Fix unaligned metadata flake (#2067)

Rkyv uses aligned access by default, but our metadata flatbufferis
&[u8], therefore has no alignment requirements.

The unaligned feature is marginally slower (unaligned load
instructions), but handles this case.

We can use aligned access again when we move to 8-byte metadata and
store it as a u64 in flatbuffer.

---------

Co-authored-by: Robert Kruszewski <[email protected]>

Author: gatesn

Cargo.toml

@@ -121,7 +121,8 @@ rand_distr = "0.4"
 rayon = "1.10.0"
 regex = "1.11.0"
 reqwest = { version = "0.12.0", features = ["blocking"] }
-rkyv = { version = "0.8", features = ["little_endian", "pointer_width_32", "bytecheck"] }
+# TODO(ngates): remove the unaligned rkyv feature when we move to 8-byte metadata.
+rkyv = { version = "0.8", features = ["little_endian", "pointer_width_32", "bytecheck", "unaligned"] }
 rstest = "0.24"
 serde = "1.0.197"
 serde_json = "1.0.116"

docs/quickstart.rst

@@ -35,7 +35,7 @@ Vortex array:
    >>> parquet = pq.read_table("_static/example.parquet")
    >>> vtx = vortex.array(parquet)
    >>> vtx.nbytes
-   141057
+   141055
 
 Compress
 ^^^^^^^^
@@ -46,7 +46,7 @@ Use :func:`~vortex.encoding.compress` to compress the Vortex array and check the
 
    >>> cvtx = vortex.compress(vtx)
    >>> cvtx.nbytes
-   15888
+   15755
    >>> cvtx.nbytes / vtx.nbytes
    0.11...
 

encodings/fastlanes/goldenfiles/bitpacked.metadata

@@ -168,7 +168,7 @@ mod tests {
             assert_eq!(chunk.encoding().id(), VarBinEncoding::ID);
             assert_eq!(
                 chunk.statistics().get(Stat::UncompressedSizeInBytes),
-                Some(Scalar::from(134357009_u64))
+                Some(Scalar::from(134357007_u64))
             );
         }