fix: Use bytemuck in slice reinterpret for Parquet ArrayChunks (#17700)

pola-rs · Jul 18, 2024 · 235cad3 · 235cad3
1 parent 05d4703
commit 235cad3
Show file tree

Hide file tree

Showing 4 changed files with 4 additions and 3 deletions.
diff --git a/Cargo.lock b/Cargo.lock
diff --git a/crates/polars-parquet/Cargo.toml b/crates/polars-parquet/Cargo.toml
@@ -16,6 +16,7 @@ description = "Apache Parquet I/O operations for Polars"
 ahash = { workspace = true }
 arrow = { workspace = true, features = ["io_ipc"] }
 base64 = { workspace = true }
+bytemuck = { workspace = true }
 ethnum = { workspace = true }
 fallible-streaming-iterator = { workspace = true, optional = true }
 futures = { workspace = true, optional = true }

diff --git a/crates/polars-parquet/src/arrow/read/deserialize/utils/array_chunks.rs b/crates/polars-parquet/src/arrow/read/deserialize/utils/array_chunks.rs
@@ -20,9 +20,7 @@ impl<'a, P: ParquetNativeType> ArrayChunks<'a, P> {
             return None;
         }
 
-        // SAFETY:
-        // We know that that the alignment, size and provenance are the same.
-        let bytes = unsafe { std::mem::transmute::<&[u8], &[P::Bytes]>(bytes) };
+        let bytes = bytemuck::cast_slice(bytes);
 
         Some(Self { bytes })
     }

diff --git a/crates/polars-parquet/src/parquet/types.rs b/crates/polars-parquet/src/parquet/types.rs
@@ -3,6 +3,7 @@ use crate::parquet::schema::types::PhysicalType;
 /// A physical native representation of a Parquet fixed-sized type.
 pub trait NativeType: std::fmt::Debug + Send + Sync + 'static + Copy + Clone {
     type Bytes: AsRef<[u8]>
+        + bytemuck::Pod
         + for<'a> TryFrom<&'a [u8], Error = std::array::TryFromSliceError>
         + std::fmt::Debug
         + Clone