Add WriteZeroOptionalFields option for Snowflake compatibility

encoding/thrift/encode.go

@@ -278,7 +278,10 @@ encodeFields:
 			}
 		}
 
-		if !f.flags.have(required) && x.IsZero() {
+		// Always write optional fields for regular structs, even if zero-valued.
+		// This ensures fields like NullCount=0 are present in Parquet statistics.
+		// However, unions must skip zero values to maintain exactly one field.
+		if !f.flags.have(required) && x.IsZero() && enc.union {
 			continue encodeFields
 		}
 

format/parquet_test.go

@@ -17,7 +17,11 @@ func TestMarshalUnmarshalSchemaMetadata(t *testing.T) {
 				Name: "hello",
 			},
 		},
-		RowGroups: []format.RowGroup{},
+		RowGroups:                []format.RowGroup{},
+		KeyValueMetadata:         []format.KeyValue{},
+		CreatedBy:                "",
+		ColumnOrders:             []format.ColumnOrder{},
+		FooterSigningKeyMetadata: []byte{},
 	}
 
 	b, err := thrift.Marshal(protocol, metadata)

nullcount_test.go

@@ -0,0 +1,62 @@
+package parquet_test
+
+import (
+	"bytes"
+	"testing"
+
+	"github.com/parquet-go/parquet-go"
+)
+
+// TestNullCountAlwaysWritten verifies that NullCount=0 is always written to
+// the file statistics, ensuring compatibility with strict Parquet consumers
+// like Snowflake that require explicit null counts.
+func TestNullCountAlwaysWritten(t *testing.T) {
+	type Row struct {
+		ID   int64  `parquet:"id"`
+		Name string `parquet:"name"`
+	}
+
+	rows := []Row{
+		{ID: 1, Name: "Alice"},
+		{ID: 2, Name: "Bob"},
+		{ID: 3, Name: "Charlie"},
+	}
+
+	buf := new(bytes.Buffer)
+	writer := parquet.NewGenericWriter[Row](buf)
+
+	if _, err := writer.Write(rows); err != nil {
+		t.Fatalf("failed to write rows: %v", err)
+	}
+
+	if err := writer.Close(); err != nil {
+		t.Fatalf("failed to close writer: %v", err)
+	}
+
+	// Read the file back
+	fileBytes := buf.Bytes()
+	reader := bytes.NewReader(fileBytes)
+	pf, err := parquet.OpenFile(reader, int64(len(fileBytes)))
+	if err != nil {
+		t.Fatalf("failed to open parquet file: %v", err)
+	}
+
+	// Verify that column chunks have NullCount set to 0
+	metadata := pf.Metadata()
+	if len(metadata.RowGroups) == 0 {
+		t.Fatal("No row groups found")
+	}
+
+	for rgIdx, rowGroup := range metadata.RowGroups {
+		for colIdx, col := range rowGroup.Columns {
+			// NullCount should be 0 for all non-nullable columns
+			stats := col.MetaData.Statistics
+			nullCount := stats.NullCount
+			if nullCount != 0 {
+				t.Errorf("RowGroup[%d].Column[%d]: expected NullCount=0, got %d",
+					rgIdx, colIdx, nullCount)
+			}
+			t.Logf("RowGroup[%d].Column[%d]: NullCount=%d ✓", rgIdx, colIdx, nullCount)
+		}
+	}
+}