Meeting minutes

Shift handover

Marta --> Vincenzo

• Builds are red mostly due to TMVA
• ROOT forum is all handled

PoW status

Looking good for now, 19%, the outlook is positive considering more people will come

Meeting reports

• LIM: Some bugs regarding PyROOT due to the move to latest cppyy
• Experiments: short report by Danilo on last week's trip to FermiLab
  • Important to show our presence in person there
  • LPC group at FNAL organized an unsolicited informal meeting to connect scientists there with the ROOT developers
  • The direction of ROOT regarding RNTuple is confirmed and strenghtened
  • We visited Argonne, met with Peter Van Gemmeren's group. People were positively engaging with us
  • Generally, people showed a lot of respect for the project and they wanted to engage

RooFit workshop summary

• 2 day workshop, participation from experiments (ATLAS,CMS,LHCb) and NIKHEF (~20 people)
• Concrete discussions regarding the features, limitations and future directions of RooFit
• Minuit2:
  • Designed for a few dozens parameters, now using O(1000)
  • This leads to problems that only ROOT devs can fix
  • New Strategy 3 for Minuit helped ATLAS analysis converge (contribution by Will B.)

• Clad:

  • Presentation by Vaibhav
  • presentation well received
  • Integrated in Jupyter notebook in SWAN via ROOT: this was particularly liked!

• CMS input:

  • Hackathon session to help HiggsCombine up to speed with new ROOT version (right now using 6.22)
  • CMS published a paper to explain HiggsCombine, last week they published a docker image with all histograms and data cards as input for Run1 Higgs discovery.
  • The workflow runs perfectly well with ROOT master, in roughly 1 minute

Comment: Let's put the workflow in our CI!

• ATLAS

  • ATLAS has many different frameworks for statistical analysis, most of them are based on RooFit
  • Q: Why so many?
  • A: There is a lot of "art" involved in the statistical treatment, that domain knowledge has to be encoded somewhere. Ideally, this could be written in a high-level script. But often what started as simple descriptive analysis frameworks started to incorporate some workarounds for RooFit and Minuit, growing organically in that direction. Sometimes they grow from different Physics subgroups (e.g. TRexFitter from the top analysis group).
  • Will Buttinger is starting an effort called xroofit to provide higher-level abstractions for all other ATLAS statistical analysis tools.

• Q: How do the conclusions of the RooFit workshop impact the PoW?

• A: The PoW was actually discussed at the workshop. I don't think this year the PoW will be impacted, we already have itmes that go in the direction of improving the minimizers (performance, Python interfaces)

• Q: Are the numerical instabilities of the minimizer understood? If something goes wrong, do we know at "which line" we should look or is there more of a holistic approach?

• A:

  • Some of the numerical effects are well understood, those related to high statistics (high statistics, small uncertainties). Combining parameters with small uncertainties with parameters with large uncertainties is problematic.
  • Something much less understood is related to a new trend called "effective field theory fits". In the end it boils down to your likelihood not looking very parabolic, which is one of the main assumptions of the Minuit2 algorithm.
  • Some other parameters are ATLAS internal, it would benefit from an open repository to test it.
  • Something we did is improve the algorithm that computes the numerical derivative (Strategy 3 previously mentioned)

• comment: Next time the workshop should be organized a bit more in advance.

• A: I wasn't sure how large of an audience we wanted to have. I preferred a more compact format with a very focused audience.

Topic: TMVA breaking CI builds

Last week there were many PRs put into place to mitigate the many problems arisen in the CI due to TMVA, which is broken in different ways depending on which Python packages are installed (and which versions). The current approach in the tests is to just not run (silently) if the Python package is not found, which is a ticking bomb.

Goal: We need green builds, now.

Action item: Create one CMake build option per type of test that is required by TMVA (see what distrdf does for test_distrdf_dask).

Action item: Disable the last remaining test that is constantly failing on Alma9, open a github issue to keep track of it.