Saving and loading composite experiments

[coruscating]

Saving composite experiments should save sub-experiments, but right now it only saves the parent experiment. Loading a composite experiment from DB should allow us to retrieve the sub-experiment data with component_experiment_data, but that doesn't work right now either.

[coruscating]

When saving, the parent experiment and sub-experiments should all be tagged with the experiment ID of the parent.

[yaelbh]

The request to load data into a CompositeExperimentData reveals a delicate point, because normally for a regular experiment we never load data into an ExperimentData object: DbExperimentDataV1.load returns a DbExperimentDataV1 object, and ExperimentData doesn't override load. So, a user cannot run specific ExperimentData methods on loaded data, like completion_times.

I propose the following solution:

1. Don't store experiment in ExperimentData; all data members in classes inheriting from DbExperimentDataV1 must be able to be saved and loaded.
2. When initializing an ExperimentData object, one will be able to provide either the experiment, or the experiment type + metadata (or none of them, resulting in the same behavior like today when experiment is None).
3. Similarly, it will be possible to initialize a CompositeExperimentData object by providing the components instead of the experiment.
4. Override ExperimentData.load: it will call load of the super class, which returns a DbExperimentDataV1 object; this will be used to create and return a new ExperimentData object.
5. CompositeExperimentData.save will do the following:
   • Call save of the super class.
   • Call save for each component.
   • Save a list of ids of components, in the entry of the composite experiment.
6. CompositeExperimentData.load will:
   • Retrieve the list of components.
   • Call load for each component.
   • Call load of the super class.
   • Use the loaded ExperimentData object, together with the loaded components, to create and return a new CompositeExperimentData object.

[yaelbh]

So, a user cannot run specific ExperimentData methods on loaded data

An additional example: a user cannot run Amplitude.update on loaded data

[yaelbh]

One option is to replace the type of the update argument from ExperimentData to Experiment. Namely, to make it explicit that if you don't have the experiment itself, then you cannot run update.

I don't like this solution, because all that update wants to know is whether the experiment is a Rabi or a FineAmplitude one. It's fair to expect this information to be stored in the experiment data. I prefer to introduce a new data member to ExperimentData, named experiment_extra_types, which will be a set of all the types that can be associated with the experiment, beside the main type experiment_type.

[yaelbh]

When saving, the parent experiment and sub-experiments should all be tagged with the experiment ID of the parent.

@coruscating How to tag a parent who's also a child? And how to tag an experiment that has nothing to do with composite experiments?

[coruscating]

@yaelbh Maybe tagging each child with their immediate parent and also the root parent? What do you think about that? And I don't think experiments that are not parent or child in any way need to be tagged.

[yaelbh]

I don't know what to think; what are the considerations, what's the purpose?