if there are usages in Math we want to be recoverable, we should change those to match the others

[SteveBronder]

Okay looking at this again I think this is fine to merge 👍