From a0c5612aee26173b0f993d185c386c480348aa30 Mon Sep 17 00:00:00 2001 From: Vivek Date: Thu, 17 Sep 2026 10:59:41 -0700 Subject: [PATCH] feat(rust): make MetadataEraserExec public (#4213) Export `MetadataEraserExec` and its constructor from `lancedb::table::datafusion`. Engines that serialise a physical plan containing a LanceDB scan have to rebuild the operator outside this crate, which a private type makes impossible. --- rust/lancedb/src/table/datafusion.rs | 21 +++++++++++++++++---- 1 file changed, 17 insertions(+), 4 deletions(-) diff --git a/rust/lancedb/src/table/datafusion.rs b/rust/lancedb/src/table/datafusion.rs index d60bce7dc..85366910b 100644 --- a/rust/lancedb/src/table/datafusion.rs +++ b/rust/lancedb/src/table/datafusion.rs @@ -33,11 +33,23 @@ use crate::{ use arrow_schema::{DataType, Field}; use lance_index::scalar::FullTextSearchQuery; -/// Datafusion attempts to maintain batch metadata +/// An execution plan that erases Arrow schema-level metadata from its input's batches. /// -/// This is needless and it triggers bugs in DF. This operator erases metadata from the batches. +/// DataFusion attempts to maintain batch metadata. This is needless and it triggers bugs in +/// DF, so [`BaseTableAdapter::scan`] wraps every scan it produces in one of these. +/// +/// ``` +/// use std::sync::Arc; +/// +/// use datafusion_physical_plan::ExecutionPlan; +/// use lancedb::table::datafusion::MetadataEraserExec; +/// +/// # fn erase_metadata(scan: Arc) -> Arc { +/// Arc::new(MetadataEraserExec::new(scan)) +/// # } +/// ``` #[derive(Debug)] -struct MetadataEraserExec { +pub struct MetadataEraserExec { input: Arc, schema: Arc, properties: Arc, @@ -62,7 +74,8 @@ impl MetadataEraserExec { ) } - fn new(input: Arc) -> Self { + /// Wrap `input` in an operator that strips schema-level metadata from its batches. + pub fn new(input: Arc) -> Self { let schema = Arc::new( input .schema()