storage can create ensemble with state HAS_DATA when no data present

[eivindjahren]

The following fails:

def test_empty_ensemble_should_be_undefined():
    with open_storage("/tmp/storage", mode="w") as storage:
        experiment = storage.create_experiment(
            parameters=[
                GenKwConfig(
                    template_file="//tmp/tmp280u73a6/template_file",
                    transfer_function_definitions=[],
                    name="0",
                    forward_init=False,
                    output_file=None,
                )
            ],
            responses=[
                SummaryConfig(
                    name="A",
                    input_file="D",
                    keys=[],
                    refcase=set(),
                ),
            ],
        )
        ensemble = storage.create_ensemble(ensemble_size=1, experiment=experiment.id)
        assert ensemble.get_ensemble_state() == [RealizationStorageState.UNDEFINED]

The same issue reappears when you do create_ensemble with prior:

def test_creating_ensemble_from_empty_prior_results_in_parent_failure():
    with open_storage("/tmp/storage", mode="w") as storage:
        experiment = storage.create_experiment(
            parameters=[
                GenKwConfig(
                    template_file="//tmp/tmp280u73a6/template_file",
                    transfer_function_definitions=[],
                    name="0",
                    forward_init=False,
                    output_file=None,
                )
            ],
            responses=[
                SummaryConfig(
                    name="A",
                    input_file="D",
                    keys=[],
                    refcase=set(),
                ),
            ],
        )
        ensemble1 = storage.create_ensemble(ensemble_size=1, experiment=experiment.id)
        ensemble2 = storage.create_ensemble(
            ensemble_size=1, experiment=experiment.id, prior_ensemble=ensemble1
        )
        assert ensemble2.get_ensemble_state() == [
            RealizationStorageState.PARENT_FAILURE
        ]

This situation can be provoked from a config by setting

NUM_REALIZATIONS  100
ECLBASE BASE

and then not having any summary observations or history observations.

[frode-aarstad]

@eivindjahren are the tests above part of a branch about to be merged - or should I add similar tests to this PR?

[eivindjahren]

@frode-aarstad You can just uncomment the lines marked with issue 7046 here:

ert/tests/unit_tests/storage/test_local_storage.py

Lines 385 to 389 in 4743427

	# https://github.com/equinor/ert/issues/7046
	# assert (
	# ensemble.get_ensemble_state()
	# == [RealizationStorageState.UNDEFINED] * ensemble_size
	# )

[frode-aarstad]

The issue here is if we should allow construction of SummaryConfig with no keys. When reading summary and this is the case we throw an exception.

(https://github.com/equinor/ert/blob/6895be5119c0769fdd5899f75d6be571459df206/src/ert/config/summary_config.py)

[eivindjahren]

The issue here is if we should allow construction of SummaryConfig with no keys. When reading summary and this is the case we throw an exception.

(https://github.com/equinor/ert/blob/6895be5119c0769fdd5899f75d6be571459df206/src/ert/config/summary_config.py)

Can the bug be triggered by putting SUMMARY "" in the config file?

[eivindjahren]

It might be that all we need is to update the documentation of RealizationStorageState. I'll have a check.

[eivindjahren]

Storage is quite inconsistent with what it considers "HAS_DATA". Usually, a forall quantifier in front of an empty collection is considered True (see https://en.wikipedia.org/wiki/Vacuous_truth). That is why all([]) == True.

However, when LocalEnsemble.response_configuration is empty, _all_responses_exist_for_realization returns False, but if LocalEnsemble.response_configuration contains an SummaryConfig without keys, you get True.

This is also inconsistent with _get_parameter_mask which uses

all(
     path / f"{parameter}.nc").exists()
     for parameter in self.experiment.parameter_configuration
)

[eivindjahren]

I suggest we do the following: #7150 which enforces that SummaryConfig must have keys and changes the name and documentation of the existence checks in local_ensemble.

Also, I think the comment in _filter_response_configuration is incorrect as it dis produce a file, but that file actually ended up the triggering the bug #6974 so now that is a forward model failure. So now that SummaryConfig cannot have empty keys, I removed _filter_response_configuration as it does nothing.

[frode-aarstad]

I think the proposed solution of forcing SummaryConfig to have keys is good and leads to the proposed simplifications to local_ensemble

[eivindjahren]

After looking a bit closer at it, that solution seems to not be possible as it would break with current behavior that
summary keys from HISTORY_OBSERVATION are added to the summary config if missing.

ert/src/ert/config/observations.py

Lines 110 to 111 in 1ff65a5

	if summary_key not in response_config.keys:
	response_config.keys.append(summary_key)

[eivindjahren]

It seems like this issue is less problematic than first anticipated so we can lower the severity of the bug at least.