diff --git a/src/hats_import/margin_cache/margin_cache_map_reduce.py b/src/hats_import/margin_cache/margin_cache_map_reduce.py
index f4cb5c21..fc12998c 100644
--- a/src/hats_import/margin_cache/margin_cache_map_reduce.py
+++ b/src/hats_import/margin_cache/margin_cache_map_reduce.py
@@ -25,8 +25,10 @@ def map_pixel_shards(
     """Creates margin cache shards from a source partition file."""
     try:
         schema = file_io.read_parquet_metadata(original_catalog_metadata).schema.to_arrow_schema()
+
+        # first pass to compute margin filter
         data = pq.read_table(
-            partition_file.path, filesystem=partition_file.fs, schema=schema
+            partition_file.path, filesystem=partition_file.fs, schema=schema, columns=[healpix_column]
         ).combine_chunks()
 
         # Constrain the possible margin pairs, first by only those `margin_order` pixels
@@ -49,6 +51,16 @@ def map_pixel_shards(
             {"margin_pixel": margin_pixel_list, "filter_value": np.arange(0, len(margin_pixel_list))}
         ).merge(margin_pairs, on="margin_pixel")
 
+        # If no row falls into any margin pixel, skip the full-schema read entirely.
+        if len(margin_pixel_filter) == 0:
+            MarginCachePlan.mapping_key_done(output_path, mapping_key, 0)
+            return
+
+        # Second pass: read the full table once, now that we know rows are needed.
+        data = pq.read_table(
+            partition_file.path, filesystem=partition_file.fs, schema=schema
+        ).combine_chunks()
+
         # For every possible output pixel, find the full margin_order pixel filter list,
         # perform the filter, and pass along to helper method to compute fine filter
         # and write out shard file.
diff --git a/tests/hats_import/margin_cache/test_margin_cache_map_reduce.py b/tests/hats_import/margin_cache/test_margin_cache_map_reduce.py
index b0d79494..6f9a72b8 100644
--- a/tests/hats_import/margin_cache/test_margin_cache_map_reduce.py
+++ b/tests/hats_import/margin_cache/test_margin_cache_map_reduce.py
@@ -1,8 +1,10 @@
 import os
+from unittest.mock import patch
 
 import numpy as np
 import pandas as pd
 import pyarrow as pa
+import pyarrow.parquet as pq
 import pytest
 from hats import pixel_math
 from hats.io import paths
@@ -197,3 +199,34 @@ def test_reduce_margin_shards(tmp_path):
 
     validate_result_dataframe(result_path, 720)
     assert not os.path.exists(shard_dir)
+
+
+def test_map_pixel_shards_no_match(tmp_path, test_data_dir, small_sky_source_catalog):
+    """When no row falls into any margin pixel, the full-schema read is skipped.
+
+    Regression guard for issue #685: the early-exit branch must short-circuit
+    before the second pq.read_table call, not just produce the same output.
+    """
+    intermediate_dir = tmp_path / "intermediate"
+    os.makedirs(intermediate_dir / "mapping")
+
+    with patch.object(pq, "read_table", wraps=pq.read_table) as read_table_spy:
+        margin_cache_map_reduce.map_pixel_shards(
+            paths.pixel_catalog_file(small_sky_source_catalog, HealpixPixel(1, 47)),
+            source_pixel=HealpixPixel(1, 47),
+            mapping_key="1_47",
+            original_catalog_metadata=small_sky_source_catalog / "dataset" / "_common_metadata",
+            margin_pair_file=test_data_dir / "margin_pairs" / "small_sky_source_pairs.csv",
+            output_path=intermediate_dir,
+            margin_order=1,
+            healpix_column="_healpix_29",
+            healpix_order=29,
+        )
+
+    # Exactly one read happened, and it asked for only the healpix column.
+    assert read_table_spy.call_count == 1
+    assert read_table_spy.call_args.kwargs.get("columns") == ["_healpix_29"]
+
+    # And the observable behavior: marker written, no shards.
+    assert os.path.exists(intermediate_dir / "mapping" / "1_47_done")
+    assert not any(p.name.startswith("order_") for p in intermediate_dir.iterdir())