bonsai reconstruction crashes on an event generated by ratpac

[davehadley]

The version of BONSAI included within rat-pac crashes on the first event of the attached file.

example_problematic_ratpac_file.zip

The RAT-PAC macro to generate this file is:

/glg4debug/glg4param omit_muon_processes  0.0
/glg4debug/glg4param omit_hadronic_processes  0.0

/rat/db/set DETECTOR experiment "Watchman"
/rat/db/set DETECTOR geo_file "Watchman/Watchman.geo"


/run/initialize

# BEGIN EVENT LOOP
/rat/proc lesssimpledaq
# /rat/proc fitbonsai
/rat/proc fitcentroid
/rat/proc fitpath
/rat/proc count
/rat/procset update 200

# Use IO.default_output_filename
/rat/proclast outroot
#END EVENT LOOP


/generator/add decaychain 214Bi:regexfill:poisson


/generator/pos/set shield+
#/generator/rate/set 1.0

/generator/rate/set 1
/run/beamOn 1000

Inspecting the crash inside gdb the issue seems to be because inside fourhitgrid::fourhitgrid only 3 hits are selected. This leads to a seg fault since this grid object is used without checking that it is in a valid state.

I don't understand the reconstruction enough to know whether this is due to invalid input data that the reconstruction should not be expected to handle, or whether this is a valid input and the problem is in bonsai.

[MarcFBergevin]

Bonsai requires at least 4 hit for any realistic fit to be made. There used to be a condition that required at least 4 nhits prior to dark noise before. This seemed to have been taken out. It was added recently with a requirement of 9 nhits. This has not yet been ported to the version of rat-pac within Watchman.

[MarcFBergevin]

Based on conversations with the DAQ group, a detector threshold of 6 nhits is realistic. There are two potential approach here:

1. Enforce that no events below 6 hits are processed by bonsai with a if condition in the bonsai loop
2. Implement a DAQ triggering scheme.

For the second option:

1. @ekneale Is implementing a trigger scheme in lesssimpledaq.
2. @MorganAskins has two pull request open (noise generator and splitevdaq

Finally, @katfra noticed that reconstruction from the bonsai version in rat-pac has much poorer reconstructions than the bonsai standalone. From my understanding, this was not completly explained by the change of likelihood.

[davehadley]

This has not yet been ported to the version of rat-pac within Watchman.

@MarcFBergevin, thanks for looking at this issue. For now, does that mean it is recommended to use the HEAD version from: https://github.com/AIT-WATCHMAN/bonsai/ ?

On this particular issue, I may be misunderstanding, but in this case, the crashing event has nhits=59.
BONSAI constructs a goodness object (a subclass of hitselection) that selects only 3 hits (i.e. nselected() returns 3).
This leads the call to fourhitgrid::fourhitgrid and likelihood::maximize to return early (since there aren't enough hits).

The segfault happens on the line:

x = 10. * bsfit->xfit();

Since likelihood::maximize did not complete the bsfit object contains no fit result.

So, naively, it seems like a bug in BONSAI and that the correct action is to modify BONSAI to check that the bsfit object is in a valid state before using its results.