Part 1, Read transforms via expressions: Just compute the expression and return it.

ffi/src/scan.rs

@@ -230,7 +230,7 @@ fn kernel_scan_data_next_impl(
         .data
         .lock()
         .map_err(|_| Error::generic("poisoned mutex"))?;
-    if let Some((data, sel_vec)) = data.next().transpose()? {
+    if let Some((data, sel_vec, _transforms)) = data.next().transpose()? {
         let bool_slice = KernelBoolSlice::from(sel_vec);
         (engine_visitor)(engine_context, data.into(), bool_slice);
         Ok(true)

kernel/examples/inspect-table/src/main.rs

@@ -209,7 +209,7 @@ fn try_main() -> DeltaResult<()> {
             let scan = ScanBuilder::new(snapshot).build()?;
             let scan_data = scan.scan_data(&engine)?;
             for res in scan_data {
-                let (data, vector) = res?;
+                let (data, vector, _transforms) = res?;
                 delta_kernel::scan::state::visit_scan_files(
                     data.as_ref(),
                     &vector,

kernel/examples/read-table-multi-threaded/src/main.rs

@@ -210,7 +210,7 @@ fn try_main() -> DeltaResult<()> {
     drop(record_batch_tx);
 
     for res in scan_data {
-        let (data, vector) = res?;
+        let (data, vector, _transforms) = res?;
         scan_file_tx = delta_kernel::scan::state::visit_scan_files(
             data.as_ref(),
             &vector,

kernel/src/engine_data.rs

@@ -129,7 +129,9 @@ pub trait TypedGetData<'a, T> {
     fn get_opt(&'a self, row_index: usize, field_name: &str) -> DeltaResult<Option<T>>;
     fn get(&'a self, row_index: usize, field_name: &str) -> DeltaResult<T> {
         let val = self.get_opt(row_index, field_name)?;
-        val.ok_or_else(|| Error::MissingData(format!("Data missing for field {field_name}")))
+        val.ok_or_else(|| {
+            Error::MissingData(format!("Data missing for field {field_name}")).with_backtrace()
+        })
     }
 }
 

kernel/src/scan/log_replay.rs

@@ -1,15 +1,16 @@
 use std::clone::Clone;
-use std::collections::HashSet;
+use std::collections::{HashMap, HashSet};
 use std::sync::{Arc, LazyLock};
 
+use itertools::Itertools;
 use tracing::debug;
 
 use super::data_skipping::DataSkippingFilter;
-use super::ScanData;
+use super::{ScanData, Transform};
 use crate::actions::get_log_add_schema;
 use crate::engine_data::{GetData, RowVisitor, TypedGetData as _};
 use crate::expressions::{column_expr, column_name, ColumnName, Expression, ExpressionRef};
-use crate::scan::DeletionVectorDescriptor;
+use crate::scan::{DeletionVectorDescriptor, TransformExpr};
 use crate::schema::{ColumnNamesAndTypes, DataType, MapType, SchemaRef, StructField, StructType};
 use crate::utils::require;
 use crate::{DeltaResult, Engine, EngineData, Error, ExpressionEvaluator};
@@ -44,12 +45,17 @@ struct LogReplayScanner {
 struct AddRemoveDedupVisitor<'seen> {
     seen: &'seen mut HashSet<FileActionKey>,
     selection_vector: Vec<bool>,
+    logical_schema: SchemaRef,
+    transform: Option<Arc<Transform>>,
+    transforms: HashMap<usize, Expression>,
     is_log_batch: bool,
 }
 
 impl AddRemoveDedupVisitor<'_> {
     /// Checks if log replay already processed this logical file (in which case the current action
     /// should be ignored). If not already seen, register it so we can recognize future duplicates.
+    /// Returns `true` if we have seen the file and should ignore it, `false` if we have not seen it
+    /// and should process it.
     fn check_and_record_seen(&mut self, key: FileActionKey) -> bool {
         // Note: each (add.path + add.dv_unique_id()) pair has a
         // unique Add + Remove pair in the log. For example:
@@ -83,11 +89,11 @@ impl AddRemoveDedupVisitor<'_> {
         // have a remove with a path at index 4. In either case, extract the three dv getters at
         // indexes that immediately follow a valid path index.
         let (path, dv_getters, is_add) = if let Some(path) = getters[0].get_str(i, "add.path")? {
-            (path, &getters[1..4], true)
+            (path, &getters[2..5], true)
         } else if !self.is_log_batch {
             return Ok(false);
-        } else if let Some(path) = getters[4].get_opt(i, "remove.path")? {
-            (path, &getters[5..8], false)
+        } else if let Some(path) = getters[5].get_opt(i, "remove.path")? {
+            (path, &getters[6..9], false)
         } else {
             return Ok(false);
         };
@@ -103,7 +109,33 @@ impl AddRemoveDedupVisitor<'_> {
 
         // Process both adds and removes, but only return not already-seen adds
         let file_key = FileActionKey::new(path, dv_unique_id);
-        Ok(!self.check_and_record_seen(file_key) && is_add)
+        let have_seen = self.check_and_record_seen(file_key);
+        if is_add && !have_seen {
+            // compute transform here
+            if let Some(ref transform) = self.transform {
+                let partition_values: HashMap<_, _> = getters[1].get(i, "add.partitionValues")?;
+                let transforms = transform
+                    .iter()
+                    .map(|transform_expr| match transform_expr {
+                        TransformExpr::Partition(field_idx) => {
+                            let field = self.logical_schema.fields.get_index(*field_idx);
+                            let Some((_, field)) = field else {
+                                return Err(Error::generic(
+                                    "logical schema did not contain expected field, can't transform data",
+                                ));
+                            };
+                            let name = field.physical_name();
+                            let value_expression =
+                                super::parse_partition_value(partition_values.get(name), field.data_type())?;
+                            Ok(value_expression.into())
+                        }
+                        TransformExpr::Static(field_expr) => Ok(field_expr.clone()),
+                    })
+                    .try_collect()?;
+                self.transforms.insert(i, Expression::Struct(transforms));
+            }
+        }
+        Ok(!have_seen && is_add)
     }
 }
 
@@ -113,8 +145,10 @@ impl RowVisitor for AddRemoveDedupVisitor<'_> {
         static NAMES_AND_TYPES: LazyLock<ColumnNamesAndTypes> = LazyLock::new(|| {
             const STRING: DataType = DataType::STRING;
             const INTEGER: DataType = DataType::INTEGER;
+            let ss_map: DataType = MapType::new(STRING, STRING, true).into();
             let types_and_names = vec![
                 (STRING, column_name!("add.path")),
+                (ss_map, column_name!("add.partitionValues")),
                 (STRING, column_name!("add.deletionVector.storageType")),
                 (STRING, column_name!("add.deletionVector.pathOrInlineDv")),
                 (INTEGER, column_name!("add.deletionVector.offset")),
@@ -132,12 +166,12 @@ impl RowVisitor for AddRemoveDedupVisitor<'_> {
         } else {
             // All checkpoint actions are already reconciled and Remove actions in checkpoint files
             // only serve as tombstones for vacuum jobs. So we only need to examine the adds here.
-            (&names[..4], &types[..4])
+            (&names[..5], &types[..5])
         }
     }
 
     fn visit<'a>(&mut self, row_count: usize, getters: &[&'a dyn GetData<'a>]) -> DeltaResult<()> {
-        let expected_getters = if self.is_log_batch { 8 } else { 4 };
+        let expected_getters = if self.is_log_batch { 9 } else { 5 };
         require!(
             getters.len() == expected_getters,
             Error::InternalError(format!(
@@ -207,6 +241,8 @@ impl LogReplayScanner {
         &mut self,
         add_transform: &dyn ExpressionEvaluator,
         actions: &dyn EngineData,
+        logical_schema: SchemaRef,
+        transform: Option<Arc<Transform>>,
         is_log_batch: bool,
     ) -> DeltaResult<ScanData> {
         // Apply data skipping to get back a selection vector for actions that passed skipping. We
@@ -220,24 +256,29 @@ impl LogReplayScanner {
         let mut visitor = AddRemoveDedupVisitor {
             seen: &mut self.seen,
             selection_vector,
+            logical_schema,
+            transform,
+            transforms: HashMap::new(),
             is_log_batch,
         };
         visitor.visit_rows_of(actions)?;
 
         // TODO: Teach expression eval to respect the selection vector we just computed so carefully!
         let selection_vector = visitor.selection_vector;
         let result = add_transform.evaluate(actions)?;
-        Ok((result, selection_vector))
+        Ok((result, selection_vector, visitor.transforms))
     }
 }
 
 /// Given an iterator of (engine_data, bool) tuples and a predicate, returns an iterator of
 /// `(engine_data, selection_vec)`. Each row that is selected in the returned `engine_data` _must_
 /// be processed to complete the scan. Non-selected rows _must_ be ignored. The boolean flag
 /// indicates whether the record batch is a log or checkpoint batch.
-pub fn scan_action_iter(
+pub(crate) fn scan_action_iter(
     engine: &dyn Engine,
     action_iter: impl Iterator<Item = DeltaResult<(Box<dyn EngineData>, bool)>>,
+    logical_schema: SchemaRef,
+    transform: Option<Arc<Transform>>,
     physical_predicate: Option<(ExpressionRef, SchemaRef)>,
 ) -> impl Iterator<Item = DeltaResult<ScanData>> {
     let mut log_scanner = LogReplayScanner::new(engine, physical_predicate);
@@ -249,20 +290,39 @@ pub fn scan_action_iter(
     action_iter
         .map(move |action_res| {
             let (batch, is_log_batch) = action_res?;
-            log_scanner.process_scan_batch(add_transform.as_ref(), batch.as_ref(), is_log_batch)
+            log_scanner.process_scan_batch(
+                add_transform.as_ref(),
+                batch.as_ref(),
+                logical_schema.clone(),
+                transform.clone(),
+                is_log_batch,
+            )
         })
-        .filter(|res| res.as_ref().map_or(true, |(_, sv)| sv.contains(&true)))
+        .filter(|res| res.as_ref().map_or(true, |(_, sv, _)| sv.contains(&true)))
 }
 
 #[cfg(test)]
 mod tests {
-    use std::collections::HashMap;
+    use std::{collections::HashMap, sync::Arc};
 
+    use crate::expressions::{column_name, Scalar};
     use crate::scan::{
+        get_state_info,
         state::{DvInfo, Stats},
-        test_utils::{add_batch_simple, add_batch_with_remove, run_with_validate_callback},
+        test_utils::{
+            add_batch_simple, add_batch_with_partition_col, add_batch_with_remove,
+            run_with_validate_callback,
+        },
+        Scan,
+    };
+    use crate::{
+        engine::sync::SyncEngine,
+        schema::{DataType, SchemaRef, StructField, StructType},
+        Expression,
     };
 
+    use super::scan_action_iter;
+
     // dv-info is more complex to validate, we validate that works in the test for visit_scan_files
     // in state.rs
     fn validate_simple(
@@ -288,6 +348,8 @@ mod tests {
     fn test_scan_action_iter() {
         run_with_validate_callback(
             vec![add_batch_simple()],
+            None, // not testing schema
+            None, // not testing transform
             &[true, false],
             (),
             validate_simple,
@@ -298,9 +360,76 @@ mod tests {
     fn test_scan_action_iter_with_remove() {
         run_with_validate_callback(
             vec![add_batch_with_remove()],
+            None, // not testing schema
+            None, // not testing transform
             &[false, false, true, false],
             (),
             validate_simple,
         );
     }
+
+    #[test]
+    fn test_no_transforms() {
+        let batch = vec![add_batch_simple()];
+        let logical_schema = Arc::new(crate::schema::StructType::new(vec![]));
+        let iter = scan_action_iter(
+            &SyncEngine::new(),
+            batch.into_iter().map(|batch| Ok((batch as _, true))),
+            logical_schema,
+            None,
+            None,
+        );
+        for res in iter {
+            let (_batch, _sel, transforms) = res.unwrap();
+            assert!(transforms.is_empty(), "Should have no transforms");
+        }
+    }
+
+    #[test]
+    fn test_simple_transform() {
+        let schema: SchemaRef = Arc::new(StructType::new([
+            StructField::new("value", DataType::INTEGER, true),
+            StructField::new("date", DataType::DATE, true),
+        ]));
+        let partition_cols = ["date".to_string()];
+        let state_info = get_state_info(schema.as_ref(), &partition_cols).unwrap();
+        let static_transform = Some(Arc::new(Scan::get_static_transform(&state_info.all_fields)));
+        let batch = vec![add_batch_with_partition_col()];
+        let iter = scan_action_iter(
+            &SyncEngine::new(),
+            batch.into_iter().map(|batch| Ok((batch as _, true))),
+            schema,
+            static_transform,
+            None,
+        );
+
+        fn validate_transform(transform: Option<&Expression>, expected_date_offset: i32) {
+            assert!(transform.is_some());
+            if let Expression::Struct(inner) = transform.unwrap() {
+                if let Expression::Column(ref name) = inner[0] {
+                    assert_eq!(name, &column_name!("value"), "First col should be 'value'");
+                } else {
+                    panic!("Expected first expression to be a column");
+                }
+                if let Expression::Literal(ref scalar) = inner[1] {
+                    assert_eq!(
+                        scalar,
+                        &Scalar::Date(expected_date_offset),
+                        "Didn't get expected date offset"
+                    );
+                } else {
+                    panic!("Expected second expression to be a literal");
+                }
+            } else {
+                panic!("Transform should always be a struct expr");
+            }
+        }
+
+        for res in iter {
+            let (_batch, _sel, transforms) = res.unwrap();
+            assert_eq!(transforms.len(), 2, "Should have two transforms");
+            validate_transform(transforms.get(&0), 17511);
+            validate_transform(transforms.get(&1), 17510);
+        }
+    }
 }