Store each response key as a column for responses

src/ert/analysis/_es_update.py

@@ -165,10 +165,19 @@ def _get_observations_and_responses(
             response_type, realizations=tuple(iens_active_index)
         )
 
+        # Make realizations into columns,
+        # and add response_key column
+        unpivoted = responses_for_type.unpivot(
+            on=response_cls.keys,
+            variable_name="response_key",
+            value_name="values",
+            index=["realization", *response_cls.primary_key],
+        )
+
         # Note that if there are duplicate entries for one
         # response at one index, they are aggregated together
         # with "mean" by default
-        pivoted = responses_for_type.pivot(
+        pivoted = unpivoted.pivot(
             on="realization",
             index=["response_key", *response_cls.primary_key],
             aggregate_function="mean",

src/ert/config/gen_data_config.py

@@ -187,6 +187,13 @@ def _read_file(filename: Path, report_step: int) -> polars.DataFrame:
                 )
 
         combined = polars.concat(datasets_per_name)
+
+        if combined.is_empty():
+            raise InvalidResponseFile(
+                f"No data found within response files: {', '.join(self.input_files)}"
+            )
+
+        combined = combined.pivot(on="response_key", index=self.primary_key)
         return combined
 
     def get_args_for_key(self, key: str) -> Tuple[Optional[str], Optional[List[int]]]:

src/ert/config/summary_config.py

@@ -59,6 +59,9 @@ def read_from_file(self, run_path: str, iens: int) -> polars.DataFrame:
             }
         )
         df = df.explode("values", "time")
+        df = df.pivot(
+            on="response_key", index=self.primary_key, aggregate_function="mean"
+        )
         return df
 
     @property

src/ert/dark_storage/common.py

@@ -130,17 +130,15 @@ def data_for_key(
             if summary_data.is_empty():
                 return pd.DataFrame()
 
-            df = (
-                summary_data.rename({"time": "Date", "realization": "Realization"})
-                .drop("response_key")
+            data = (
+                summary_data.pivot(
+                    on="time", values=response_key, aggregate_function="mean"
+                )
+                .rename({"realization": "Realization"})
                 .to_pandas()
             )
-            df = df.set_index(["Date", "Realization"])
-            # This performs the same aggragation by mean of duplicate values
-            # as in ert/analysis/_es_update.py
-            df = df.groupby(["Date", "Realization"]).mean()
-            data = df.unstack(level="Date")
-            data.columns = data.columns.droplevel(0)
+            data.set_index("Realization", inplace=True)
+            data.columns = data.columns.astype("datetime64[ms]")
             try:
                 return data.astype(float)
             except ValueError:
@@ -162,8 +160,8 @@ def data_for_key(
 
             try:
                 vals = data.filter(polars.col("report_step").eq(report_step))
-                pivoted = vals.drop("response_key", "report_step").pivot(
-                    on="index", values="values"
+                pivoted = vals.drop(["report_step"]).pivot(
+                    on=["index"], values=response_key, aggregate_function="mean"
                 )
                 data = pivoted.to_pandas().set_index("realization")
                 data.columns = data.columns.astype(int)

src/ert/data/_measured_data.py

@@ -120,10 +120,19 @@ def _get_data(
                     f"No response loaded for observation type: {response_type}"
                 )
 
+            # Make realizations into columns,
+            # and add response_key column
+            unpivoted = responses_for_type.unpivot(
+                on=response_cls.keys,
+                variable_name="response_key",
+                value_name="values",
+                index=["realization", *response_cls.primary_key],
+            )
+
             # Note that if there are duplicate entries for one
             # response at one index, they are aggregated together
             # with "mean" by default
-            pivoted = responses_for_type.pivot(
+            pivoted = unpivoted.pivot(
                 on="realization",
                 index=["response_key", *response_cls.primary_key],
                 aggregate_function="mean",

src/ert/simulator/batch_simulator_context.py

@@ -280,7 +280,7 @@ def results(self) -> List[Optional[Dict[str, "npt.NDArray[np.float64]"]]]:
             d = {}
             for key in self.result_keys:
                 data = self.ensemble.load_responses(key, (sim_id,))
-                d[key] = data["values"].to_numpy()
+                d[key] = data[key].to_numpy()
             res.append(d)
 
         return res

src/ert/storage/local_ensemble.py

@@ -661,15 +661,21 @@ def load_responses(self, key: str, realizations: Tuple[int]) -> polars.DataFrame
             response_type = self.experiment.response_key_to_response_type[key]
             select_key = True
 
+        response_config = self.experiment.response_configuration[response_type]
+
         loaded = []
         for realization in realizations:
             input_path = self._realization_dir(realization) / f"{response_type}.parquet"
             if not input_path.exists():
                 raise KeyError(f"No response for key {key}, realization: {realization}")
-            df = polars.read_parquet(input_path)
+            lazy_df = polars.scan_parquet(input_path)
 
             if select_key:
-                df = df.filter(polars.col("response_key") == key)
+                df = lazy_df.select(
+                    ["realization", *response_config.primary_key, key]
+                ).collect()
+            else:
+                df = lazy_df.collect()
 
             loaded.append(df)
 
@@ -707,9 +713,6 @@ def load_all_summary_data(
 
         except (ValueError, KeyError):
             return pd.DataFrame()
-        df_pl = df_pl.pivot(
-            on="response_key", index=["realization", "time"], sort_columns=True
-        )
         df_pl = df_pl.rename({"time": "Date", "realization": "Realization"})
 
         df_pandas = (
@@ -845,11 +848,16 @@ def save_response(
         data : polars DataFrame
             polars DataFrame to save.
         """
+        response_config = self.experiment.response_configuration[response_type]
 
-        if "values" not in data.columns:
+        num_response_columns = (
+            len(data.columns)
+            - len(response_config.primary_key)
+            - (1 if "realization" in data.columns else 0)
+        )
+        if num_response_columns <= 0:
             raise ValueError(
-                f"Dataset for response group '{response_type}' "
-                f"must contain a 'values' variable"
+                f"Dataset for response type '{response_type}' must contain values for at least one response key"
             )
 
         if len(data) == 0:
@@ -873,7 +881,7 @@ def save_response(
         )
 
         if not self.experiment._has_finalized_response_keys(response_type):
-            response_keys = data["response_key"].unique().to_list()
+            response_keys = data.columns[(len(response_config.primary_key) + 1) :]
             self.experiment._update_response_keys(response_type, response_keys)
 
     def calculate_std_dev_for_parameter(self, parameter_group: str) -> xr.Dataset:

src/ert/storage/local_storage.py

@@ -46,7 +46,7 @@
 
 logger = logging.getLogger(__name__)
 
-_LOCAL_STORAGE_VERSION = 8
+_LOCAL_STORAGE_VERSION = 9
 
 
 class _Migrations(BaseModel):
@@ -472,6 +472,7 @@ def _migrate(self, version: int) -> None:
             to6,
             to7,
             to8,
+            to9,
         )
 
         try:
@@ -516,7 +517,7 @@ def _migrate(self, version: int) -> None:
 
             elif version < _LOCAL_STORAGE_VERSION:
                 migrations = list(
-                    enumerate([to2, to3, to4, to5, to6, to7, to8], start=1)
+                    enumerate([to2, to3, to4, to5, to6, to7, to8, to9], start=1)
                 )
                 for from_version, migration in migrations[version - 1 :]:
                     print(f"* Updating storage to version: {from_version+1}")

src/ert/storage/migration/to8.py

@@ -78,7 +78,10 @@ def _migrate_responses_from_netcdf_to_parquet(path: Path) -> None:
                         )
 
                         pandas_df = gen_data_ds.to_dataframe().dropna()
-                        polars_df = polars.from_pandas(pandas_df.reset_index())
+                        polars_df = polars.from_pandas(
+                            pandas_df.reset_index(),
+                            schema_overrides={"values": polars.Float32},
+                        )
                         polars_df = polars_df.rename({"name": "response_key"})
 
                         if "time" in polars_df:

src/ert/storage/migration/to9.py

@@ -0,0 +1,43 @@
+import json
+import os
+from pathlib import Path
+
+import polars
+
+info = "Make response datasets have one column per response"
+
+
+def _migrate_responses_to_one_col_per_response(path: Path) -> None:
+    for experiment in path.glob("experiments/*"):
+        ensembles = path.glob("ensembles/*")
+
+        experiment_id = None
+        with open(experiment / "index.json", encoding="utf-8") as f:
+            exp_index = json.load(f)
+            experiment_id = exp_index["id"]
+
+        for ens in ensembles:
+            with open(ens / "index.json", encoding="utf-8") as f:
+                ens_file = json.load(f)
+                if ens_file["experiment_id"] != experiment_id:
+                    continue
+
+            real_dirs = [*ens.glob("realization-*")]
+
+            for real_dir in real_dirs:
+                for df_name, columns in [
+                    ("gen_data", ["report_step", "index"]),
+                    ("summary", ["time"]),
+                ]:
+                    if (real_dir / f"{df_name}.parquet").exists():
+                        df = polars.read_parquet(real_dir / f"{df_name}.parquet")
+                        pivoted = df.pivot(
+                            on="response_key", index=["realization", *columns]
+                        )
+
+                        os.remove(real_dir / f"{df_name}.parquet")
+                        pivoted.write_parquet(real_dir / f"{df_name}.parquet")
+
+
+def migrate(path: Path) -> None:
+    _migrate_responses_to_one_col_per_response(path)

tests/ert/performance_tests/test_analysis.py

@@ -139,7 +139,7 @@ def g(X):
                     "index": range(len(Y[:, iens])),
                     "values": Y[:, iens],
                 }
-            ),
+            ).pivot(on="response_key", index=["report_step", "index"]),
             iens,
         )
 

tests/ert/performance_tests/test_dark_storage_performance.py

@@ -141,7 +141,7 @@ def test_direct_dark_performance(
             if ensemble_json.userdata["name"] == "default":
                 ensemble_id_default = ensemble_id
 
-        benchmark(function, storage, ensemble_id_default, key, template_config)
+        function(storage, ensemble_id_default, key, template_config)
 
 
 @pytest.mark.parametrize(
@@ -177,4 +177,4 @@ def test_direct_dark_performance_with_storage(
             if ensemble_json.userdata["name"] == "default":
                 ensemble_id_default = ensemble_id
 
-        benchmark(function, storage, ensemble_id_default, key, template_config)
+        function(storage, ensemble_id_default, key, template_config)

tests/ert/performance_tests/test_memory_usage.py

@@ -89,16 +89,21 @@ def fill_storage_with_data(poly_template: Path, ert_config: ErtConfig) -> None:
             gendatas = []
             gen_obs = ert_config.observations["gen_data"]
             for response_key, df in gen_obs.group_by("response_key"):
-                gendata_df = make_gen_data(df["index"].max() + 1)
+                gendata_df = make_gen_data(response_key[0], df["index"].max() + 1)
                 gendata_df = gendata_df.insert_column(
                     0,
                     polars.Series(np.full(len(gendata_df), response_key)).alias(
                         "response_key"
                     ),
                 )
-                gendatas.append(gendata_df)
+                gendatas.append((response_key, gendata_df))
 
-            source.save_response("gen_data", polars.concat(gendatas), real)
+            gendatas.sort(key=lambda info: info[0])
+
+            wide_gendatas = polars.concat([df for _, df in gendatas]).pivot(
+                on="response_key", index=["report_step", "index"]
+            )
+            source.save_response("gen_data", wide_gendatas, real)
 
             obs_time_list = ens_config.refcase.all_dates
 
@@ -123,13 +128,15 @@ def fill_storage_with_data(poly_template: Path, ert_config: ErtConfig) -> None:
         )
 
 
-def make_gen_data(obs: int, min_val: float = 0, max_val: float = 5) -> polars.DataFrame:
+def make_gen_data(
+    response_key: str, obs: int, min_val: float = 0, max_val: float = 5
+) -> polars.DataFrame:
     data = np.random.default_rng().uniform(min_val, max_val, obs)
     return polars.DataFrame(
         {
             "report_step": polars.Series(np.full(len(data), 0), dtype=polars.UInt16),
             "index": polars.Series(range(len(data)), dtype=polars.UInt16),
-            "values": data,
+            "values": polars.Series(data, dtype=polars.Float32),
         }
     )
 
@@ -148,9 +155,9 @@ def make_summary_data(
             "time": polars.Series(
                 np.tile(dates, len(obs_keys)).tolist()
             ).dt.cast_time_unit("ms"),
-            "values": data,
+            "values": polars.Series(data, dtype=polars.Float32),
         }
-    )
+    ).pivot(on="response_key", index="time")
 
 
 @pytest.mark.limit_memory("130 MB")

tests/ert/unit_tests/analysis/test_es_update.py

@@ -415,7 +415,7 @@ def test_smoother_snapshot_alpha(
                     "index": polars.Series(range(len(data)), dtype=polars.UInt16),
                     "values": data,
                 }
-            ),
+            ).pivot(on="response_key", index=["report_step", "index"]),
             iens,
         )
     posterior_storage = storage.create_ensemble(
@@ -719,7 +719,7 @@ def test_gen_data_obs_data_mismatch(storage, uniform_parameter):
                     "index": polars.Series(range(len(data)), dtype=polars.UInt16),
                     "values": polars.Series(data, dtype=polars.Float32),
                 }
-            ),
+            ).pivot(on="response_key", index=["report_step", "index"]),
             iens,
         )
     posterior_ens = storage.create_ensemble(
@@ -782,7 +782,7 @@ def test_gen_data_missing(storage, uniform_parameter, obs):
                     "index": polars.Series(range(len(data)), dtype=polars.UInt16),
                     "values": polars.Series(data, dtype=polars.Float32),
                 }
-            ),
+            ).pivot(on="response_key", index=["report_step", "index"]),
             iens,
         )
     posterior_ens = storage.create_ensemble(
@@ -876,7 +876,7 @@ def test_update_subset_parameters(storage, uniform_parameter, obs):
                     "index": polars.Series(range(len(data)), dtype=polars.UInt16),
                     "values": polars.Series(data, dtype=polars.Float32),
                 }
-            ),
+            ).pivot(on="response_key", index=["report_step", "index"]),
             iens,
         )
     posterior_ens = storage.create_ensemble(

tests/ert/unit_tests/dark_storage/test_common.py

@@ -95,7 +95,7 @@ def test_data_for_key_returns_empty_gen_data_config(tmp_path):
                     "index": polars.Series([0], dtype=polars.UInt16),
                     "values": polars.Series([0.0], dtype=polars.Float32),
                 }
-            ),
+            ).pivot(on="response_key", index=["report_step", "index"]),
             0,
         )
         ensemble.refresh_ensemble_state()

tests/ert/unit_tests/gui/tools/plot/test_plot_api.py

@@ -250,7 +250,7 @@ def test_plot_api_big_summary_memory_usage(
             "time": dates_df,
             "values": values_df,
         }
-    )
+    ).pivot(on=["response_key"], index="time")
 
     experiment = storage.create_experiment(
         parameters=[],
@@ -396,7 +396,7 @@ def test_plot_api_handles_urlescape(api_and_storage):
             "values": [polars.Series([1.0], dtype=polars.Float32)],
         }
     )
-    df = df.explode("values", "time")
+    df = df.explode("values", "time").pivot(on="response_key", index="time")
     ensemble.save_response(
         "summary",
         df,