diff --git a/rust/lance-core/src/deepsize.rs b/rust/lance-core/src/deepsize.rs
index b6c145bb504..350c52b146c 100644
--- a/rust/lance-core/src/deepsize.rs
+++ b/rust/lance-core/src/deepsize.rs
@@ -87,6 +87,16 @@ impl DeepSizeOf for String {
     }
 }
 
+impl DeepSizeOf for bytes::Bytes {
+    fn deep_size_of_children(&self, context: &mut Context) -> usize {
+        if context.mark_seen(self.as_ptr() as usize) {
+            self.len()
+        } else {
+            0
+        }
+    }
+}
+
 impl DeepSizeOf for AtomicU64 {
     fn deep_size_of_children(&self, _context: &mut Context) -> usize {
         0
diff --git a/rust/lance-encoding/src/decoder.rs b/rust/lance-encoding/src/decoder.rs
index a30d5ed93a9..bb46e725098 100644
--- a/rust/lance-encoding/src/decoder.rs
+++ b/rust/lance-encoding/src/decoder.rs
@@ -225,13 +225,14 @@ use futures::future::{BoxFuture, MaybeDone, maybe_done};
 use futures::stream::{self, BoxStream};
 use futures::{FutureExt, StreamExt};
 use lance_arrow::DataTypeExt;
-use lance_core::cache::LanceCache;
+use lance_core::cache::{Context, DeepSizeOf, LanceCache};
 use lance_core::datatypes::{
     BLOB_DESC_LANCE_FIELD, Field, Schema, validate_fixed_size_list_dimensions,
 };
 use lance_core::utils::futures::{FinallyStreamExt, StreamOnDropExt};
 use lance_core::utils::parse::parse_env_as_bool;
 use log::{debug, trace, warn};
+use prost::Message;
 use tokio::sync::mpsc::error::SendError;
 use tokio::sync::mpsc::{self, unbounded_channel};
 
@@ -299,6 +300,15 @@ pub enum PageEncoding {
     Structural(pb21::PageLayout),
 }
 
+impl DeepSizeOf for PageEncoding {
+    fn deep_size_of_children(&self, _context: &mut Context) -> usize {
+        match self {
+            Self::Legacy(encoding) => encoding.encoded_len() * 4,
+            Self::Structural(encoding) => encoding.encoded_len() * 4,
+        }
+    }
+}
+
 impl PageEncoding {
     pub fn as_legacy(&self) -> &pb::ArrayEncoding {
         match self {
@@ -336,6 +346,13 @@ pub struct PageInfo {
     pub buffer_offsets_and_sizes: Arc<[(u64, u64)]>,
 }
 
+impl DeepSizeOf for PageInfo {
+    fn deep_size_of_children(&self, context: &mut Context) -> usize {
+        self.encoding.deep_size_of_children(context)
+            + self.buffer_offsets_and_sizes.deep_size_of_children(context)
+    }
+}
+
 /// Metadata describing a column in a file
 ///
 /// This is typically created by reading the metadata section of a Lance file
@@ -350,6 +367,14 @@ pub struct ColumnInfo {
     pub encoding: pb::ColumnEncoding,
 }
 
+impl DeepSizeOf for ColumnInfo {
+    fn deep_size_of_children(&self, context: &mut Context) -> usize {
+        self.page_infos.deep_size_of_children(context)
+            + self.buffer_offsets_and_sizes.deep_size_of_children(context)
+            + self.encoding.encoded_len() * 4
+    }
+}
+
 impl ColumnInfo {
     /// Create a new instance
     pub fn new(
diff --git a/rust/lance-encoding/src/version.rs b/rust/lance-encoding/src/version.rs
index 41640d0bf2d..cd8f09b011b 100644
--- a/rust/lance-encoding/src/version.rs
+++ b/rust/lance-encoding/src/version.rs
@@ -5,6 +5,7 @@ use std::str::FromStr;
 
 use lance_arrow::DataTypeExt;
 use lance_core::datatypes::Field;
+use lance_core::deepsize::{Context, DeepSizeOf};
 use lance_core::{Error, Result};
 
 pub const LEGACY_FORMAT_VERSION: &str = "0.1";
@@ -38,6 +39,12 @@ pub enum LanceFileVersion {
     V2_3,
 }
 
+impl DeepSizeOf for LanceFileVersion {
+    fn deep_size_of_children(&self, _context: &mut Context) -> usize {
+        0
+    }
+}
+
 impl LanceFileVersion {
     /// Convert Stable or Next to the actual version
     pub fn resolve(&self) -> Self {
diff --git a/rust/lance-file/src/reader.rs b/rust/lance-file/src/reader.rs
index c454f73819e..d5355c089b2 100644
--- a/rust/lance-file/src/reader.rs
+++ b/rust/lance-file/src/reader.rs
@@ -2,6 +2,7 @@
 // SPDX-FileCopyrightText: Copyright The Lance Authors
 
 use std::{
+    borrow::Cow,
     collections::{BTreeMap, BTreeSet},
     io::Cursor,
     ops::Range,
@@ -31,7 +32,7 @@ use prost::{Message, Name};
 
 use lance_core::{
     Error, Result,
-    cache::LanceCache,
+    cache::{CacheKey, LanceCache},
     datatypes::{Field, Schema},
 };
 use lance_encoding::format::pb as pbenc;
@@ -127,6 +128,14 @@ impl CachedFileMetadata {
     }
 }
 
+fn column_metadata_deep_size(column_metadatas: &[pbfile::ColumnMetadata]) -> usize {
+    column_metadatas
+        .iter()
+        .map(|cm| cm.encoded_len() * 4)
+        .sum::<usize>()
+        + std::mem::size_of_val(column_metadatas)
+}
+
 impl DeepSizeOf for CachedFileMetadata {
     fn deep_size_of_children(&self, context: &mut Context) -> usize {
         let schema_size = self.file_schema.deep_size_of_children(context);
@@ -142,49 +151,15 @@ impl DeepSizeOf for CachedFileMetadata {
         // as a proxy for in-memory size. The decoded representation is typically
         // several times larger than the wire format due to heap-allocated
         // repeated/string/bytes fields, so we apply a 4x multiplier.
-        let column_metadatas_size: usize = self
-            .column_metadatas
-            .iter()
-            .map(|cm| cm.encoded_len() * 4)
-            .sum::<usize>()
-            + std::mem::size_of_val(self.column_metadatas.as_slice());
+        let column_metadatas_size = column_metadata_deep_size(self.column_metadatas.as_slice());
 
         // column_infos is Vec<Arc<ColumnInfo>>. Each ColumnInfo contains
         // page_infos (with protobuf PageEncoding), buffer offsets, and a
         // column-level ColumnEncoding protobuf.
-        let column_infos_size: usize = self
-            .column_infos
-            .iter()
-            .map(|ci| {
-                let pages_size: usize = ci
-                    .page_infos
-                    .iter()
-                    .map(|pi| {
-                        let enc_size = match &pi.encoding {
-                            lance_encoding::decoder::PageEncoding::Legacy(e) => e.encoded_len() * 4,
-                            lance_encoding::decoder::PageEncoding::Structural(e) => {
-                                e.encoded_len() * 4
-                            }
-                        };
-                        enc_size
-                            + std::mem::size_of_val(pi.buffer_offsets_and_sizes.as_ref())
-                            + std::mem::size_of::<u64>() * 2 // num_rows + priority
-                    })
-                    .sum();
-                pages_size
-                    + std::mem::size_of_val(ci.buffer_offsets_and_sizes.as_ref())
-                    + ci.encoding.encoded_len() * 4
-                    + std::mem::size_of::<u32>() // index
-                    + std::mem::size_of::<usize>() * 2 // Arc overhead
-            })
-            .sum();
+        let column_infos_size = self.column_infos.deep_size_of_children(context);
 
         // Global buffer bytes retained for zero-IO reads (copied out of the tail).
-        let retained_buffers_size: usize = self
-            .retained_global_buffers
-            .values()
-            .map(|buf| buf.len())
-            .sum();
+        let retained_buffers_size = self.retained_global_buffers.deep_size_of_children(context);
 
         schema_size
             + buffers_size
@@ -194,6 +169,65 @@ impl DeepSizeOf for CachedFileMetadata {
     }
 }
 
+/// Lightweight file metadata used to locate per-column metadata on demand.
+///
+/// This contains the file-level schema, row count, global buffer descriptors,
+/// and column metadata offset table. Unlike [`CachedFileMetadata`], it does not
+/// hold decoded metadata for every column.
+#[derive(Debug, DeepSizeOf)]
+pub struct FileMetadataIndex {
+    file_schema: Arc<Schema>,
+    num_rows: u64,
+    file_buffers: Vec<BufferDescriptor>,
+    column_metadata_offsets: Arc<[(u64, u64)]>,
+    num_columns: u32,
+    version: LanceFileVersion,
+    file_size_bytes: u64,
+    retained_global_buffers: BTreeMap<u32, Bytes>,
+}
+
+impl FileMetadataIndex {
+    /// Returns the total size of the file in bytes.
+    pub fn file_size(&self) -> u64 {
+        self.file_size_bytes
+    }
+
+    /// Returns the number of physical columns in the file.
+    pub fn num_columns(&self) -> u32 {
+        self.num_columns
+    }
+}
+
+#[derive(Debug)]
+struct CachedColumnMetadata {
+    column_metadata: pbfile::ColumnMetadata,
+    column_info: Arc<ColumnInfo>,
+}
+
+impl DeepSizeOf for CachedColumnMetadata {
+    fn deep_size_of_children(&self, context: &mut Context) -> usize {
+        column_metadata_deep_size(std::slice::from_ref(&self.column_metadata))
+            + self.column_info.deep_size_of_children(context)
+    }
+}
+
+#[derive(Debug, Clone)]
+struct ColumnMetadataCacheKey {
+    column_index: u32,
+}
+
+impl CacheKey for ColumnMetadataCacheKey {
+    type ValueType = CachedColumnMetadata;
+
+    fn key(&self) -> Cow<'_, str> {
+        Cow::Owned(format!("column_metadata/{}", self.column_index))
+    }
+
+    fn type_name() -> &'static str {
+        "ColumnMetadata"
+    }
+}
+
 impl CachedFileMetadata {
     pub fn version(&self) -> LanceFileVersion {
         match (self.major_version, self.minor_version) {
@@ -430,16 +464,45 @@ impl Default for FileReaderOptions {
 }
 
 #[derive(Debug, Clone)]
-pub struct FileReader {
+struct PreparedProjection {
+    column_infos: Vec<Arc<ColumnInfo>>,
+    decoder_projection: ReaderProjection,
+}
+
+#[derive(Debug, Clone)]
+enum FileMetadataProvider {
+    Full(Arc<CachedFileMetadata>),
+    Indexed(Arc<FileMetadataIndex>),
+}
+
+#[derive(Debug, Clone)]
+struct FileReadCore {
     scheduler: Arc<dyn EncodingsIo>,
-    // The default projection to be applied to all reads
     base_projection: ReaderProjection,
-    num_rows: u64,
-    metadata: Arc<CachedFileMetadata>,
+    metadata_provider: FileMetadataProvider,
     decoder_plugins: Arc<DecoderPlugins>,
     cache: Arc<LanceCache>,
     options: FileReaderOptions,
 }
+
+/// A projection-scoped reader for Lance files.
+///
+/// This reader fixes a base projection at construction time. All later reads
+/// must stay within that projection, which lets the reader load only the column
+/// metadata needed by the base projection when opening from a [`FileMetadataIndex`].
+/// It intentionally does not expose APIs that require synchronous access to full
+/// file metadata.
+#[derive(Debug, Clone)]
+pub struct ProjectedFileReader {
+    core: FileReadCore,
+}
+
+/// A Lance file reader backed by fully decoded file metadata.
+#[derive(Debug, Clone)]
+pub struct FileReader {
+    core: FileReadCore,
+    metadata: Arc<CachedFileMetadata>,
+}
 #[derive(Debug)]
 struct Footer {
     #[allow(dead_code)]
@@ -460,13 +523,8 @@ const FOOTER_LEN: usize = 40;
 impl FileReader {
     pub fn with_scheduler(&self, scheduler: Arc<dyn EncodingsIo>) -> Self {
         Self {
-            scheduler,
-            base_projection: self.base_projection.clone(),
-            cache: self.cache.clone(),
-            decoder_plugins: self.decoder_plugins.clone(),
+            core: self.core.with_scheduler(scheduler),
             metadata: self.metadata.clone(),
-            options: self.options.clone(),
-            num_rows: self.num_rows,
         }
     }
 
@@ -481,23 +539,23 @@ impl FileReader {
         &self,
         stats: Arc<dyn lance_core::utils::io_stats::IoStatsRecorder>,
     ) -> Self {
-        match self.scheduler.with_io_stats(stats) {
+        match self.core.scheduler.with_io_stats(stats) {
             Some(scheduler) => self.with_scheduler(scheduler),
             None => self.clone(),
         }
     }
 
     pub fn num_rows(&self) -> u64 {
-        self.num_rows
+        self.core.num_rows()
     }
 
     pub fn metadata(&self) -> &Arc<CachedFileMetadata> {
         &self.metadata
     }
 
-    pub fn file_statistics(&self) -> FileStatistics {
-        let column_metadatas = &self.metadata().column_metadatas;
-
+    fn statistics_from_column_metadata(
+        column_metadatas: &[pbfile::ColumnMetadata],
+    ) -> FileStatistics {
         let column_stats = column_metadatas
             .iter()
             .map(|col_metadata| {
@@ -519,22 +577,12 @@ impl FileReader {
         }
     }
 
-    pub async fn read_global_buffer(&self, index: u32) -> Result<Bytes> {
-        let buffer_desc = self.metadata.file_buffers.get(index as usize).ok_or_else(||Error::invalid_input(format!("request for global buffer at index {} but there were only {} global buffers in the file", index, self.metadata.file_buffers.len())))?;
-
-        // If the buffer's bytes were captured by the tail read at open, serve them
-        // from memory with no additional I/O. Larger buffers (outside the window)
-        // are not retained and fall back to a dedicated read.
-        if let Some(bytes) = self.metadata.retained_global_buffers.get(&index) {
-            return Ok(bytes.clone());
-        }
+    pub fn file_statistics(&self) -> FileStatistics {
+        Self::statistics_from_column_metadata(&self.metadata().column_metadatas)
+    }
 
-        self.scheduler
-            .submit_single(
-                buffer_desc.position..buffer_desc.position + buffer_desc.size,
-                0,
-            )
-            .await
+    pub async fn read_global_buffer(&self, index: u32) -> Result<Bytes> {
+        self.core.read_global_buffer(index).await
     }
 
     async fn read_tail(scheduler: &FileScheduler) -> Result<(Bytes, u64)> {
@@ -548,6 +596,47 @@ impl FileReader {
         Ok((tail_bytes, file_size))
     }
 
+    async fn read_range_from_tail_or_scheduler(
+        tail_bytes: &Bytes,
+        tail_offset: u64,
+        scheduler: &FileScheduler,
+        range: Range<u64>,
+    ) -> Result<Bytes> {
+        let tail_end = tail_offset + tail_bytes.len() as u64;
+        if range.start >= tail_offset && range.end <= tail_end {
+            let rel_start = (range.start - tail_offset) as usize;
+            let rel_end = (range.end - tail_offset) as usize;
+            Ok(tail_bytes.slice(rel_start..rel_end))
+        } else {
+            scheduler.submit_single(range, 0).await
+        }
+    }
+
+    fn retained_global_buffers_from_tail(
+        gbo_table: &[BufferDescriptor],
+        tail_bytes: &Bytes,
+        tail_offset: u64,
+    ) -> BTreeMap<u32, Bytes> {
+        let tail_end = tail_offset + tail_bytes.len() as u64;
+        gbo_table
+            .iter()
+            .enumerate()
+            .skip(1)
+            .filter_map(|(index, buffer)| {
+                let start = buffer.position;
+                let end = buffer.position + buffer.size;
+                if start >= tail_offset && end <= tail_end {
+                    let rel_start = (start - tail_offset) as usize;
+                    let rel_end = (end - tail_offset) as usize;
+                    let bytes = Bytes::copy_from_slice(&tail_bytes[rel_start..rel_end]);
+                    Some((index as u32, bytes))
+                } else {
+                    None
+                }
+            })
+            .collect()
+    }
+
     // Checks to make sure the footer is written correctly and returns the
     // position of the file descriptor (which comes from the footer)
     fn decode_footer(footer_bytes: &Bytes) -> Result<Footer> {
@@ -602,15 +691,22 @@ impl FileReader {
         let column_metadata_start = footer.column_meta_start;
         // cmo == column_metadata_offsets
         let cmo_table_size = 16 * footer.num_columns as usize;
+        if column_metadata_bytes.len() < cmo_table_size {
+            return Err(Error::invalid_input(format!(
+                "column metadata region has {} bytes but CMO table needs {} bytes for {} columns",
+                column_metadata_bytes.len(),
+                cmo_table_size,
+                footer.num_columns
+            )));
+        }
         let cmo_table = column_metadata_bytes.slice(column_metadata_bytes.len() - cmo_table_size..);
+        let column_metadata_offsets = Self::decode_cmo_table(cmo_table, footer)?;
 
-        (0..footer.num_columns)
-            .map(|col_idx| {
-                let offset = (col_idx * 16) as usize;
-                let position = LittleEndian::read_u64(&cmo_table[offset..offset + 8]);
-                let length = LittleEndian::read_u64(&cmo_table[offset + 8..offset + 16]);
-                let normalized_position = (position - column_metadata_start) as usize;
-                let normalized_end = normalized_position + (length as usize);
+        column_metadata_offsets
+            .iter()
+            .map(|(position, length)| {
+                let normalized_position = (*position - column_metadata_start) as usize;
+                let normalized_end = normalized_position + (*length as usize);
                 Ok(pbfile::ColumnMetadata::decode(
                     &column_metadata_bytes[normalized_position..normalized_end],
                 )?)
@@ -618,6 +714,44 @@ impl FileReader {
             .collect::<Result<Vec<_>>>()
     }
 
+    fn decode_cmo_table(cmo_table: Bytes, footer: &Footer) -> Result<Arc<[(u64, u64)]>> {
+        let expected_size = 16 * footer.num_columns as usize;
+        if cmo_table.len() != expected_size {
+            return Err(Error::invalid_input(format!(
+                "column metadata offset table has {} bytes but expected {} bytes for {} columns",
+                cmo_table.len(),
+                expected_size,
+                footer.num_columns
+            )));
+        }
+
+        let mut offsets = Vec::with_capacity(footer.num_columns as usize);
+        for col_idx in 0..footer.num_columns {
+            let offset = (col_idx * 16) as usize;
+            let position = LittleEndian::read_u64(&cmo_table[offset..offset + 8]);
+            let length = LittleEndian::read_u64(&cmo_table[offset + 8..offset + 16]);
+            let end = position.checked_add(length).ok_or_else(|| {
+                Error::invalid_input(format!(
+                    "column metadata range overflows for column index {}, position={}, length={}",
+                    col_idx, position, length
+                ))
+            })?;
+            if position < footer.column_meta_start || end > footer.column_meta_offsets_start {
+                return Err(Error::invalid_input(format!(
+                    "column metadata range for column index {} is outside metadata region: position={}, length={}, metadata_start={}, cmo_start={}",
+                    col_idx,
+                    position,
+                    length,
+                    footer.column_meta_start,
+                    footer.column_meta_offsets_start
+                )));
+            }
+            offsets.push((position, length));
+        }
+
+        Ok(Arc::from(offsets))
+    }
+
     async fn optimistic_tail_read(
         data: &Bytes,
         start_pos: u64,
@@ -757,24 +891,8 @@ impl FileReader {
         // is decoded above and never fetched via read_global_buffer) out of the tail
         // so read_global_buffer can serve them without I/O. We copy rather than slice
         // so the much larger tail allocation can be released once decoding is done.
-        let tail_end = tail_offset + tail_bytes.len() as u64;
-        let retained_global_buffers = gbo_table
-            .iter()
-            .enumerate()
-            .skip(1)
-            .filter_map(|(index, buffer)| {
-                let start = buffer.position;
-                let end = buffer.position + buffer.size;
-                if start >= tail_offset && end <= tail_end {
-                    let rel_start = (start - tail_offset) as usize;
-                    let rel_end = (end - tail_offset) as usize;
-                    let bytes = Bytes::copy_from_slice(&tail_bytes[rel_start..rel_end]);
-                    Some((index as u32, bytes))
-                } else {
-                    None
-                }
-            })
-            .collect();
+        let retained_global_buffers =
+            Self::retained_global_buffers_from_tail(&gbo_table, &tail_bytes, tail_offset);
 
         Ok(CachedFileMetadata {
             file_schema: Arc::new(schema),
@@ -793,6 +911,87 @@ impl FileReader {
         })
     }
 
+    async fn read_metadata_index_with_known_schema(
+        scheduler: &FileScheduler,
+        known_schema: Option<(Arc<Schema>, u64)>,
+    ) -> Result<FileMetadataIndex> {
+        let (tail_bytes, file_len) = Self::read_tail(scheduler).await?;
+        let tail_offset = file_len - tail_bytes.len() as u64;
+        let footer = Self::decode_footer(&tail_bytes)?;
+
+        let file_version = LanceFileVersion::try_from_major_minor(
+            footer.major_version as u32,
+            footer.minor_version as u32,
+        )?;
+
+        let gbo_table =
+            Self::decode_gbo_table(&tail_bytes, file_len, scheduler, &footer, file_version).await?;
+        if gbo_table.is_empty() {
+            return Err(Error::internal(
+                "File did not contain any global buffers, schema expected".to_string(),
+            ));
+        }
+        let (file_schema, num_rows) = match known_schema {
+            Some((file_schema, num_rows)) => (file_schema, num_rows),
+            None => {
+                let schema_buffer = &gbo_table[0];
+                let schema_bytes = Self::read_range_from_tail_or_scheduler(
+                    &tail_bytes,
+                    tail_offset,
+                    scheduler,
+                    schema_buffer.position..schema_buffer.position + schema_buffer.size,
+                )
+                .await?;
+                let (num_rows, schema) = Self::decode_schema(schema_bytes)?;
+                (Arc::new(schema), num_rows)
+            }
+        };
+
+        let cmo_table = Self::read_range_from_tail_or_scheduler(
+            &tail_bytes,
+            tail_offset,
+            scheduler,
+            footer.column_meta_offsets_start..footer.global_buff_offsets_start,
+        )
+        .await?;
+        let column_metadata_offsets = Self::decode_cmo_table(cmo_table, &footer)?;
+
+        let retained_global_buffers =
+            Self::retained_global_buffers_from_tail(&gbo_table, &tail_bytes, tail_offset);
+
+        Ok(FileMetadataIndex {
+            file_schema,
+            num_rows,
+            file_buffers: gbo_table,
+            column_metadata_offsets,
+            num_columns: footer.num_columns,
+            version: file_version,
+            file_size_bytes: file_len,
+            retained_global_buffers,
+        })
+    }
+
+    /// Reads the lightweight metadata index from a file.
+    ///
+    /// This reads the file schema from the schema global buffer. Use
+    /// [`Self::read_metadata_index_with_schema`] when the caller already has
+    /// the schema and row count from a higher-level metadata source.
+    pub async fn read_metadata_index(scheduler: &FileScheduler) -> Result<FileMetadataIndex> {
+        Self::read_metadata_index_with_known_schema(scheduler, None).await
+    }
+
+    /// Reads the metadata index without fetching the schema global buffer.
+    ///
+    /// Use this when the caller already has the file schema and physical row
+    /// count from an enclosing metadata layer, such as a dataset manifest.
+    pub async fn read_metadata_index_with_schema(
+        scheduler: &FileScheduler,
+        file_schema: Arc<Schema>,
+        num_rows: u64,
+    ) -> Result<FileMetadataIndex> {
+        Self::read_metadata_index_with_known_schema(scheduler, Some((file_schema, num_rows))).await
+    }
+
     fn fetch_encoding<M: Default + Name + Sized>(encoding: &pbfile::Encoding) -> M {
         match &encoding.location {
             Some(pbfile::encoding::Location::Indirect(_)) => todo!(),
@@ -814,61 +1013,67 @@ impl FileReader {
             .iter()
             .enumerate()
             .map(|(col_idx, col_meta)| {
-                let page_infos = col_meta
-                    .pages
-                    .iter()
-                    .map(|page| {
-                        let num_rows = page.length;
-                        let encoding = match file_version {
-                            LanceFileVersion::V2_0 => {
-                                PageEncoding::Legacy(Self::fetch_encoding::<pbenc::ArrayEncoding>(
-                                    page.encoding.as_ref().unwrap(),
-                                ))
-                            }
-                            _ => PageEncoding::Structural(Self::fetch_encoding::<
-                                pbenc21::PageLayout,
-                            >(
-                                page.encoding.as_ref().unwrap()
-                            )),
-                        };
-                        let buffer_offsets_and_sizes = Arc::from(
-                            page.buffer_offsets
-                                .iter()
-                                .zip(page.buffer_sizes.iter())
-                                .map(|(offset, size)| {
-                                    // Starting with version 2.1 we can assert that page buffers are aligned
-                                    assert!(
-                                        file_version < LanceFileVersion::V2_1
-                                            || offset % PAGE_BUFFER_ALIGNMENT as u64 == 0
-                                    );
-                                    (*offset, *size)
-                                })
-                                .collect::<Vec<_>>(),
-                        );
-                        PageInfo {
-                            buffer_offsets_and_sizes,
-                            encoding,
-                            num_rows,
-                            priority: page.priority,
-                        }
-                    })
-                    .collect::<Vec<_>>();
+                Self::meta_to_col_info(col_idx as u32, col_meta, file_version)
+            })
+            .collect::<Vec<_>>()
+    }
+
+    fn meta_to_col_info(
+        col_idx: u32,
+        col_meta: &pbfile::ColumnMetadata,
+        file_version: LanceFileVersion,
+    ) -> Arc<ColumnInfo> {
+        let page_infos = col_meta
+            .pages
+            .iter()
+            .map(|page| {
+                let num_rows = page.length;
+                let encoding = match file_version {
+                    LanceFileVersion::V2_0 => {
+                        PageEncoding::Legacy(Self::fetch_encoding::<pbenc::ArrayEncoding>(
+                            page.encoding.as_ref().unwrap(),
+                        ))
+                    }
+                    _ => PageEncoding::Structural(Self::fetch_encoding::<pbenc21::PageLayout>(
+                        page.encoding.as_ref().unwrap(),
+                    )),
+                };
                 let buffer_offsets_and_sizes = Arc::from(
-                    col_meta
-                        .buffer_offsets
+                    page.buffer_offsets
                         .iter()
-                        .zip(col_meta.buffer_sizes.iter())
-                        .map(|(offset, size)| (*offset, *size))
+                        .zip(page.buffer_sizes.iter())
+                        .map(|(offset, size)| {
+                            // Starting with version 2.1 we can assert that page buffers are aligned
+                            assert!(
+                                file_version < LanceFileVersion::V2_1
+                                    || offset % PAGE_BUFFER_ALIGNMENT as u64 == 0
+                            );
+                            (*offset, *size)
+                        })
                         .collect::<Vec<_>>(),
                 );
-                Arc::new(ColumnInfo {
-                    index: col_idx as u32,
-                    page_infos: Arc::from(page_infos),
+                PageInfo {
                     buffer_offsets_and_sizes,
-                    encoding: Self::fetch_encoding(col_meta.encoding.as_ref().unwrap()),
-                })
+                    encoding,
+                    num_rows,
+                    priority: page.priority,
+                }
             })
-            .collect::<Vec<_>>()
+            .collect::<Vec<_>>();
+        let buffer_offsets_and_sizes = Arc::from(
+            col_meta
+                .buffer_offsets
+                .iter()
+                .zip(col_meta.buffer_sizes.iter())
+                .map(|(offset, size)| (*offset, *size))
+                .collect::<Vec<_>>(),
+        );
+        Arc::new(ColumnInfo {
+            index: col_idx,
+            page_infos: Arc::from(page_infos),
+            buffer_offsets_and_sizes,
+            encoding: Self::fetch_encoding(col_meta.encoding.as_ref().unwrap()),
+        })
     }
 
     fn validate_projection(
@@ -937,6 +1142,7 @@ impl FileReader {
     /// This method also can accept any kind of `EncodingsIo` implementation allowing
     /// for custom strategies to be used for I/O scheduling (e.g. for takes on fast
     /// disks it may be better to avoid asynchronous overhead).
+    /// Opens a data reader backed by fully decoded file metadata.
     pub async fn try_open_with_file_metadata(
         scheduler: Arc<dyn EncodingsIo>,
         path: Path,
@@ -947,22 +1153,17 @@ impl FileReader {
         options: FileReaderOptions,
     ) -> Result<Self> {
         let cache = Arc::new(cache.with_key_prefix(path.as_ref()));
-
-        if let Some(base_projection) = base_projection.as_ref() {
-            Self::validate_projection(base_projection, &file_metadata)?;
-        }
-        let num_rows = file_metadata.num_rows;
-        Ok(Self {
+        let core = FileReadCore::try_new(
             scheduler,
-            base_projection: base_projection.unwrap_or(ReaderProjection::from_whole_schema(
-                file_metadata.file_schema.as_ref(),
-                file_metadata.version(),
-            )),
-            num_rows,
-            metadata: file_metadata,
+            base_projection,
             decoder_plugins,
+            FileMetadataProvider::Full(file_metadata.clone()),
             cache,
             options,
+        )?;
+        Ok(Self {
+            core,
+            metadata: file_metadata,
         })
     }
 
@@ -1031,30 +1232,6 @@ impl FileReader {
         .await
     }
 
-    async fn read_range(
-        &self,
-        range: Range<u64>,
-        batch_size: u32,
-        projection: ReaderProjection,
-        filter: FilterExpression,
-    ) -> Result<BoxStream<'static, ReadBatchTask>> {
-        // Create and initialize the stream
-        Self::do_read_range(
-            self.collect_columns_from_projection(&projection)?,
-            self.scheduler.clone(),
-            self.cache.clone(),
-            self.num_rows,
-            self.decoder_plugins.clone(),
-            range,
-            batch_size,
-            projection,
-            filter,
-            self.options.decoder_config.clone(),
-            self.options.batch_size_bytes,
-        )
-        .await
-    }
-
     #[allow(clippy::too_many_arguments)]
     async fn do_take_rows(
         column_infos: Vec<Arc<ColumnInfo>>,
@@ -1099,28 +1276,6 @@ impl FileReader {
         .await
     }
 
-    async fn take_rows(
-        &self,
-        indices: Vec<u64>,
-        batch_size: u32,
-        projection: ReaderProjection,
-    ) -> Result<BoxStream<'static, ReadBatchTask>> {
-        // Create and initialize the stream
-        Self::do_take_rows(
-            self.collect_columns_from_projection(&projection)?,
-            self.scheduler.clone(),
-            self.cache.clone(),
-            self.decoder_plugins.clone(),
-            indices,
-            batch_size,
-            projection,
-            FilterExpression::no_filter(),
-            self.options.decoder_config.clone(),
-            self.options.batch_size_bytes,
-        )
-        .await
-    }
-
     #[allow(clippy::too_many_arguments)]
     async fn do_read_ranges(
         column_infos: Vec<Arc<ColumnInfo>>,
@@ -1167,28 +1322,6 @@ impl FileReader {
         .await
     }
 
-    async fn read_ranges(
-        &self,
-        ranges: Vec<Range<u64>>,
-        batch_size: u32,
-        projection: ReaderProjection,
-        filter: FilterExpression,
-    ) -> Result<BoxStream<'static, ReadBatchTask>> {
-        Self::do_read_ranges(
-            self.collect_columns_from_projection(&projection)?,
-            self.scheduler.clone(),
-            self.cache.clone(),
-            self.decoder_plugins.clone(),
-            ranges,
-            batch_size,
-            projection,
-            filter,
-            self.options.decoder_config.clone(),
-            self.options.batch_size_bytes,
-        )
-        .await
-    }
-
     /// Creates a stream of "read tasks" to read the data from the file
     ///
     /// The arguments are similar to [`Self::read_stream_projected`] but instead of returning a stream
@@ -1219,72 +1352,9 @@ impl FileReader {
         projection: Option<ReaderProjection>,
         filter: FilterExpression,
     ) -> Result<Pin<Box<dyn Stream<Item = ReadBatchTask> + Send>>> {
-        let projection = projection.unwrap_or_else(|| self.base_projection.clone());
-        Self::validate_projection(&projection, &self.metadata)?;
-        let verify_bound = |params: &ReadBatchParams, bound: u64, inclusive: bool| {
-            if bound > self.num_rows || bound == self.num_rows && inclusive {
-                Err(Error::invalid_input(format!(
-                    "cannot read {:?} from file with {} rows",
-                    params, self.num_rows
-                )))
-            } else {
-                Ok(())
-            }
-        };
-        match &params {
-            ReadBatchParams::Indices(indices) => {
-                for idx in indices {
-                    match idx {
-                        None => {
-                            return Err(Error::invalid_input("Null value in indices array"));
-                        }
-                        Some(idx) => {
-                            verify_bound(&params, idx as u64, true)?;
-                        }
-                    }
-                }
-                let indices = indices.iter().map(|idx| idx.unwrap() as u64).collect();
-                self.take_rows(indices, batch_size, projection).await
-            }
-            ReadBatchParams::Range(range) => {
-                verify_bound(&params, range.end as u64, false)?;
-                self.read_range(
-                    range.start as u64..range.end as u64,
-                    batch_size,
-                    projection,
-                    filter,
-                )
-                .await
-            }
-            ReadBatchParams::Ranges(ranges) => {
-                let mut ranges_u64 = Vec::with_capacity(ranges.len());
-                for range in ranges.as_ref() {
-                    verify_bound(&params, range.end, false)?;
-                    ranges_u64.push(range.start..range.end);
-                }
-                self.read_ranges(ranges_u64, batch_size, projection, filter)
-                    .await
-            }
-            ReadBatchParams::RangeFrom(range) => {
-                verify_bound(&params, range.start as u64, true)?;
-                self.read_range(
-                    range.start as u64..self.num_rows,
-                    batch_size,
-                    projection,
-                    filter,
-                )
-                .await
-            }
-            ReadBatchParams::RangeTo(range) => {
-                verify_bound(&params, range.end as u64, false)?;
-                self.read_range(0..range.end as u64, batch_size, projection, filter)
-                    .await
-            }
-            ReadBatchParams::RangeFull => {
-                self.read_range(0..self.num_rows, batch_size, projection, filter)
-                    .await
-            }
-        }
+        self.core
+            .read_tasks(params, batch_size, projection, filter)
+            .await
     }
 
     /// Reads data from the file as a stream of record batches
@@ -1357,11 +1427,11 @@ impl FileReader {
 
         let config = SchedulerDecoderConfig {
             batch_size,
-            cache: self.cache.clone(),
-            decoder_plugins: self.decoder_plugins.clone(),
-            io: self.scheduler.clone(),
-            decoder_config: self.options.decoder_config.clone(),
-            batch_size_bytes: self.options.batch_size_bytes,
+            cache: self.core.cache.clone(),
+            decoder_plugins: self.core.decoder_plugins.clone(),
+            io: self.core.scheduler.clone(),
+            decoder_config: self.core.options.decoder_config.clone(),
+            batch_size_bytes: self.core.options.batch_size_bytes,
         };
 
         let requested_rows = RequestedRows::Indices(indices);
@@ -1397,11 +1467,11 @@ impl FileReader {
 
         let config = SchedulerDecoderConfig {
             batch_size,
-            cache: self.cache.clone(),
-            decoder_plugins: self.decoder_plugins.clone(),
-            io: self.scheduler.clone(),
-            decoder_config: self.options.decoder_config.clone(),
-            batch_size_bytes: self.options.batch_size_bytes,
+            cache: self.core.cache.clone(),
+            decoder_plugins: self.core.decoder_plugins.clone(),
+            io: self.core.scheduler.clone(),
+            decoder_config: self.core.options.decoder_config.clone(),
+            batch_size_bytes: self.core.options.batch_size_bytes,
         };
 
         let requested_rows = RequestedRows::Ranges(ranges);
@@ -1424,7 +1494,7 @@ impl FileReader {
         filter: FilterExpression,
     ) -> Result<Box<dyn RecordBatchReader + Send + 'static>> {
         let column_infos = self.collect_columns_from_projection(&projection)?;
-        let num_rows = self.num_rows;
+        let num_rows = self.core.num_rows();
 
         debug!(
             "Reading range {:?} with batch_size {} from file with {} rows and {} columns into schema with {} columns",
@@ -1437,11 +1507,11 @@ impl FileReader {
 
         let config = SchedulerDecoderConfig {
             batch_size,
-            cache: self.cache.clone(),
-            decoder_plugins: self.decoder_plugins.clone(),
-            io: self.scheduler.clone(),
-            decoder_config: self.options.decoder_config.clone(),
-            batch_size_bytes: self.options.batch_size_bytes,
+            cache: self.core.cache.clone(),
+            decoder_plugins: self.core.decoder_plugins.clone(),
+            io: self.core.scheduler.clone(),
+            decoder_config: self.core.options.decoder_config.clone(),
+            batch_size_bytes: self.core.options.batch_size_bytes,
         };
 
         let requested_rows = RequestedRows::Ranges(vec![range]);
@@ -1474,13 +1544,14 @@ impl FileReader {
         projection: Option<ReaderProjection>,
         filter: FilterExpression,
     ) -> Result<Box<dyn RecordBatchReader + Send + 'static>> {
-        let projection = projection.unwrap_or_else(|| self.base_projection.clone());
+        let projection = projection.unwrap_or_else(|| self.core.base_projection.clone());
         Self::validate_projection(&projection, &self.metadata)?;
         let verify_bound = |params: &ReadBatchParams, bound: u64, inclusive: bool| {
-            if bound > self.num_rows || bound == self.num_rows && inclusive {
+            if bound > self.core.num_rows() || bound == self.core.num_rows() && inclusive {
                 Err(Error::invalid_input(format!(
                     "cannot read {:?} from file with {} rows",
-                    params, self.num_rows
+                    params,
+                    self.core.num_rows()
                 )))
             } else {
                 Ok(())
@@ -1521,7 +1592,7 @@ impl FileReader {
             ReadBatchParams::RangeFrom(range) => {
                 verify_bound(&params, range.start as u64, true)?;
                 self.read_range_blocking(
-                    range.start as u64..self.num_rows,
+                    range.start as u64..self.core.num_rows(),
                     batch_size,
                     projection,
                     filter,
@@ -1532,41 +1603,637 @@ impl FileReader {
                 self.read_range_blocking(0..range.end as u64, batch_size, projection, filter)
             }
             ReadBatchParams::RangeFull => {
-                self.read_range_blocking(0..self.num_rows, batch_size, projection, filter)
+                self.read_range_blocking(0..self.core.num_rows(), batch_size, projection, filter)
             }
         }
     }
 
-    /// Reads data from the file as a stream of record batches
-    ///
-    /// This is similar to [`Self::read_stream_projected`] but uses the base projection
-    /// provided when the file was opened (or reads all columns if the file was
-    /// opened without a base projection)
-    ///
-    /// # Why is this async?
+    /// Reads data from the file as a stream of record batches
+    ///
+    /// This is similar to [`Self::read_stream_projected`] but uses the base projection
+    /// provided when the file was opened (or reads all columns if the file was
+    /// opened without a base projection)
+    ///
+    /// # Why is this async?
+    ///
+    /// This delegates to [`Self::read_stream_projected`], which awaits the
+    /// decode scheduler's `initialize` step before returning the stream.
+    /// See [`Self::read_tasks`] for the rationale.
+    pub async fn read_stream(
+        &self,
+        params: ReadBatchParams,
+        batch_size: u32,
+        batch_readahead: u32,
+        filter: FilterExpression,
+    ) -> Result<Pin<Box<dyn RecordBatchStream>>> {
+        self.read_stream_projected(
+            params,
+            batch_size,
+            batch_readahead,
+            self.core.base_projection.clone(),
+            filter,
+        )
+        .await
+    }
+
+    pub fn schema(&self) -> &Arc<Schema> {
+        self.core.schema()
+    }
+}
+
+impl FileMetadataProvider {
+    fn version(&self) -> LanceFileVersion {
+        match self {
+            Self::Full(metadata) => metadata.version(),
+            Self::Indexed(metadata_index) => metadata_index.version,
+        }
+    }
+
+    fn num_rows(&self) -> u64 {
+        match self {
+            Self::Full(metadata) => metadata.num_rows,
+            Self::Indexed(metadata_index) => metadata_index.num_rows,
+        }
+    }
+
+    fn schema(&self) -> &Arc<Schema> {
+        match self {
+            Self::Full(metadata) => &metadata.file_schema,
+            Self::Indexed(metadata_index) => &metadata_index.file_schema,
+        }
+    }
+
+    fn file_buffers(&self) -> &Vec<BufferDescriptor> {
+        match self {
+            Self::Full(metadata) => &metadata.file_buffers,
+            Self::Indexed(metadata_index) => &metadata_index.file_buffers,
+        }
+    }
+
+    fn retained_global_buffers(&self) -> &BTreeMap<u32, Bytes> {
+        match self {
+            Self::Full(metadata) => &metadata.retained_global_buffers,
+            Self::Indexed(metadata_index) => &metadata_index.retained_global_buffers,
+        }
+    }
+
+    fn file_statistics(&self) -> Option<FileStatistics> {
+        let metadata = match self {
+            Self::Full(metadata) => metadata,
+            Self::Indexed(_) => return None,
+        };
+        Some(FileReader::statistics_from_column_metadata(
+            &metadata.column_metadatas,
+        ))
+    }
+
+    fn supports_indexed_projection(
+        projection: &ReaderProjection,
+        version: LanceFileVersion,
+    ) -> bool {
+        version >= LanceFileVersion::V2_1
+            && !projection.schema.fields.is_empty()
+            && projection.schema.fields.len() == projection.column_indices.len()
+            && projection.schema.fields.iter().all(|field| {
+                field.children.is_empty() && !field.is_blob() && !field.is_packed_struct()
+            })
+    }
+
+    fn validate_indexed_projection(
+        projection: &ReaderProjection,
+        metadata_index: &FileMetadataIndex,
+    ) -> Result<()> {
+        if projection.schema.fields.is_empty() {
+            return Err(Error::invalid_input(
+                "Attempt to read zero columns from the file, at least one column must be specified"
+                    .to_string(),
+            ));
+        }
+        let mut column_indices_seen = BTreeSet::new();
+        for column_index in &projection.column_indices {
+            if !column_indices_seen.insert(*column_index) {
+                return Err(Error::invalid_input(format!(
+                    "The projection specified the column index {} more than once",
+                    column_index
+                )));
+            }
+            if *column_index >= metadata_index.num_columns {
+                return Err(Error::invalid_input(format!(
+                    "The projection specified the column index {} but there are only {} columns in the file",
+                    column_index, metadata_index.num_columns
+                )));
+            }
+        }
+        if !Self::supports_indexed_projection(projection, metadata_index.version) {
+            return Err(Error::not_supported(format!(
+                "lazy column metadata loading only supports direct V2.1+ top-level physical column projections; got file version {:?}, {} schema fields, and {} column indices",
+                metadata_index.version,
+                projection.schema.fields.len(),
+                projection.column_indices.len()
+            )));
+        }
+        Ok(())
+    }
+
+    fn validate_projection(&self, projection: &ReaderProjection) -> Result<()> {
+        match self {
+            Self::Full(metadata) => FileReader::validate_projection(projection, metadata),
+            Self::Indexed(metadata_index) => {
+                Self::validate_indexed_projection(projection, metadata_index)
+            }
+        }
+    }
+
+    fn column_metadata_range(
+        metadata_index: &FileMetadataIndex,
+        column_index: u32,
+    ) -> Result<Range<u64>> {
+        let (position, length) = metadata_index
+            .column_metadata_offsets
+            .get(column_index as usize)
+            .copied()
+            .ok_or_else(|| {
+                Error::invalid_input(format!(
+                    "The projection specified the column index {} but there are only {} columns in the file",
+                    column_index, metadata_index.num_columns
+                ))
+            })?;
+        let end = position.checked_add(length).ok_or_else(|| {
+            Error::invalid_input(format!(
+                "column metadata range overflows for column index {}, position={}, length={}",
+                column_index, position, length
+            ))
+        })?;
+        Ok(position..end)
+    }
+
+    async fn load_indexed_column_infos(
+        metadata_index: &FileMetadataIndex,
+        io: &Arc<dyn EncodingsIo>,
+        cache: &Arc<LanceCache>,
+        column_indices: &[u32],
+    ) -> Result<Vec<Arc<ColumnInfo>>> {
+        let mut column_infos = vec![None; column_indices.len()];
+        let mut missing_columns = Vec::new();
+
+        for (result_index, column_index) in column_indices.iter().copied().enumerate() {
+            let cache_key = ColumnMetadataCacheKey { column_index };
+            if let Some(cached) = cache.get_with_key(&cache_key).await {
+                column_infos[result_index] = Some(cached.column_info.clone());
+            } else {
+                let range = Self::column_metadata_range(metadata_index, column_index)?;
+                missing_columns.push((result_index, column_index, range));
+            }
+        }
+
+        missing_columns.sort_by_key(|(_, _, range)| range.start);
+        if !missing_columns.is_empty() {
+            let ranges = missing_columns
+                .iter()
+                .map(|(_, _, range)| range.clone())
+                .collect::<Vec<_>>();
+            let metadata_bytes = io.submit_request(ranges, 0).await?;
+            for ((result_index, column_index, _), bytes) in
+                missing_columns.into_iter().zip(metadata_bytes.into_iter())
+            {
+                let column_metadata = pbfile::ColumnMetadata::decode(bytes)?;
+                let column_info = FileReader::meta_to_col_info(
+                    column_index,
+                    &column_metadata,
+                    metadata_index.version,
+                );
+                let cached = Arc::new(CachedColumnMetadata {
+                    column_metadata,
+                    column_info: column_info.clone(),
+                });
+                let cache_key = ColumnMetadataCacheKey { column_index };
+                cache.insert_with_key(&cache_key, cached).await;
+                column_infos[result_index] = Some(column_info);
+            }
+        }
+
+        column_infos
+            .into_iter()
+            .enumerate()
+            .map(|(idx, column_info)| {
+                column_info.ok_or_else(|| {
+                    Error::internal(format!(
+                        "lazy metadata loader did not load requested projection column at position {}",
+                        idx
+                    ))
+                })
+            })
+            .collect()
+    }
+
+    async fn prepare_projection(
+        &self,
+        projection: &ReaderProjection,
+        io: &Arc<dyn EncodingsIo>,
+        cache: &Arc<LanceCache>,
+    ) -> Result<PreparedProjection> {
+        self.validate_projection(projection)?;
+        match self {
+            Self::Full(metadata) => Ok(PreparedProjection {
+                column_infos: metadata.column_infos.clone(),
+                decoder_projection: projection.clone(),
+            }),
+            Self::Indexed(metadata_index) => {
+                let column_infos = Self::load_indexed_column_infos(
+                    metadata_index,
+                    io,
+                    cache,
+                    &projection.column_indices,
+                )
+                .await?;
+                let decoder_projection = ReaderProjection {
+                    schema: projection.schema.clone(),
+                    column_indices: (0..projection.column_indices.len())
+                        .map(|idx| idx as u32)
+                        .collect(),
+                };
+                Ok(PreparedProjection {
+                    column_infos,
+                    decoder_projection,
+                })
+            }
+        }
+    }
+}
+
+impl FileReadCore {
+    fn try_new(
+        scheduler: Arc<dyn EncodingsIo>,
+        base_projection: Option<ReaderProjection>,
+        decoder_plugins: Arc<DecoderPlugins>,
+        metadata_provider: FileMetadataProvider,
+        cache: Arc<LanceCache>,
+        options: FileReaderOptions,
+    ) -> Result<Self> {
+        if let Some(base_projection) = base_projection.as_ref() {
+            metadata_provider.validate_projection(base_projection)?;
+        }
+        let base_projection = base_projection.unwrap_or(ReaderProjection::from_whole_schema(
+            metadata_provider.schema().as_ref(),
+            metadata_provider.version(),
+        ));
+        Ok(Self {
+            scheduler,
+            base_projection,
+            metadata_provider,
+            decoder_plugins,
+            cache,
+            options,
+        })
+    }
+
+    fn with_scheduler(&self, scheduler: Arc<dyn EncodingsIo>) -> Self {
+        Self {
+            scheduler,
+            base_projection: self.base_projection.clone(),
+            metadata_provider: self.metadata_provider.clone(),
+            decoder_plugins: self.decoder_plugins.clone(),
+            cache: self.cache.clone(),
+            options: self.options.clone(),
+        }
+    }
+
+    fn version(&self) -> LanceFileVersion {
+        self.metadata_provider.version()
+    }
+
+    fn num_rows(&self) -> u64 {
+        self.metadata_provider.num_rows()
+    }
+
+    fn schema(&self) -> &Arc<Schema> {
+        self.metadata_provider.schema()
+    }
+
+    async fn read_global_buffer(&self, index: u32) -> Result<Bytes> {
+        let file_buffers = self.metadata_provider.file_buffers();
+        let buffer_desc = file_buffers.get(index as usize).ok_or_else(|| {
+            Error::invalid_input(format!(
+                "request for global buffer at index {} but there were only {} global buffers in the file",
+                index,
+                file_buffers.len()
+            ))
+        })?;
+
+        if let Some(bytes) = self.metadata_provider.retained_global_buffers().get(&index) {
+            return Ok(bytes.clone());
+        }
+
+        let bytes = self
+            .scheduler
+            .submit_request(
+                vec![buffer_desc.position..buffer_desc.position + buffer_desc.size],
+                0,
+            )
+            .await?;
+        bytes.into_iter().next().ok_or_else(|| {
+            Error::internal(format!(
+                "global buffer read for index {} returned no bytes",
+                index
+            ))
+        })
+    }
+
+    async fn read_range(
+        &self,
+        range: Range<u64>,
+        batch_size: u32,
+        projection: ReaderProjection,
+        filter: FilterExpression,
+    ) -> Result<BoxStream<'static, ReadBatchTask>> {
+        let prepared = self
+            .metadata_provider
+            .prepare_projection(&projection, &self.scheduler, &self.cache)
+            .await?;
+        FileReader::do_read_range(
+            prepared.column_infos,
+            self.scheduler.clone(),
+            self.cache.clone(),
+            self.num_rows(),
+            self.decoder_plugins.clone(),
+            range,
+            batch_size,
+            prepared.decoder_projection,
+            filter,
+            self.options.decoder_config.clone(),
+            self.options.batch_size_bytes,
+        )
+        .await
+    }
+
+    async fn take_rows(
+        &self,
+        indices: Vec<u64>,
+        batch_size: u32,
+        projection: ReaderProjection,
+    ) -> Result<BoxStream<'static, ReadBatchTask>> {
+        let prepared = self
+            .metadata_provider
+            .prepare_projection(&projection, &self.scheduler, &self.cache)
+            .await?;
+        FileReader::do_take_rows(
+            prepared.column_infos,
+            self.scheduler.clone(),
+            self.cache.clone(),
+            self.decoder_plugins.clone(),
+            indices,
+            batch_size,
+            prepared.decoder_projection,
+            FilterExpression::no_filter(),
+            self.options.decoder_config.clone(),
+            self.options.batch_size_bytes,
+        )
+        .await
+    }
+
+    async fn read_ranges(
+        &self,
+        ranges: Vec<Range<u64>>,
+        batch_size: u32,
+        projection: ReaderProjection,
+        filter: FilterExpression,
+    ) -> Result<BoxStream<'static, ReadBatchTask>> {
+        let prepared = self
+            .metadata_provider
+            .prepare_projection(&projection, &self.scheduler, &self.cache)
+            .await?;
+        FileReader::do_read_ranges(
+            prepared.column_infos,
+            self.scheduler.clone(),
+            self.cache.clone(),
+            self.decoder_plugins.clone(),
+            ranges,
+            batch_size,
+            prepared.decoder_projection,
+            filter,
+            self.options.decoder_config.clone(),
+            self.options.batch_size_bytes,
+        )
+        .await
+    }
+
+    async fn read_tasks(
+        &self,
+        params: ReadBatchParams,
+        batch_size: u32,
+        projection: Option<ReaderProjection>,
+        filter: FilterExpression,
+    ) -> Result<Pin<Box<dyn Stream<Item = ReadBatchTask> + Send>>> {
+        let projection = projection.unwrap_or_else(|| self.base_projection.clone());
+        let verify_bound = |params: &ReadBatchParams, bound: u64, inclusive: bool| {
+            if bound > self.num_rows() || bound == self.num_rows() && inclusive {
+                Err(Error::invalid_input(format!(
+                    "cannot read {:?} from file with {} rows",
+                    params,
+                    self.num_rows()
+                )))
+            } else {
+                Ok(())
+            }
+        };
+        match &params {
+            ReadBatchParams::Indices(indices) => {
+                for idx in indices {
+                    match idx {
+                        None => {
+                            return Err(Error::invalid_input("Null value in indices array"));
+                        }
+                        Some(idx) => {
+                            verify_bound(&params, idx as u64, true)?;
+                        }
+                    }
+                }
+                let indices = indices.iter().map(|idx| idx.unwrap() as u64).collect();
+                self.take_rows(indices, batch_size, projection).await
+            }
+            ReadBatchParams::Range(range) => {
+                verify_bound(&params, range.end as u64, false)?;
+                self.read_range(
+                    range.start as u64..range.end as u64,
+                    batch_size,
+                    projection,
+                    filter,
+                )
+                .await
+            }
+            ReadBatchParams::Ranges(ranges) => {
+                let mut ranges_u64 = Vec::with_capacity(ranges.len());
+                for range in ranges.as_ref() {
+                    verify_bound(&params, range.end, false)?;
+                    ranges_u64.push(range.start..range.end);
+                }
+                self.read_ranges(ranges_u64, batch_size, projection, filter)
+                    .await
+            }
+            ReadBatchParams::RangeFrom(range) => {
+                verify_bound(&params, range.start as u64, true)?;
+                self.read_range(
+                    range.start as u64..self.num_rows(),
+                    batch_size,
+                    projection,
+                    filter,
+                )
+                .await
+            }
+            ReadBatchParams::RangeTo(range) => {
+                verify_bound(&params, range.end as u64, false)?;
+                self.read_range(0..range.end as u64, batch_size, projection, filter)
+                    .await
+            }
+            ReadBatchParams::RangeFull => {
+                self.read_range(0..self.num_rows(), batch_size, projection, filter)
+                    .await
+            }
+        }
+    }
+}
+
+impl ProjectedFileReader {
+    /// Opens a data reader backed by indexed column metadata.
+    ///
+    /// `base_projection` must be a supported indexed projection. Reads that do
+    /// not pass an explicit projection use this base projection.
+    pub async fn try_open(
+        scheduler: FileScheduler,
+        base_projection: Option<ReaderProjection>,
+        decoder_plugins: Arc<DecoderPlugins>,
+        cache: &LanceCache,
+        options: FileReaderOptions,
+    ) -> Result<Self> {
+        let base_projection = Self::require_indexed_base_projection(base_projection)?;
+        let metadata_index = Arc::new(FileReader::read_metadata_index(&scheduler).await?);
+        let path = scheduler.reader().path().clone();
+        let encodings_io =
+            LanceEncodingsIo::new(scheduler).with_read_chunk_size(options.read_chunk_size);
+        Self::try_open_with_metadata_index(
+            Arc::new(encodings_io),
+            path,
+            Some(base_projection),
+            decoder_plugins,
+            metadata_index,
+            cache,
+            options,
+        )
+        .await
+    }
+
+    /// Opens a data reader from a previously loaded metadata index.
     ///
-    /// This delegates to [`Self::read_stream_projected`], which awaits the
-    /// decode scheduler's `initialize` step before returning the stream.
-    /// See [`Self::read_tasks`] for the rationale.
-    pub async fn read_stream(
+    /// `base_projection` must be a supported indexed projection. Use
+    /// [`Self::try_open_with_file_metadata`] when the default read should cover
+    /// the whole file schema.
+    pub async fn try_open_with_metadata_index(
+        scheduler: Arc<dyn EncodingsIo>,
+        path: Path,
+        base_projection: Option<ReaderProjection>,
+        decoder_plugins: Arc<DecoderPlugins>,
+        metadata_index: Arc<FileMetadataIndex>,
+        cache: &LanceCache,
+        options: FileReaderOptions,
+    ) -> Result<Self> {
+        let base_projection = Self::require_indexed_base_projection(base_projection)?;
+        let cache = Arc::new(cache.with_key_prefix(path.as_ref()));
+        let core = FileReadCore::try_new(
+            scheduler,
+            Some(base_projection),
+            decoder_plugins,
+            FileMetadataProvider::Indexed(metadata_index),
+            cache,
+            options,
+        )?;
+        Ok(Self { core })
+    }
+
+    fn require_indexed_base_projection(
+        base_projection: Option<ReaderProjection>,
+    ) -> Result<ReaderProjection> {
+        base_projection.ok_or_else(|| {
+            Error::invalid_input("ProjectedFileReader requires an explicit base projection")
+        })
+    }
+
+    pub async fn try_open_with_file_metadata(
+        scheduler: Arc<dyn EncodingsIo>,
+        path: Path,
+        base_projection: Option<ReaderProjection>,
+        decoder_plugins: Arc<DecoderPlugins>,
+        file_metadata: Arc<CachedFileMetadata>,
+        cache: &LanceCache,
+        options: FileReaderOptions,
+    ) -> Result<Self> {
+        let cache = Arc::new(cache.with_key_prefix(path.as_ref()));
+        let core = FileReadCore::try_new(
+            scheduler,
+            base_projection,
+            decoder_plugins,
+            FileMetadataProvider::Full(file_metadata),
+            cache,
+            options,
+        )?;
+        Ok(Self { core })
+    }
+
+    /// Returns whether a projection can be served by indexed column metadata.
+    pub fn supports_projection(projection: &ReaderProjection, version: LanceFileVersion) -> bool {
+        FileMetadataProvider::supports_indexed_projection(projection, version)
+    }
+
+    /// Returns a clone of this reader using a different scheduler.
+    pub fn with_scheduler(&self, scheduler: Arc<dyn EncodingsIo>) -> Self {
+        Self {
+            core: self.core.with_scheduler(scheduler),
+        }
+    }
+
+    /// Returns the Lance file version.
+    pub fn version(&self) -> LanceFileVersion {
+        self.core.version()
+    }
+
+    /// Returns the number of rows in the file.
+    pub fn num_rows(&self) -> u64 {
+        self.core.num_rows()
+    }
+
+    /// Returns the file schema visible to this reader.
+    pub fn schema(&self) -> &Arc<Schema> {
+        self.core.schema()
+    }
+
+    /// Returns file statistics when this reader has full file metadata.
+    pub fn file_statistics(&self) -> Option<FileStatistics> {
+        self.core.metadata_provider.file_statistics()
+    }
+
+    #[cfg(test)]
+    fn metadata_index(&self) -> Option<&Arc<FileMetadataIndex>> {
+        match &self.core.metadata_provider {
+            FileMetadataProvider::Indexed(metadata_index) => Some(metadata_index),
+            FileMetadataProvider::Full(_) => None,
+        }
+    }
+
+    /// Reads a global buffer by index.
+    pub async fn read_global_buffer(&self, index: u32) -> Result<Bytes> {
+        self.core.read_global_buffer(index).await
+    }
+
+    /// Creates a stream of read tasks for the requested rows and projection.
+    pub async fn read_tasks(
         &self,
         params: ReadBatchParams,
         batch_size: u32,
-        batch_readahead: u32,
+        projection: Option<ReaderProjection>,
         filter: FilterExpression,
-    ) -> Result<Pin<Box<dyn RecordBatchStream>>> {
-        self.read_stream_projected(
-            params,
-            batch_size,
-            batch_readahead,
-            self.base_projection.clone(),
-            filter,
-        )
-        .await
-    }
-
-    pub fn schema(&self) -> &Arc<Schema> {
-        &self.metadata.file_schema
+    ) -> Result<Pin<Box<dyn Stream<Item = ReadBatchTask> + Send>>> {
+        self.core
+            .read_tasks(params, batch_size, projection, filter)
+            .await
     }
 }
 
@@ -1736,7 +2403,9 @@ mod tests {
     use lance_core::{ArrowResult, datatypes::Schema};
     use lance_datagen::{BatchCount, ByteCount, RowCount, array, gen_batch};
     use lance_encoding::{
-        decoder::{DecodeBatchScheduler, DecoderPlugins, FilterExpression, decode_batch},
+        decoder::{
+            DecodeBatchScheduler, DecoderPlugins, FilterExpression, ReadBatchTask, decode_batch,
+        },
         encoder::{EncodedBatch, EncodingOptions, default_encoding_strategy, encode_batch},
         version::LanceFileVersion,
     };
@@ -1745,7 +2414,9 @@ mod tests {
     use rstest::rstest;
     use tokio::sync::mpsc;
 
-    use crate::reader::{EncodedBatchReaderExt, FileReader, FileReaderOptions, ReaderProjection};
+    use crate::reader::{
+        EncodedBatchReaderExt, FileReader, FileReaderOptions, ProjectedFileReader, ReaderProjection,
+    };
     use crate::testing::{FsFixture, WrittenFile, test_cache, write_lance_file};
     use crate::writer::{EncodedBatchWriteExt, FileWriter, FileWriterOptions};
     use lance_encoding::decoder::DecoderConfig;
@@ -1778,6 +2449,24 @@ mod tests {
         .await
     }
 
+    async fn create_wide_direct_file(fs: &FsFixture, num_columns: usize) -> WrittenFile {
+        let mut reader = gen_batch();
+        for column_idx in 0..num_columns {
+            reader = reader.col(format!("c{column_idx}"), array::step::<Int32Type>());
+        }
+        let reader = reader.into_reader_rows(RowCount::from(1000), BatchCount::from(100));
+
+        write_lance_file(
+            reader,
+            fs,
+            FileWriterOptions {
+                format_version: Some(LanceFileVersion::V2_1),
+                ..Default::default()
+            },
+        )
+        .await
+    }
+
     type Transformer = Box<dyn Fn(&RecordBatch) -> RecordBatch>;
 
     async fn verify_expected(
@@ -1830,6 +2519,18 @@ mod tests {
         assert_eq!(remaining, 0);
     }
 
+    async fn collect_read_tasks(
+        tasks: Pin<Box<dyn futures::Stream<Item = ReadBatchTask> + Send>>,
+        readahead: usize,
+    ) -> Vec<RecordBatch> {
+        tasks
+            .map(|task| task.task)
+            .buffered(readahead)
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap()
+    }
+
     #[tokio::test]
     async fn test_round_trip() {
         let fs = FsFixture::default();
@@ -2110,6 +2811,237 @@ mod tests {
         );
     }
 
+    #[tokio::test]
+    async fn test_lazy_reader_direct_projection_matches_eager_reader() {
+        let fs = FsFixture::default();
+        let written_file = create_wide_direct_file(&fs, 16).await;
+
+        let file_scheduler = fs
+            .scheduler
+            .open_file(&fs.tmp_path, &CachedFileSize::unknown())
+            .await
+            .unwrap();
+        let projection = ReaderProjection::from_column_names(
+            LanceFileVersion::V2_1,
+            &written_file.schema,
+            &["c10"],
+        )
+        .unwrap();
+
+        let eager_reader = FileReader::try_open(
+            file_scheduler.clone(),
+            None,
+            Arc::<DecoderPlugins>::default(),
+            &test_cache(),
+            FileReaderOptions::default(),
+        )
+        .await
+        .unwrap();
+        let expected = eager_reader
+            .read_stream_projected(
+                lance_io::ReadBatchParams::RangeFull,
+                127,
+                16,
+                projection.clone(),
+                FilterExpression::no_filter(),
+            )
+            .await
+            .unwrap()
+            .try_collect::<Vec<_>>()
+            .await
+            .unwrap();
+
+        let cache = test_cache();
+        let lazy_reader = ProjectedFileReader::try_open(
+            file_scheduler,
+            Some(projection.clone()),
+            Arc::<DecoderPlugins>::default(),
+            &cache,
+            FileReaderOptions::default(),
+        )
+        .await
+        .unwrap();
+        let tasks = lazy_reader
+            .read_tasks(
+                lance_io::ReadBatchParams::RangeFull,
+                127,
+                None,
+                FilterExpression::no_filter(),
+            )
+            .await
+            .unwrap();
+        let actual = collect_read_tasks(tasks, 16).await;
+
+        assert_eq!(expected, actual);
+    }
+
+    #[tokio::test]
+    async fn test_lazy_reader_loads_only_requested_column_metadata() {
+        let fs = FsFixture::default();
+        let written_file = create_wide_direct_file(&fs, 512).await;
+
+        let projection = ReaderProjection::from_column_names(
+            LanceFileVersion::V2_1,
+            &written_file.schema,
+            &["c0"],
+        )
+        .unwrap();
+        let file_scheduler = fs
+            .scheduler
+            .open_file(&fs.tmp_path, &CachedFileSize::unknown())
+            .await
+            .unwrap();
+        let lazy_reader = ProjectedFileReader::try_open(
+            file_scheduler,
+            Some(projection.clone()),
+            Arc::<DecoderPlugins>::default(),
+            &test_cache(),
+            FileReaderOptions::default(),
+        )
+        .await
+        .unwrap();
+        let selected_column = projection.column_indices[0] as usize;
+        let requested_metadata_bytes = lazy_reader
+            .metadata_index()
+            .unwrap()
+            .column_metadata_offsets[selected_column]
+            .1;
+        let total_metadata_bytes = lazy_reader
+            .metadata_index()
+            .unwrap()
+            .column_metadata_offsets
+            .iter()
+            .map(|(_, length)| *length)
+            .sum::<u64>();
+        assert!(
+            total_metadata_bytes > 8 * fs.object_store.block_size() as u64,
+            "test file metadata is too small to prove lazy loading: {total_metadata_bytes} bytes"
+        );
+
+        fs.object_store.io_stats_incremental();
+        let tasks = lazy_reader
+            .read_tasks(
+                lance_io::ReadBatchParams::Range(0..0),
+                1024,
+                Some(projection.clone()),
+                FilterExpression::no_filter(),
+            )
+            .await
+            .unwrap();
+        let batches = collect_read_tasks(tasks, 1).await;
+        assert!(batches.is_empty());
+
+        let stats = fs.object_store.io_stats_incremental();
+        assert!(
+            stats.read_bytes < total_metadata_bytes / 2,
+            "lazy read fetched too much metadata: read {} bytes, requested column metadata is {} bytes, total column metadata is {} bytes",
+            stats.read_bytes,
+            requested_metadata_bytes,
+            total_metadata_bytes
+        );
+
+        fs.object_store.io_stats_incremental();
+        let tasks = lazy_reader
+            .read_tasks(
+                lance_io::ReadBatchParams::Range(0..0),
+                1024,
+                Some(projection),
+                FilterExpression::no_filter(),
+            )
+            .await
+            .unwrap();
+        let batches = collect_read_tasks(tasks, 1).await;
+        assert!(batches.is_empty());
+
+        let stats = fs.object_store.io_stats_incremental();
+        assert_eq!(
+            stats.read_iops, 0,
+            "cached column metadata should avoid repeat metadata I/O"
+        );
+        assert_eq!(
+            stats.read_bytes, 0,
+            "cached column metadata should avoid repeat metadata reads"
+        );
+    }
+
+    #[rstest]
+    #[case::before_metadata_region(90, 5)]
+    #[case::after_metadata_region(190, 20)]
+    fn test_decode_cmo_table_rejects_out_of_range_offsets(
+        #[case] position: u64,
+        #[case] length: u64,
+    ) {
+        let mut cmo_table = [0; 16];
+        cmo_table[0..8].copy_from_slice(&position.to_le_bytes());
+        cmo_table[8..16].copy_from_slice(&length.to_le_bytes());
+        let footer = super::Footer {
+            column_meta_start: 100,
+            column_meta_offsets_start: 200,
+            global_buff_offsets_start: 200,
+            num_global_buffers: 0,
+            num_columns: 1,
+            major_version: 2,
+            minor_version: 1,
+        };
+
+        let err = FileReader::decode_cmo_table(Bytes::copy_from_slice(&cmo_table), &footer)
+            .expect_err("out-of-range CMO entries must be rejected");
+        assert!(
+            matches!(err, lance_core::Error::InvalidInput { .. }),
+            "expected InvalidInput, got {err:?}"
+        );
+    }
+
+    #[tokio::test]
+    async fn test_lazy_reader_rejects_unsupported_projection() {
+        let fs = FsFixture::default();
+        let written_file = create_some_file(&fs, LanceFileVersion::V2_1).await;
+
+        let projection = ReaderProjection::from_column_names(
+            LanceFileVersion::V2_1,
+            &written_file.schema,
+            &["location"],
+        )
+        .unwrap();
+        assert!(!ProjectedFileReader::supports_projection(
+            &projection,
+            LanceFileVersion::V2_1
+        ));
+
+        let file_scheduler = fs
+            .scheduler
+            .open_file(&fs.tmp_path, &CachedFileSize::unknown())
+            .await
+            .unwrap();
+        let err = ProjectedFileReader::try_open(
+            file_scheduler.clone(),
+            None,
+            Arc::<DecoderPlugins>::default(),
+            &test_cache(),
+            FileReaderOptions::default(),
+        )
+        .await
+        .unwrap_err();
+        assert!(
+            matches!(err, lance_core::Error::InvalidInput { .. }),
+            "expected InvalidInput, got {err:?}"
+        );
+
+        let err = ProjectedFileReader::try_open(
+            file_scheduler,
+            Some(projection),
+            Arc::<DecoderPlugins>::default(),
+            &test_cache(),
+            FileReaderOptions::default(),
+        )
+        .await
+        .unwrap_err();
+        assert!(
+            matches!(err, lance_core::Error::NotSupported { .. }),
+            "expected NotSupported, got {err:?}"
+        );
+    }
+
     #[test_log::test(tokio::test)]
     async fn test_compressing_buffer() {
         let fs = FsFixture::default();
@@ -2334,7 +3266,7 @@ mod tests {
             &vec![],
             total_rows as u64,
             Arc::<DecoderPlugins>::default(),
-            file_reader.scheduler.clone(),
+            file_reader.core.scheduler.clone(),
             test_cache(),
             &FilterExpression::no_filter(),
             &DecoderConfig::default(),
@@ -2354,7 +3286,7 @@ mod tests {
             range,
             &FilterExpression::no_filter(),
             tx,
-            file_reader.scheduler.clone(),
+            file_reader.core.scheduler.clone(),
         )
     }
 
diff --git a/rust/lance/src/dataset/fragment.rs b/rust/lance/src/dataset/fragment.rs
index eb165e5f612..4ea1b51f073 100644
--- a/rust/lance/src/dataset/fragment.rs
+++ b/rust/lance/src/dataset/fragment.rs
@@ -37,7 +37,9 @@ use lance_encoding::decoder::DecoderPlugins;
 use lance_file::previous::reader::{
     FileReader as PreviousFileReader, read_batch as previous_read_batch,
 };
-use lance_file::reader::{CachedFileMetadata, FileReaderOptions, ReaderProjection};
+use lance_file::reader::{
+    CachedFileMetadata, FileMetadataIndex, FileReaderOptions, ProjectedFileReader, ReaderProjection,
+};
 use lance_file::version::LanceFileVersion;
 use lance_file::{LanceEncodingsIo, determine_file_version};
 use lance_io::ReadBatchParams;
@@ -133,7 +135,7 @@ pub trait GenericFileReader: std::fmt::Debug + Send + Sync {
     fn projection(&self) -> &Arc<Schema>;
 
     /// Get storage statistics for this file (ignored by v1 reader)
-    fn storage_stats(&self) -> Vec<(u32, u64)>;
+    fn storage_stats(&self) -> Result<Vec<(u32, u64)>>;
 
     // Helper functions to fallback to the legacy implementation while we
     // slowly migrate functionality over to the generic reader
@@ -299,9 +301,9 @@ impl GenericFileReader for V1Reader {
         self.reader.len() as u32
     }
 
-    fn storage_stats(&self) -> Vec<(u32, u64)> {
+    fn storage_stats(&self) -> Result<Vec<(u32, u64)>> {
         // No-op for v1 files
-        Vec::new()
+        Ok(Vec::new())
     }
 
     fn clone_box(&self) -> Box<dyn GenericFileReader> {
@@ -328,7 +330,7 @@ mod v2_adapter {
 
     #[derive(Debug, Clone)]
     pub struct Reader {
-        reader: Arc<lance_file::reader::FileReader>,
+        reader: Arc<ProjectedFileReader>,
         projection: Arc<Schema>,
         field_id_to_column_idx: Arc<BTreeMap<u32, u32>>,
         default_priority: u32,
@@ -337,7 +339,7 @@ mod v2_adapter {
 
     impl Reader {
         pub fn new(
-            reader: Arc<lance_file::reader::FileReader>,
+            reader: Arc<ProjectedFileReader>,
             projection: Arc<Schema>,
             field_id_to_column_idx: Arc<BTreeMap<u32, u32>>,
             default_priority: u32,
@@ -363,7 +365,7 @@ mod v2_adapter {
         ) -> BoxFuture<'_, Result<ReadBatchTaskStream>> {
             async move {
                 let projection = ReaderProjection::from_field_ids(
-                    self.reader.metadata().version(),
+                    self.reader.version(),
                     projection.as_ref(),
                     self.field_id_to_column_idx.as_ref(),
                 )?;
@@ -393,7 +395,7 @@ mod v2_adapter {
         ) -> BoxFuture<'_, Result<ReadBatchTaskStream>> {
             async move {
                 let projection = ReaderProjection::from_field_ids(
-                    self.reader.metadata().version(),
+                    self.reader.version(),
                     projection.as_ref(),
                     self.field_id_to_column_idx.as_ref(),
                 )?;
@@ -422,7 +424,7 @@ mod v2_adapter {
         ) -> BoxFuture<'_, Result<ReadBatchTaskStream>> {
             async move {
                 let projection = ReaderProjection::from_field_ids(
-                    self.reader.metadata().version(),
+                    self.reader.version(),
                     projection.as_ref(),
                     self.field_id_to_column_idx.as_ref(),
                 )?;
@@ -454,7 +456,7 @@ mod v2_adapter {
             let indices = UInt32Array::from(indices.to_vec());
             async move {
                 let projection = ReaderProjection::from_field_ids(
-                    self.reader.metadata().version(),
+                    self.reader.version(),
                     projection.as_ref(),
                     self.field_id_to_column_idx.as_ref(),
                 )?;
@@ -487,8 +489,10 @@ mod v2_adapter {
             .boxed()
         }
 
-        fn storage_stats(&self) -> Vec<(u32, u64)> {
-            let file_statistics = self.reader.file_statistics();
+        fn storage_stats(&self) -> Result<Vec<(u32, u64)>> {
+            let file_statistics = self.reader.file_statistics().ok_or_else(|| {
+                Error::internal("storage_stats requires full file metadata".to_string())
+            })?;
             let column_idx_to_field_id = self
                 .field_id_to_column_idx
                 .iter()
@@ -505,7 +509,7 @@ mod v2_adapter {
                 }
                 stats.push((current_field_id, col_stats.size_bytes));
             }
-            stats
+            Ok(stats)
         }
 
         fn projection(&self) -> &Arc<Schema> {
@@ -514,7 +518,7 @@ mod v2_adapter {
 
         /// Return the number of rows in the file
         fn len(&self) -> u32 {
-            self.reader.metadata().num_rows as u32
+            self.reader.num_rows() as u32
         }
 
         fn clone_box(&self) -> Box<dyn GenericFileReader> {
@@ -614,9 +618,9 @@ impl GenericFileReader for NullReader {
         self.read_ranges_tasks(vec![0..num_rows].into(), batch_size, projection)
     }
 
-    fn storage_stats(&self) -> Vec<(u32, u64)> {
+    fn storage_stats(&self) -> Result<Vec<(u32, u64)>> {
         // No-op for null reader
-        Vec::new()
+        Ok(Vec::new())
     }
 
     fn projection(&self) -> &Arc<Schema> {
@@ -715,6 +719,12 @@ impl FragReadConfig {
     }
 }
 
+#[derive(Debug, Clone, Copy, PartialEq, Eq)]
+enum MetadataMode {
+    LazyAllowed,
+    Full,
+}
+
 impl FileFragment {
     /// Creates a new FileFragment.
     pub fn new(dataset: Arc<Dataset>, metadata: Fragment) -> Self {
@@ -839,13 +849,13 @@ impl FileFragment {
     ) -> Result<Vec<(u32, u64)>> {
         let mut stats = Vec::new();
         for reader in self
-            .open_readers(
+            .open_readers_with_full_metadata(
                 dataset_schema,
                 &FragReadConfig::default().with_scan_scheduler(scan_scheduler),
             )
             .await?
         {
-            stats.extend(reader.storage_stats());
+            stats.extend(reader.storage_stats()?);
         }
         Ok(stats)
     }
@@ -958,123 +968,225 @@ impl FileFragment {
         data_file.fields.first().copied().unwrap_or(0) as u32
     }
 
+    fn should_try_indexed_metadata(
+        data_file: &DataFile,
+        projection: &ReaderProjection,
+        file_version: LanceFileVersion,
+    ) -> bool {
+        if !ProjectedFileReader::supports_projection(projection, file_version) {
+            return false;
+        }
+        let total_columns = data_file
+            .column_indices
+            .iter()
+            .filter(|column_index| **column_index >= 0)
+            .count();
+        let selected_columns = projection.column_indices.len();
+        selected_columns.saturating_mul(4) < total_columns
+    }
+
     async fn open_reader(
         &self,
         data_file: &DataFile,
         projection: Option<&Schema>,
         read_config: &FragReadConfig,
     ) -> Result<Option<Box<dyn GenericFileReader>>> {
-        let full_schema = self.dataset.schema();
-        // The data file may contain fields that are not part of the dataset any longer, remove those
-        let data_file_schema = data_file.schema(full_schema);
-        let projection = projection.unwrap_or(full_schema);
-        // Also remove any fields that are not part of the user's provided projection
-        let schema_per_file = Arc::new(projection.intersection_ignore_types(&data_file_schema)?);
-
-        if data_file.is_legacy_file() {
-            let max_field_id = data_file.fields.iter().max().unwrap();
-            if !schema_per_file.fields.is_empty() {
+        self.open_reader_impl(
+            data_file,
+            projection,
+            read_config,
+            MetadataMode::LazyAllowed,
+        )
+        .await
+    }
+
+    async fn open_reader_with_full_metadata(
+        &self,
+        data_file: &DataFile,
+        projection: Option<&Schema>,
+        read_config: &FragReadConfig,
+    ) -> Result<Option<Box<dyn GenericFileReader>>> {
+        self.open_reader_impl(data_file, projection, read_config, MetadataMode::Full)
+            .await
+    }
+
+    fn open_reader_impl<'a>(
+        &'a self,
+        data_file: &'a DataFile,
+        projection: Option<&'a Schema>,
+        read_config: &'a FragReadConfig,
+        metadata_mode: MetadataMode,
+    ) -> BoxFuture<'a, Result<Option<Box<dyn GenericFileReader>>>> {
+        async move {
+            let full_schema = self.dataset.schema();
+            // The data file may contain fields that are not part of the dataset any longer, remove those
+            let data_file_schema = Arc::new(data_file.schema(full_schema));
+            let projection = projection.unwrap_or(full_schema);
+            // Also remove any fields that are not part of the user's provided projection
+            let schema_per_file =
+                Arc::new(projection.intersection_ignore_types(data_file_schema.as_ref())?);
+
+            if data_file.is_legacy_file() {
+                let max_field_id = data_file.fields.iter().max().unwrap();
+                if !schema_per_file.fields.is_empty() {
+                    let path = self
+                        .dataset
+                        .data_file_dir(data_file)?
+                        .join(data_file.path.as_str());
+                    let object_store = self.dataset.object_store_for_data_file(data_file).await?;
+                    let field_id_offset = Self::get_field_id_offset(data_file);
+                    let reader = PreviousFileReader::try_new_with_fragment_id(
+                        &object_store,
+                        &path,
+                        self.schema().clone(),
+                        self.id() as u32,
+                        field_id_offset as i32,
+                        *max_field_id,
+                        Some(&self.dataset.metadata_cache.file_metadata_cache(&path)),
+                    )
+                    .await?;
+                    let initialized_schema = reader.schema().project_by_schema(
+                        schema_per_file.as_ref(),
+                        OnMissing::Error,
+                        OnTypeMismatch::Error,
+                    )?;
+                    let reader = V1Reader::new(reader, Arc::new(initialized_schema));
+                    let reader: Box<dyn GenericFileReader> = Box::new(reader);
+                    Ok(Some(reader))
+                } else {
+                    Ok(None)
+                }
+            } else if schema_per_file.fields.is_empty() {
+                Ok(None)
+            } else {
                 let path = self
                     .dataset
                     .data_file_dir(data_file)?
                     .join(data_file.path.as_str());
-                let object_store = self.dataset.object_store_for_data_file(data_file).await?;
-                let field_id_offset = Self::get_field_id_offset(data_file);
-                let reader = PreviousFileReader::try_new_with_fragment_id(
-                    &object_store,
-                    &path,
-                    self.schema().clone(),
-                    self.id() as u32,
-                    field_id_offset as i32,
-                    *max_field_id,
-                    Some(&self.dataset.metadata_cache.file_metadata_cache(&path)),
-                )
-                .await?;
-                let initialized_schema = reader.schema().project_by_schema(
+                let (store_scheduler, reader_priority) = if let Some(base_id) = data_file.base_id {
+                    // TODO: make object stores for non-default bases reuse the same scan scheduler
+                    //  currently we always create a new one
+                    let object_store = self.dataset.object_store(Some(base_id)).await?;
+                    let config = SchedulerConfig::max_bandwidth(&object_store);
+                    (
+                        ScanScheduler::new(object_store, config),
+                        read_config.reader_priority.unwrap_or(0),
+                    )
+                } else if let Some(scan_scheduler) = read_config.scan_scheduler.as_ref() {
+                    (
+                        scan_scheduler.clone(),
+                        read_config.reader_priority.unwrap_or(0),
+                    )
+                } else {
+                    (
+                        ScanScheduler::new(
+                            self.dataset.object_store.clone(),
+                            SchedulerConfig::max_bandwidth(&self.dataset.object_store),
+                        ),
+                        0,
+                    )
+                };
+                let file_scheduler = store_scheduler
+                    .open_file_with_priority(
+                        &path,
+                        reader_priority as u64,
+                        &data_file.file_size_bytes,
+                    )
+                    .await?;
+                let path = file_scheduler.reader().path().clone();
+                let metadata_cache = self.dataset.metadata_cache.file_metadata_cache(&path);
+                let field_id_to_column_idx = Arc::new(BTreeMap::from_iter(
+                    data_file
+                        .fields
+                        .iter()
+                        .copied()
+                        .zip(data_file.column_indices.iter().copied())
+                        .filter_map(|(field_id, column_index)| {
+                            if column_index < 0 {
+                                None
+                            } else {
+                                Some((field_id as u32, column_index as u32))
+                            }
+                        }),
+                ));
+                let file_version = LanceFileVersion::try_from_major_minor(
+                    data_file.file_major_version,
+                    data_file.file_minor_version,
+                )?;
+                let reader_projection = ReaderProjection::from_field_ids(
+                    file_version,
                     schema_per_file.as_ref(),
-                    OnMissing::Error,
-                    OnTypeMismatch::Error,
+                    field_id_to_column_idx.as_ref(),
                 )?;
-                let reader = V1Reader::new(reader, Arc::new(initialized_schema));
-                Ok(Some(Box::new(reader)))
-            } else {
-                Ok(None)
+                let file_reader_options = read_config
+                    .file_reader_options
+                    .clone()
+                    .or_else(|| self.dataset.file_reader_options.clone())
+                    .unwrap_or_default();
+                let metadata_index = if metadata_mode == MetadataMode::LazyAllowed
+                    && Self::should_try_indexed_metadata(
+                        data_file,
+                        &reader_projection,
+                        file_version,
+                    ) {
+                    let known_schema = self
+                        .metadata
+                        .physical_rows
+                        .map(|num_rows| (data_file_schema.clone(), num_rows as u64));
+                    let metadata_index = self
+                        .get_file_metadata_index(&file_scheduler, known_schema)
+                        .await?;
+                    if (reader_projection.column_indices.len() as u32).saturating_mul(4)
+                        < metadata_index.num_columns()
+                    {
+                        Some(metadata_index)
+                    } else {
+                        None
+                    }
+                } else {
+                    None
+                };
+
+                let encodings_io = Arc::new(
+                    LanceEncodingsIo::new(file_scheduler.clone())
+                        .with_read_chunk_size(file_reader_options.read_chunk_size),
+                );
+                let reader = if let Some(metadata_index) = metadata_index {
+                    ProjectedFileReader::try_open_with_metadata_index(
+                        encodings_io.clone(),
+                        path.clone(),
+                        Some(reader_projection.clone()),
+                        Arc::<DecoderPlugins>::default(),
+                        metadata_index,
+                        &metadata_cache,
+                        file_reader_options.clone(),
+                    )
+                    .await?
+                } else {
+                    let file_metadata = self.get_file_metadata(&file_scheduler).await?;
+                    ProjectedFileReader::try_open_with_file_metadata(
+                        encodings_io,
+                        path.clone(),
+                        None,
+                        Arc::<DecoderPlugins>::default(),
+                        file_metadata,
+                        &metadata_cache,
+                        file_reader_options,
+                    )
+                    .await?
+                };
+                let reader = v2_adapter::Reader::new(
+                    Arc::new(reader),
+                    schema_per_file,
+                    field_id_to_column_idx,
+                    reader_priority,
+                    file_scheduler,
+                );
+                let reader: Box<dyn GenericFileReader> = Box::new(reader);
+                Ok(Some(reader))
             }
-        } else if schema_per_file.fields.is_empty() {
-            Ok(None)
-        } else {
-            let path = self
-                .dataset
-                .data_file_dir(data_file)?
-                .join(data_file.path.as_str());
-            let (store_scheduler, reader_priority) = if let Some(base_id) = data_file.base_id {
-                // TODO: make object stores for non-default bases reuse the same scan scheduler
-                //  currently we always create a new one
-                let object_store = self.dataset.object_store(Some(base_id)).await?;
-                let config = SchedulerConfig::max_bandwidth(&object_store);
-                (
-                    ScanScheduler::new(object_store, config),
-                    read_config.reader_priority.unwrap_or(0),
-                )
-            } else if let Some(scan_scheduler) = read_config.scan_scheduler.as_ref() {
-                (
-                    scan_scheduler.clone(),
-                    read_config.reader_priority.unwrap_or(0),
-                )
-            } else {
-                (
-                    ScanScheduler::new(
-                        self.dataset.object_store.clone(),
-                        SchedulerConfig::max_bandwidth(&self.dataset.object_store),
-                    ),
-                    0,
-                )
-            };
-            let file_scheduler = store_scheduler
-                .open_file_with_priority(&path, reader_priority as u64, &data_file.file_size_bytes)
-                .await?;
-            let file_metadata = self.get_file_metadata(&file_scheduler).await?;
-            let path = file_scheduler.reader().path().clone();
-            let metadata_cache = self.dataset.metadata_cache.file_metadata_cache(&path);
-            let reader = Arc::new(
-                lance_file::reader::FileReader::try_open_with_file_metadata(
-                    Arc::new(LanceEncodingsIo::new(file_scheduler.clone())),
-                    path,
-                    None,
-                    Arc::<DecoderPlugins>::default(),
-                    file_metadata,
-                    &metadata_cache,
-                    read_config
-                        .file_reader_options
-                        .clone()
-                        .or_else(|| self.dataset.file_reader_options.clone())
-                        .unwrap_or_default(),
-                )
-                .await?,
-            );
-            let field_id_to_column_idx = Arc::new(BTreeMap::from_iter(
-                data_file
-                    .fields
-                    .iter()
-                    .copied()
-                    .zip(data_file.column_indices.iter().copied())
-                    .filter_map(|(field_id, column_index)| {
-                        if column_index < 0 {
-                            None
-                        } else {
-                            Some((field_id as u32, column_index as u32))
-                        }
-                    }),
-            ));
-            let reader = v2_adapter::Reader::new(
-                reader,
-                schema_per_file,
-                field_id_to_column_idx,
-                reader_priority,
-                file_scheduler,
-            );
-            Ok(Some(Box::new(reader)))
         }
+        .boxed()
     }
 
     async fn open_readers(
@@ -1082,35 +1194,68 @@ impl FileFragment {
         projection: &Schema,
         read_config: &FragReadConfig,
     ) -> Result<Vec<Box<dyn GenericFileReader>>> {
-        let mut opened_files = vec![];
-        for data_file in &self.metadata.files {
-            if let Some(reader) = self
-                .open_reader(data_file, Some(projection), read_config)
-                .await?
-            {
-                opened_files.push(reader);
+        self.open_readers_impl(projection, read_config, MetadataMode::LazyAllowed)
+            .await
+    }
+
+    async fn open_readers_with_full_metadata(
+        &self,
+        projection: &Schema,
+        read_config: &FragReadConfig,
+    ) -> Result<Vec<Box<dyn GenericFileReader>>> {
+        self.open_readers_impl(projection, read_config, MetadataMode::Full)
+            .await
+    }
+
+    fn open_readers_impl<'a>(
+        &'a self,
+        projection: &'a Schema,
+        read_config: &'a FragReadConfig,
+        metadata_mode: MetadataMode,
+    ) -> BoxFuture<'a, Result<Vec<Box<dyn GenericFileReader>>>> {
+        async move {
+            let mut opened_files = vec![];
+            for data_file in &self.metadata.files {
+                let reader = match metadata_mode {
+                    MetadataMode::LazyAllowed => {
+                        self.open_reader(data_file, Some(projection), read_config)
+                            .await?
+                    }
+                    MetadataMode::Full => {
+                        self.open_reader_with_full_metadata(
+                            data_file,
+                            Some(projection),
+                            read_config,
+                        )
+                        .await?
+                    }
+                };
+                if let Some(reader) = reader {
+                    opened_files.push(reader);
+                }
             }
-        }
 
-        // This should return immediately on modern datasets.  Need to use physical_rows because
-        // deletions will be applied later
-        let num_rows = self.physical_rows().await?;
+            // This should return immediately on modern datasets.  Need to use physical_rows because
+            // deletions will be applied later
+            let num_rows = self.physical_rows().await?;
 
-        // Check if there are any fields that are not in any data files
-        let field_ids_in_files = opened_files
-            .iter()
-            .flat_map(|r| r.projection().fields_pre_order().map(|f| f.id))
-            .filter(|id| *id >= 0)
-            .collect::<HashSet<_>>();
-        let mut missing_fields = projection.field_ids();
-        missing_fields.retain(|f| !field_ids_in_files.contains(f) && *f >= 0);
-        if !missing_fields.is_empty() {
-            let missing_projection = projection.project_by_ids(&missing_fields, true);
-            let null_reader = NullReader::new(Arc::new(missing_projection), num_rows as u32);
-            opened_files.push(Box::new(null_reader));
-        }
+            // Check if there are any fields that are not in any data files
+            let field_ids_in_files = opened_files
+                .iter()
+                .flat_map(|r| r.projection().fields_pre_order().map(|f| f.id))
+                .filter(|id| *id >= 0)
+                .collect::<HashSet<_>>();
+            let mut missing_fields = projection.field_ids();
+            missing_fields.retain(|f| !field_ids_in_files.contains(f) && *f >= 0);
+            if !missing_fields.is_empty() {
+                let missing_projection = projection.project_by_ids(&missing_fields, true);
+                let null_reader = NullReader::new(Arc::new(missing_projection), num_rows as u32);
+                opened_files.push(Box::new(null_reader));
+            }
 
-        Ok(opened_files)
+            Ok(opened_files)
+        }
+        .boxed()
     }
 
     /// Count the rows in this fragment.
@@ -1461,6 +1606,32 @@ impl FileFragment {
         Ok(file_metadata)
     }
 
+    async fn get_file_metadata_index(
+        &self,
+        file_scheduler: &FileScheduler,
+        known_schema: Option<(Arc<Schema>, u64)>,
+    ) -> Result<Arc<FileMetadataIndex>> {
+        let path = file_scheduler.reader().path();
+        let cache = self.dataset.metadata_cache.file_metadata_cache(path);
+
+        let metadata_index = cache
+            .get_or_insert_with_key(FileMetadataIndexCacheKey, || async {
+                let metadata_index = if let Some((file_schema, num_rows)) = known_schema {
+                    lance_file::reader::FileReader::read_metadata_index_with_schema(
+                        file_scheduler,
+                        file_schema,
+                        num_rows,
+                    )
+                    .await?
+                } else {
+                    lance_file::reader::FileReader::read_metadata_index(file_scheduler).await?
+                };
+                Ok(metadata_index)
+            })
+            .await?;
+        Ok(metadata_index)
+    }
+
     /// Take rows based on internal local row offsets
     ///
     /// If the row offsets are out-of-bounds, this will return an error. But if the
@@ -1983,6 +2154,21 @@ impl CacheKey for FileMetadataCacheKey {
     }
 }
 
+#[derive(Debug, Clone)]
+struct FileMetadataIndexCacheKey;
+
+impl CacheKey for FileMetadataIndexCacheKey {
+    type ValueType = FileMetadataIndex;
+
+    fn key(&self) -> std::borrow::Cow<'_, str> {
+        "metadata_index".into()
+    }
+
+    fn type_name() -> &'static str {
+        "FileMetadataIndex"
+    }
+}
+
 impl From<FileFragment> for Fragment {
     fn from(fragment: FileFragment) -> Self {
         fragment.metadata
@@ -4060,6 +4246,105 @@ mod tests {
         );
     }
 
+    #[tokio::test]
+    async fn test_lazy_column_metadata_scan_reads_less_than_full_projection() {
+        let num_columns = 512;
+        let rows_per_batch = 100;
+        let num_batches = 10;
+        let schema = Arc::new(ArrowSchema::new(
+            (0..num_columns)
+                .map(|i| ArrowField::new(format!("col_{i}"), DataType::Int32, true))
+                .collect::<Vec<_>>(),
+        ));
+        let batches = (0..num_batches)
+            .map(|batch_idx| {
+                let columns = (0..num_columns)
+                    .map(|column_idx| {
+                        Arc::new(Int32Array::from_iter_values((0..rows_per_batch).map(
+                            |row_idx| (batch_idx * rows_per_batch + row_idx) as i32 + column_idx,
+                        ))) as ArrayRef
+                    })
+                    .collect::<Vec<_>>();
+                RecordBatch::try_new(schema.clone(), columns).unwrap()
+            })
+            .collect::<Vec<_>>();
+
+        let test_dir = TempStrDir::default();
+        let write_params = WriteParams {
+            max_rows_per_file: rows_per_batch * num_batches,
+            max_rows_per_group: rows_per_batch,
+            data_storage_version: Some(LanceFileVersion::V2_1),
+            ..Default::default()
+        };
+        let reader = RecordBatchIterator::new(batches.into_iter().map(Ok), schema);
+        let dataset = Dataset::write(reader, &test_dir, Some(write_params))
+            .await
+            .unwrap();
+
+        let projection = dataset.schema().project(&["col_0"]).unwrap();
+        let fragment = dataset.get_fragment(0).unwrap();
+
+        dataset.object_store.as_ref().io_stats_incremental();
+        let narrow_reader = fragment
+            .open(&projection, FragReadConfig::default())
+            .await
+            .unwrap();
+        let mut empty_narrow_stream = narrow_reader.take_range(0..0, 1024).await.unwrap();
+        assert!(empty_narrow_stream.next().await.is_none());
+        let narrow_metadata_stats = dataset.object_store.as_ref().io_stats_incremental();
+        assert!(
+            narrow_metadata_stats.read_iops <= 3,
+            "expected lazy metadata open to skip the schema buffer read, iops={}, bytes={}",
+            narrow_metadata_stats.read_iops,
+            narrow_metadata_stats.read_bytes
+        );
+
+        let full_projection = dataset.schema().clone();
+        let full_reader = fragment
+            .open(&full_projection, FragReadConfig::default())
+            .await
+            .unwrap();
+        let mut empty_full_stream = full_reader.take_range(0..0, 1024).await.unwrap();
+        assert!(empty_full_stream.next().await.is_none());
+        let full_metadata_stats = dataset.object_store.as_ref().io_stats_incremental();
+
+        assert!(
+            full_metadata_stats.read_bytes > narrow_metadata_stats.read_bytes * 4,
+            "expected narrow lazy metadata read to fetch much less than full metadata, narrow={} bytes, full={} bytes",
+            narrow_metadata_stats.read_bytes,
+            full_metadata_stats.read_bytes
+        );
+
+        let mut narrow_scan = dataset.scan();
+        let narrow_batch = narrow_scan
+            .project(&["col_0"])
+            .unwrap()
+            .try_into_batch()
+            .await
+            .unwrap();
+        assert_eq!(narrow_batch.num_columns(), 1);
+        assert_eq!(narrow_batch.num_rows(), rows_per_batch * num_batches);
+
+        let taken = fragment.take(&[0, 777, 999], &projection).await.unwrap();
+        let taken_values = taken
+            .column(0)
+            .as_any()
+            .downcast_ref::<Int32Array>()
+            .unwrap();
+        assert_eq!(taken_values.values(), &[0, 777, 999]);
+
+        let projected_readers = fragment
+            .open_readers(&projection, &FragReadConfig::default())
+            .await
+            .unwrap();
+        let err = projected_readers[0].storage_stats().unwrap_err();
+        assert!(
+            err.to_string()
+                .contains("storage_stats requires full file metadata"),
+            "expected storage_stats to reject projected metadata, got {err:?}"
+        );
+    }
+
     #[tokio::test]
     async fn test_iops_read_small() {
         // Create a file that has 8 columns.
diff --git a/rust/lance/src/dataset/statistics.rs b/rust/lance/src/dataset/statistics.rs
index 9e52c9653e1..4719a639068 100644
--- a/rust/lance/src/dataset/statistics.rs
+++ b/rust/lance/src/dataset/statistics.rs
@@ -82,3 +82,69 @@ impl DatasetStatisticsExt for Dataset {
         })
     }
 }
+
+#[cfg(test)]
+mod tests {
+    use std::sync::Arc;
+
+    use arrow_array::{ArrayRef, Int32Array, RecordBatch, RecordBatchIterator};
+    use arrow_schema::{DataType, Field as ArrowField, Schema as ArrowSchema};
+    use lance_core::utils::tempfile::TempStrDir;
+    use lance_file::version::LanceFileVersion;
+
+    use crate::dataset::WriteParams;
+
+    use super::*;
+
+    #[tokio::test]
+    async fn test_calculate_data_stats_after_dropping_wide_dataset_columns() {
+        let num_columns = 64;
+        let num_rows = 128;
+        let schema = Arc::new(ArrowSchema::new(
+            (0..num_columns)
+                .map(|idx| ArrowField::new(format!("col_{idx}"), DataType::Int32, true))
+                .collect::<Vec<_>>(),
+        ));
+        let batch = RecordBatch::try_new(
+            schema.clone(),
+            (0..num_columns)
+                .map(|column_idx| {
+                    Arc::new(Int32Array::from_iter_values(
+                        (0..num_rows).map(|row_idx| row_idx + column_idx),
+                    )) as ArrayRef
+                })
+                .collect::<Vec<_>>(),
+        )
+        .unwrap();
+
+        let test_dir = TempStrDir::default();
+        let reader = RecordBatchIterator::new(vec![Ok(batch)], schema);
+        let mut dataset = Dataset::write(
+            reader,
+            &test_dir,
+            Some(WriteParams {
+                data_storage_version: Some(LanceFileVersion::V2_1),
+                ..Default::default()
+            }),
+        )
+        .await
+        .unwrap();
+
+        let columns_to_drop = (1..num_columns)
+            .map(|idx| format!("col_{idx}"))
+            .collect::<Vec<_>>();
+        let column_refs = columns_to_drop
+            .iter()
+            .map(String::as_str)
+            .collect::<Vec<_>>();
+        dataset.drop_columns(&column_refs).await.unwrap();
+
+        let stats = Arc::new(dataset).calculate_data_stats().await.unwrap();
+        assert_eq!(stats.fields.len(), 1);
+        assert_eq!(stats.fields[0].id, 0);
+        assert!(
+            stats.fields[0].bytes_on_disk > 0,
+            "bytes_on_disk should include the remaining column after drop_columns"
+        );
+    }
+}
diff --git a/rust/lance/src/index/create.rs b/rust/lance/src/index/create.rs
index bbb055463dc..19f79f8dd66 100644
--- a/rust/lance/src/index/create.rs
+++ b/rust/lance/src/index/create.rs
@@ -18,7 +18,7 @@ use crate::{
         vector_index_details, vector_index_details_default,
     },
 };
-use futures::future::BoxFuture;
+use futures::{FutureExt, future::BoxFuture};
 use lance_core::datatypes::format_field_path;
 use lance_index::progress::{IndexBuildProgress, NoopIndexBuildProgress};
 use lance_index::{IndexParams, IndexType, scalar::CreatedIndex};
@@ -133,6 +133,11 @@ impl<'a> CreateIndexBuilder<'a> {
 
     #[instrument(skip_all)]
     pub async fn execute_uncommitted(&mut self) -> Result<IndexMetadata> {
+        self.execute_uncommitted_impl().await
+    }
+
+    fn execute_uncommitted_impl(&mut self) -> BoxFuture<'_, Result<IndexMetadata>> {
+        async move {
         if self.columns.len() != 1 {
             return Err(Error::index(
                 "Only support building index on 1 column at the moment".to_string(),
@@ -481,6 +486,8 @@ impl<'a> CreateIndexBuilder<'a> {
             base_id: None,
             files: Some(created_index.files),
         })
+        }
+        .boxed()
     }
 
     #[instrument(skip_all)]