Reproducibility problem

[jcoupey]

While doing some benchmarking for #1267 I noticed a behavior/solution change that turned out to be unrelated with the code changes, but somehow due to a lack of reproducibility across runs. I've been able to trigger with current master and some of the usual benchmarks:

$ for i in `seq 1 10`; do vroom -x 5 -t 32 -i vroom-scripts/benchmarks/MDVRP/cordeau/p15.json | jq .summary.cost; done
2551700
2551700
2551700
2551700
2552929
2551700
2551700
2551700
2551700
2552929

[las59]

Hello,

Thank you for the work done !
Do you have a scheduled date for resolving this bug?

[jcoupey]

Nothing scheduled here. For the record, I suspect this is due to a race condition, actually the exact same situation as in #1199. It looks like the fix there (comparing ordered route length) is not catching the difference here.

This should be quite rare, and also related to the fact that the benchmark instance does not have a big variety in costs since using euclidean distance.

@las59 did you run into a similar situation? If so sharing another instance would be useful.