fix: Improve logging for spark materialization engine (#172)

Co-authored-by: Bhargav Dodla <[email protected]>

sdk/python/feast/infra/materialization/contrib/spark/spark_materialization_engine.py

@@ -234,15 +234,25 @@ def _map_by_partition(
     spark_serialized_artifacts: _SparkSerializedArtifacts,
 ):
     feature_view, online_store, repo_config = spark_serialized_artifacts.unserialize()
+
+    total_batches = 0
+    total_time = 0.0
+    min_time = float("inf")
+    max_time = float("-inf")
+
+    total_rows = 0
+    min_batch_size = float("inf")
+    max_batch_size = float("-inf")
+
     """Load pandas df to online store"""
     for pdf in iterator:
+        start_time = time.perf_counter()
         pdf_row_count = pdf.shape[0]
-        start_time = time.time()
-        # convert to pyarrow table
         if pdf_row_count == 0:
-            print("INFO!!! Dataframe has 0 records to process")
-            return
+            print("INFO: Dataframe has 0 records to process")
+            break
 
+        # convert to pyarrow table
         table = pyarrow.Table.from_pandas(pdf)
 
         if feature_view.batch_source.field_mapping is not None:
@@ -266,10 +276,89 @@ def _map_by_partition(
             rows_to_write,
             lambda x: None,
         )
-        end_time = time.time()
-        print(
-            f"INFO!!! Processed batch with size {pdf_row_count} in {int((end_time - start_time) * 1000)} milliseconds"
+
+        batch_time = time.perf_counter() - start_time
+
+        (
+            total_batches,
+            total_time,
+            min_time,
+            max_time,
+            total_rows,
+            min_batch_size,
+            max_batch_size,
+        ) = update_exec_stats(
+            total_batches,
+            total_time,
+            min_time,
+            max_time,
+            total_rows,
+            min_batch_size,
+            max_batch_size,
+            batch_time,
+            pdf_row_count,
         )
+
+    if total_batches > 0:
+        print_exec_stats(
+            total_batches,
+            total_time,
+            min_time,
+            max_time,
+            total_rows,
+            min_batch_size,
+            max_batch_size,
+        )
+
     yield pd.DataFrame(
         [pd.Series(range(1, 2))]
     )  # dummy result because mapInPandas needs to return something
+
+
+def update_exec_stats(
+    total_batches,
+    total_time,
+    min_time,
+    max_time,
+    total_rows,
+    min_batch_size,
+    max_batch_size,
+    batch_time,
+    current_batch_size,
+):
+    total_batches += 1
+    total_time += batch_time
+    min_time = min(min_time, batch_time)
+    max_time = max(max_time, batch_time)
+
+    total_rows += current_batch_size
+    min_batch_size = min(min_batch_size, current_batch_size)
+    max_batch_size = max(max_batch_size, current_batch_size)
+
+    return (
+        total_batches,
+        total_time,
+        min_time,
+        max_time,
+        total_rows,
+        min_batch_size,
+        max_batch_size,
+    )
+
+
+def print_exec_stats(
+    total_batches,
+    total_time,
+    min_time,
+    max_time,
+    total_rows,
+    min_batch_size,
+    max_batch_size,
+):
+    # TODO: Investigate why the logger is not working in Spark Executors
+    avg_time = total_time / total_batches
+    avg_batch_size = total_rows / total_batches
+    print(
+        f"Time - Total: {total_time:.6f}s, Avg: {avg_time:.6f}s, Min: {min_time:.6f}s, Max: {max_time:.6f}s | "
+        f"Batch Size - Total: {total_rows}, Avg: {avg_batch_size:.2f}, Min: {min_batch_size}, Max: {max_batch_size}"
+    )