diff --git a/datafusion/core/src/datasource/physical_plan/parquet/mod.rs b/datafusion/core/src/datasource/physical_plan/parquet/mod.rs
index e8e186e5c45fe..1a797dbd29ae8 100644
--- a/datafusion/core/src/datasource/physical_plan/parquet/mod.rs
+++ b/datafusion/core/src/datasource/physical_plan/parquet/mod.rs
@@ -579,9 +579,33 @@ impl FileOpener for ParquetOpener {
                 builder = builder.with_limit(limit)
             }
 
+            let total_uncompressed_size = builder
+                .metadata()
+                .row_groups()
+                .iter()
+                .map(|rg| rg
+                        .columns()
+                        .iter()
+                        .enumerate()
+                        .filter(|(idx, _)| mask.leaf_included(*idx))
+                        .map(|(_, col)| col.uncompressed_size())
+                        .sum::<i64>())
+                .sum::<i64>();
+            let total_num_rows = builder
+                .metadata()
+                .row_groups()
+                .iter()
+                .map(|rg| rg.num_rows())
+                .sum::<i64>();
+            let adaptive_batch_mem_size = 16777216; // blaze: use 16MB batch mem size as default
+            let adaptive_batch_size = batch_size
+                .min((adaptive_batch_mem_size * total_num_rows / total_uncompressed_size.max(1)) as usize)
+                .max(1);
+            log::info!("executing parquet scan with adaptive batch size: {adaptive_batch_size}");
+
             let stream = builder
                 .with_projection(mask)
-                .with_batch_size(batch_size)
+                .with_batch_size(adaptive_batch_size)
                 .with_row_groups(row_groups)
                 .build()?;