speeding up Parquet_output function by vectorizing

src/troute-nwm/src/nwm_routing/output.py

@@ -64,28 +64,53 @@ def _parquet_output_format_converter(df, start_datetime, dt, configuration, pref
     --------
     - timeseries_df (DataFrame): Converted timeseries data frame
     '''
-
-    df.index.name = 'location_id'
-    df.reset_index(inplace=True)
-    timeseries_df = df.melt(id_vars=['location_id'], var_name='var')
-    timeseries_df['var'] = timeseries_df['var'].astype('string')
-    timeseries_df[['timestep', 'variable']] = timeseries_df['var'].str.strip("()").str.split(",", n=1, expand=True)
-    timeseries_df['variable'] = timeseries_df['variable'].str.strip().str.replace("'", "")
-    timeseries_df['timestep'] = timeseries_df['timestep'].astype('int')
-    timeseries_df['value_time'] = (start_datetime + pd.to_timedelta(timeseries_df['timestep'] * dt, unit='s'))
     variable_to_name_map = {"q": "streamflow", "d": "depth", "v": "velocity"}
-    timeseries_df["variable_name"] = timeseries_df["variable"].map(variable_to_name_map)
-    timeseries_df.drop(['var', 'timestep', 'variable'], axis=1, inplace=True)
-    timeseries_df['configuration'] = configuration
     variable_to_units_map = {"streamflow": "m3/s", "velocity": "m/s", "depth": "m"}
-    timeseries_df['units'] = timeseries_df['variable_name'].map(variable_to_units_map)
-    timeseries_df['reference_time'] = start_datetime.date()
-    timeseries_df['location_id'] = timeseries_df['location_id'].astype('string')
-    timeseries_df['location_id'] = prefix_ids + '-' + timeseries_df['location_id']
-    timeseries_df['value'] = timeseries_df['value'].astype('double')
-    timeseries_df['reference_time'] = timeseries_df['reference_time'].astype('datetime64[us]')
-    timeseries_df['value_time'] = timeseries_df['value_time'].astype('datetime64[us]')
 
+    # Prepare the location_id with prefix
+    df.index.name = 'location_id'
+    df.reset_index(inplace=True)
+    location_ids = prefix_ids + '-' + df['location_id'].astype(str)
+
+    # Flatten the dataframe using NumPy
+    num_locations = df.shape[0]
+    num_time_variables = df.shape[1] - 1
+    num_records = num_locations * num_time_variables
+
+    # Prepare timestep and variable arrays
+    times = df.columns[1:]
+    timesteps = np.array([t[0] for t in times], dtype=int)
+    variables = np.array([t[1] for t in times])
+
+    # Preallocate arrays
+    location_ids_repeated = np.tile(location_ids, num_time_variables)
+    value_time = np.empty(num_records, dtype='datetime64[us]')
+    variable_names = np.empty(num_records, dtype=object)
+    units = np.empty(num_records, dtype=object)
+    values = np.empty(num_records, dtype=float)
+
+    # Calculate value_time, variable_names, units, and values in a vectorized manner
+    for i in range(num_time_variables):
+        start_idx = i * num_locations
+        end_idx = start_idx + num_locations
+        value_time[start_idx:end_idx] = start_datetime + pd.to_timedelta(timesteps[i] * dt, unit='s')
+        variable_name = variable_to_name_map[variables[i]]
+        unit = variable_to_units_map[variable_name]
+        variable_names[start_idx:end_idx] = variable_name
+        units[start_idx:end_idx] = unit
+        values[start_idx:end_idx] = df.iloc[:, i + 1].values
+
+    # Create the resulting DataFrame
+    timeseries_df = pd.DataFrame({
+        'location_id': location_ids_repeated,
+        'value': values,
+        'value_time': value_time,
+        'variable_name': variable_names,
+        'units': units,
+        'reference_time': start_datetime.date(),
+        'configuration': configuration
+    })
+
     return timeseries_df