docs: changelog entry for zero-copy Arrow boundary

Author: cpsievert

CHANGELOG.md

@@ -1,5 +1,11 @@
 # Changelog
 
+## UNRELEASED
+
+### Changed
+
+- Data now crosses the Python↔Rust boundary via the zero-copy Arrow C Data Interface instead of Arrow IPC serialization, reducing memory copies and peak RSS for large results (on a 1M-row × 50-col register+read microbenchmark: ~5.7× faster, ~1 GB lower peak RSS). Set `GGSQL_ARROW_IPC=1` to restore the previous IPC behavior. The minimum pyarrow version is now 15.
+
 ## 0.5.2
 
 ### Added

src/lib.rs

@@ -272,14 +272,12 @@ fn py_to_df_stream(py: Python<'_>, obj: &Bound<'_, PyAny>) -> PyResult<DataFrame
         ))
     })?;
 
-    let batches: Vec<_> = reader
-        .collect::<Result<Vec<_>, _>>()
-        .map_err(|e| {
-            PyErr::new::<pyo3::exceptions::PyValueError, _>(format!(
-                "Failed to read RecordBatch from Arrow C stream: {}",
-                e
-            ))
-        })?;
+    let batches: Vec<_> = reader.collect::<Result<Vec<_>, _>>().map_err(|e| {
+        PyErr::new::<pyo3::exceptions::PyValueError, _>(format!(
+            "Failed to read RecordBatch from Arrow C stream: {}",
+            e
+        ))
+    })?;
 
     if batches.is_empty() {
         return Ok(DataFrame::empty());