Add dump & restore to holders & simulations.

[eraviart]

New features

• Add dump & restore to holders & simulations (fix Provide a way to dump to/restore fom disk the state of a simulation  #661).

---

EDIT (@fpagnoux)

Dump:

from openfisca_core.tools.simulation_dumper import  dump_simulation, 
dump_simulation(simulation, '/path/to/directory')

Restore:

from openfisca_core.tools.simulation_dumper import restore_simulation
simulation_2 = restore_simulation('/path/to/directory', tax_benefit_system)

[benjello]

@fpagnoux : I took the liberty to ask you to review this PR.
@eraviart coded this during a session at IPP.
We need your input before we go forward and we add higher level methods to SurveyScenario.

[fpagnoux]

Thanks for the PR @eraviart!

A few general questions @benjello before looking more into details:

• With this implem, after we restore the simulation, nothing is loaded in memory and everything is kept on disk. Is that on purpose? This could significantly slow down calculations after restoring.
• Entities compositions and roles are not stored, is that an issue?

[benjello]

@fpagnoux : I knew that your input would be valuable !

• You first point is a very good one. For performance issue we may want that the variables are kept in RAM when read.
• The second point is another good one, we definitively want to keep those if we want to compute more variables involving several entities after restoring. I completely forgot that when explaing the problem to @eraviart.

[fpagnoux]

Ok, I will try adjust the implem to take these two points into account :)

[fpagnoux]

Tested on France calculating revenu_disponible for 40000 persons, dumping and restoring takes a reasonable time (<10s)

Mahdi, do you want to try the branch sees if it works for you, or should we go ahead and ask for a review?

[benjello]

Thanks @fpagnoux !

Do you confirm that when restoring a simulation, the variables are kept back in RAM/disk according to the new memroy_config only when recalled and not restored at once ?

I think we should ask for a review so I can go on and propagate to downstream packages.
I could definitively come back and ask for a few changes if needed by a use case.

[sandcha]

Use absolute import?

[benjello]

There are many more relative imports in tests. Should we migrate them all to absolute imports
@sandcha @fpagnoux ?

[fpagnoux]

We can't use absolute imports as the tests are not packaged in openfisca_core.

[fpagnoux]

Do you confirm that when restoring a simulation, the variables are kept back in RAM/disk according to the new memroy_config only when recalled and not restored at once ?

Yes, restore_simulation takes the same arguments than the Simulation constructor, and the data are restored through the usual put_in_cache method so the memory config will be respected.

[benjello]

Looks good to merge then @fpagnoux !