We recently changed the way parameters are injected into the benchmark records. All Python-native types and containers thereof are passing unchanged, but user-defined types are serialized to string (their repr() outputs to be exact), which can be surprising.
Since we do not yet mention this anywhere in the docs, we could set up a small section on this practice, and recommend a way to handle types for best odds of reproducibility (e.g. dumping unique information into the output, or overloading the repr() method to give all unique info).
We recently changed the way parameters are injected into the benchmark records. All Python-native types and containers thereof are passing unchanged, but user-defined types are serialized to string (their repr() outputs to be exact), which can be surprising.
Since we do not yet mention this anywhere in the docs, we could set up a small section on this practice, and recommend a way to handle types for best odds of reproducibility (e.g. dumping unique information into the output, or overloading the repr() method to give all unique info).