perf: coalesce ids before executing take

python/python/benchmarks/test_search.py

@@ -58,6 +58,7 @@ def create_table(num_rows, offset) -> pa.Table:
             "vector": vectors,
             "filterable": filterable,
             "category": categories,
+            "category_no_index": categories,
             "genres": genres,
         }
     )
@@ -77,9 +78,11 @@ def create_base_dataset(data_dir: Path) -> lance.LanceDataset:
         rows_remaining -= next_batch_length
         table = create_table(next_batch_length, offset)
         if offset == 0:
-            dataset = lance.write_dataset(table, tmp_path)
+            dataset = lance.write_dataset(table, tmp_path, use_legacy_format=False)
         else:
-            dataset = lance.write_dataset(table, tmp_path, mode="append")
+            dataset = lance.write_dataset(
+                table, tmp_path, mode="append", use_legacy_format=False
+            )
         offset += next_batch_length
 
     dataset.create_index(
@@ -479,3 +482,26 @@ def test_label_list_index_prefilter(test_dataset, benchmark, filter: str):
             prefilter=True,
             filter=filter,
         )
+
+
+@pytest.mark.benchmark(group="late_materialization")
+@pytest.mark.parametrize(
+    "use_index",
+    (False, True),
+    ids=["no_index", "with_index"],
+)
+def test_late_materialization(test_dataset, benchmark, use_index):
+    column = "category" if use_index else "category_no_index"
+    print(
+        test_dataset.scanner(
+            columns=["vector"],
+            filter=f"{column} = 0",
+            batch_size=32,
+        ).explain_plan(True)
+    )
+    benchmark(
+        test_dataset.to_table,
+        columns=["vector"],
+        filter=f"{column} = 0",
+        batch_size=32,
+    )

python/python/tests/test_dataset.py

@@ -16,6 +16,7 @@
 
 import lance
 import lance.fragment
+import numpy as np
 import pandas as pd
 import pandas.testing as tm
 import polars as pl
@@ -1994,3 +1995,14 @@ def test_legacy_dataset(tmp_path: Path):
 
     fragment = list(dataset.get_fragments())[0]
     assert "minor_version: 3" not in format_fragment(fragment.metadata, dataset)
+
+
+def test_late_materialization_batch_size(tmp_path: Path):
+    table = pa.table({"filter": np.arange(32 * 32), "values": np.arange(32 * 32)})
+    dataset = lance.write_dataset(
+        table, tmp_path, data_storage_version="stable", max_rows_per_file=10000
+    )
+    for batch in dataset.to_batches(
+        columns=["values"], filter="filter % 2 == 0", batch_size=32
+    ):
+        assert batch.num_rows == 32

rust/lance-encoding/src/decoder.rs

@@ -920,7 +920,13 @@ impl DecodeBatchScheduler {
         mut schedule_action: impl FnMut(Result<DecoderMessage>),
     ) {
         let rows_requested = ranges.iter().map(|r| r.end - r.start).sum::<u64>();
-        trace!("Scheduling ranges {:?} ({} rows)", ranges, rows_requested);
+        trace!(
+            "Scheduling {} ranges across {}..{} ({} rows)",
+            ranges.len(),
+            ranges.first().unwrap().start,
+            ranges.last().unwrap().end,
+            rows_requested
+        );
 
         let mut context = SchedulerContext::new(io);
         let maybe_root_job = self.root_scheduler.schedule_ranges(ranges, filter);