Proposal: Structured MIP progress logging using event callbacks

[MySweetEden]

Motivation

For visualization, explainability, and fair comparison of MIP runs (including early termination by time or gap),
it is very useful to have a machine-readable progress log capturing:

• incumbent (primal bound)
• best bound (dual bound)
• gap
• time / nodes

While similar information is available in other solvers' MIP progress logs, in PySCIPOpt this currently requires custom event handler implementations.

Context

This proposal is inspired by and builds on the discussion in: #881

Recent PySCIPOpt versions already provide the necessary building blocks:

• attachEventHandlerCallback
• progress-related events such as GAPUPDATED and DUALBOUNDIMPROVED

This makes it possible to implement structured progress logging cleanly without modifying SCIP core behavior.

Proposal

I propose adding an official example or utility that implements structured MIP progress logging (CSV / TSV) on top of existing event callbacks.

Minimal example schema:

• time_sec
• nodes
• incumbent
• best_bound
• gap
• event

The implementation would be fully opt-in and purely event-based.

Next steps

I am currently traveling, but I plan to prepare a draft PR implementing a minimal progress logger based on this proposal next week.
Feedback on the proposed schema and triggering events would be very welcome.

[Joao-Dionisio]

Hey @MySweetEden ! Yeah, that's a great idea!! Besides leaving it in the examples, I think it should likely be included as a recipe as well, as the functionality is so neat.

It's not exactly the same, but the recent addition of the JSON statistics may be helpful as well. I don't think there's a c-SCIP public function to access these statistics, but if there were, interfacing it would be very helpful. I'll ask around 😄

[MySweetEden]

Thanks a lot for the positive feedback!

Including this both as an example and as a recipe sounds great to me.
I agree that positioning it as a clean, event-based progress logger complements the JSON statistics nicely, since this focuses on
online / incremental progress rather than post-run summaries.

I’ll keep the implementation purely on the PySCIPOpt side, based on event callbacks only, and will prepare a draft PR after my trip.
Looking forward to any additional thoughts in the meantime!

[Joao-Dionisio]

Regarding

I don't think there's a c-SCIP public function to access these statistics, but if there were, interfacing it would be very helpful.

I was told that the best way to do it would be to write the JSON statistics and later read them. So, not very elegant.

Have a nice trip!