prrao87 · Oct 31, 2026 · Oct 30, 2026 · Oct 31, 2026 · Oct 31, 2026
diff --git a/README.md b/README.md
@@ -108,7 +108,7 @@ each of the 30 queries run in the benchmark.
 
 **These are not LDBC Benchmark Results.** These independent, unaudited measurements apply to the configurations below.
 
-Neo4j, Kuzu, Ladybug and lance-graph were measured back to back on **2026-10-24** with `./run_benchmarks.sh`. lance-csr 0.3.1 was re-run on its own with `./run_benchmarks.sh lance_csr` on **2026-10-27** ([`results/environment.yml`](results/environment.yml)). DuckPGQ and FalkorDB results are from the earlier run on **2026-10-16**. All runs used the same machine:
+Neo4j, Kuzu, Ladybug and lance-graph were measured back to back on **2026-10-24** with `./run_benchmarks.sh`. lance-csr 0.5.1 was run on its own with `./run_benchmarks.sh lance_csr` on **2026-10-31** ([`results/environment.yml`](results/environment.yml)). DuckPGQ and FalkorDB results are from the earlier run on **2026-10-16**. All runs used the same machine:
 
 | | |
 | --- | --- |
@@ -125,7 +125,7 @@ Engine versions:
 - **Kuzu 0.11.3** (unchanged)
 - **Ladybug 0.21.2** (previously 0.21.1). All 30 queries use plain `conn.execute(query)`; `ANALYZE`, result cleanup and the six ART secondary indexes remain enabled
 - **lance-graph 0.5.4** (unchanged)
-- **lance-csr 0.3.1** (previously 0.2.0), a fork of lance-graph with a CSR adjacency index, running lance-graph's 30 queries over the same Lance datasets in two modes: tables in memory, and Lance datasets read lazily from disk. See the [lance-csr README](lance_csr/README.md)
+- **lance-csr 0.5.1** (previously 0.3.1), a fork of lance-graph with a CSR adjacency index, running lance-graph's 30 queries over the same Lance datasets in two modes: tables in memory, and Lance datasets read lazily from disk. See the [lance-csr README](lance_csr/README.md)
 - **DuckPGQ** on DuckDB 1.5.6 (extension build `61f5d082`), new in this run, with the six ART indexes that mirror Ladybug's
 - **FalkorDB 6.0.2**, new in this run, running Neo4j's 30 Cypher queries unmodified, with ID indexes plus seven secondary range indexes similar to Ladybug's ART indexes
 
@@ -158,38 +158,38 @@ Every Neo4j, Kuzu, Ladybug, DuckPGQ and FalkorDB query has a 10-second timeout;
 
 Times are in milliseconds. The ratio in parentheses is Neo4j mean / engine mean, so a value above 1 means the engine was faster than this Neo4j setup.
 
-| Query | neo4j-2026.09.0 (ms) | kuzu-0.11.3 (ms) | ladybug-0.21.2 (ms) | lance-graph-0.5.4 (ms) | duckpgq-1.5.6 (ms) | falkordb-6.0.2 (ms) | lance-csr-0.3.1-lazy (ms) | lance-csr-0.3.1-memory (ms) |
+| Query | neo4j-2026.09.0 (ms) | kuzu-0.11.3 (ms) | ladybug-0.21.2 (ms) | lance-graph-0.5.4 (ms) | duckpgq-1.5.6 (ms) | falkordb-6.0.2 (ms) | lance-csr-0.5.1-lazy (ms) | lance-csr-0.5.1-memory (ms) |
 | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
-| q1 | 1.277 | 1.734 (0.74x) | 1.218 (1.05x) | 1.297 (0.98x) | 0.978 (1.31x) | 0.440 (2.90x) | 1.128 (1.13x) | 0.656 (1.95x) |
-| q2 | 3.260 | 1.322 (2.47x) | 1.123 (2.90x) | 2.337 (1.39x) | 3.450 (0.94x) | 4.682 (0.70x) | 1.331 (2.45x) | 0.512 (6.37x) |
-| q3 | 0.665 | 1.034 (0.64x) | 0.789 (0.84x) | 1.862 (0.36x) | 1.120 (0.59x) | 4.420 (0.15x) | 0.773 (0.86x) | 0.616 (1.08x) |
-| q4 | 2.656 | 0.848 (3.13x) | 0.393 (6.75x) | 2.982 (0.89x) | 1.545 (1.72x) | 0.215 (12.35x) | 0.751 (3.54x) | 0.481 (5.52x) |
-| q5 | 2.785 | 3.440 (0.81x) | 4.004 (0.70x) | 2.044 (1.36x) | 2.014 (1.38x) | 1.211 (2.30x) | 2.229 (1.25x) | 0.771 (3.61x) |
-| q6 | 2.597 | 0.664 (3.91x) | 0.787 (3.30x) | 0.727 (3.57x) | 0.554 (4.69x) | 0.193 (13.44x) | 0.710 (3.66x) | 0.413 (6.29x) |
-| q7 | 0.430 | 27.735 (0.02x) | 25.112 (0.02x) | 15.677 (0.03x) | 2.666 (0.16x) | 0.649 (0.66x) | 1.053 (0.41x) | 0.982 (0.44x) |
-| q8 | 7.287 | 2.723 (2.68x) | 3.375 (2.16x) | 1.141 (6.39x) | 1.825 (3.99x) | 1.801 (4.05x) | 1.753 (4.16x) | 0.613 (11.90x) |
-| q9 | 0.758 | 1.728 (0.44x) | 1.342 (0.56x) | 1.888 (0.40x) | 1.158 (0.65x) | 0.235 (3.23x) | 1.319 (0.57x) | 0.809 (0.94x) |
-| q10 | 3.788 | 1.553 (2.44x) | 1.778 (2.13x) | 26.693 (0.14x) | 2.728 (1.39x) | 0.201 (18.84x) | 0.908 (4.17x) | 0.700 (5.41x) |
-| q11 | 10.282 | 7.960 (1.29x) | 4.635 (2.22x) | 3.084 (3.33x) | 2.690 (3.82x) | 6.742 (1.53x) | 1.893 (5.43x) | 1.466 (7.01x) |
-| q12 | 2.549 | 17.156 (0.15x) | 7.840 (0.33x) | 18.684 (0.14x) | 4.340 (0.59x) | 1.125 (2.27x) | 6.801 (0.37x) | 0.693 (3.68x) |
-| q13 | 6.654 | 42.661 (0.16x) | 5.837 (1.14x) | 9.616 (0.69x) | 3.382 (1.97x) | 2.522 (2.64x) | 0.995 (6.69x) | 0.736 (9.04x) |
-| q14 | 0.460 | 1.556 (0.30x) | 0.902 (0.51x) | 2.909 (0.16x) | 1.439 (0.32x) | 0.204 (2.26x) | 0.654 (0.70x) | 0.477 (0.96x) |
-| q15 | 2.369 | 2.379 (1.00x) | 1.164 (2.04x) | 3.023 (0.78x) | 1.295 (1.83x) | 0.226 (10.49x) | 0.703 (3.37x) | 0.505 (4.69x) |
-| q16 | 1.260 | 1.775 (0.71x) | 1.036 (1.22x) | 4.329 (0.29x) | 3.400 (0.37x) | 4.415 (0.29x) | 1.056 (1.19x) | 0.624 (2.02x) |
-| q17 | 3.086 | 2.616 (1.18x) | 1.738 (1.78x) | 2.301 (1.34x) | 1.365 (2.26x) | 1.199 (2.57x) | 1.369 (2.25x) | 0.869 (3.55x) |
-| q18 | 3.108 | 1.508 (2.06x) | 1.221 (2.55x) | 1.763 (1.76x) | 1.047 (2.97x) | 0.628 (4.95x) | 1.074 (2.89x) | 0.631 (4.93x) |
-| q19 | 4.856 | 11.657 (0.42x) | 3.433 (1.41x) | 19.927 (0.24x) | 4.327 (1.12x) | 2.129 (2.28x) | 1.332 (3.65x) | 0.900 (5.40x) |
-| q20 | 421.921 | 13.247 (31.85x) | 14.154 (29.81x) | 3.132 (134.71x) | 0.956 (441.15x) | 0.455 (926.79x) | 16.448 (25.65x) | 2.047 (206.08x) |
-| q21 | 0.511 | 0.441 (1.16x) | 0.500 (1.02x) | 1.459 (0.35x) | 0.662 (0.77x) | 0.346 (1.48x) | 0.788 (0.65x) | 0.700 (0.73x) |
-| q22 | 2.657 | 23.338 (0.11x) | 7.456 (0.36x) | 13.413 (0.20x) | 5.143 (0.52x) | 0.670 (3.97x) | 0.884 (3.01x) | 0.647 (4.11x) |
-| q23 | 3.133 | 1.187 (2.64x) | 0.791 (3.96x) | 2.419 (1.30x) | 1.131 (2.77x) | 0.299 (10.49x) | 1.175 (2.67x) | 0.662 (4.73x) |
-| q24 | 0.717 | 1.233 (0.58x) | 0.812 (0.88x) | 1.754 (0.41x) | 0.705 (1.02x) | 0.201 (3.56x) | 1.001 (0.72x) | 0.627 (1.14x) |
-| q25 | 2.858 | 1.453 (1.97x) | 1.269 (2.25x) | 1.468 (1.95x) | 0.850 (3.36x) | 0.175 (16.30x) | 1.127 (2.54x) | 0.719 (3.98x) |
-| q26 | 0.649 | 3.372 (0.19x) | 1.147 (0.57x) | 2.932 (0.22x) | 1.489 (0.44x) | 0.299 (2.17x) | 1.019 (0.64x) | 0.665 (0.98x) |
-| q27 | 2.219 | 17.960 (0.12x) | 3.708 (0.60x) | 24.840 (0.09x) | 2.881 (0.77x) | 0.767 (2.89x) | 1.236 (1.79x) | 0.801 (2.77x) |
-| q28 | 2.369 | 1.513 (1.57x) | 1.776 (1.33x) | 2.498 (0.95x) | 1.171 (2.02x) | 4.895 (0.48x) | 2.246 (1.05x) | 1.359 (1.74x) |
-| q29 | 2.621 | 1.022 (2.56x) | 0.718 (3.65x) | 2.963 (0.88x) | 1.231 (2.13x) | 4.379 (0.60x) | 0.967 (2.71x) | 0.616 (4.25x) |
-| q30 | 1,359.978 | 146.520 (9.28x) | 97.766 (13.91x) | 34.313 (39.63x) | 26.809 (50.73x) | timeout | 47.496 (28.63x) | 43.670 (31.14x) |
+| q1 | 1.277 | 1.734 (0.74x) | 1.218 (1.05x) | 1.297 (0.98x) | 0.978 (1.31x) | 0.440 (2.90x) | 1.200 (1.06x) | 0.730 (1.75x) |
+| q2 | 3.260 | 1.322 (2.47x) | 1.123 (2.90x) | 2.337 (1.39x) | 3.450 (0.94x) | 4.682 (0.70x) | 1.397 (2.33x) | 0.532 (6.13x) |
+| q3 | 0.665 | 1.034 (0.64x) | 0.789 (0.84x) | 1.862 (0.36x) | 1.120 (0.59x) | 4.420 (0.15x) | 0.810 (0.82x) | 0.654 (1.02x) |
+| q4 | 2.656 | 0.848 (3.13x) | 0.393 (6.75x) | 2.982 (0.89x) | 1.545 (1.72x) | 0.215 (12.35x) | 0.785 (3.38x) | 0.501 (5.30x) |
+| q5 | 2.785 | 3.440 (0.81x) | 4.004 (0.70x) | 2.044 (1.36x) | 2.014 (1.38x) | 1.211 (2.30x) | 2.162 (1.29x) | 0.780 (3.57x) |
+| q6 | 2.597 | 0.664 (3.91x) | 0.787 (3.30x) | 0.727 (3.57x) | 0.554 (4.69x) | 0.193 (13.44x) | 0.730 (3.56x) | 0.427 (6.08x) |
+| q7 | 0.430 | 27.735 (0.02x) | 25.112 (0.02x) | 15.677 (0.03x) | 2.666 (0.16x) | 0.649 (0.66x) | 1.102 (0.39x) | 1.034 (0.42x) |
+| q8 | 7.287 | 2.723 (2.68x) | 3.375 (2.16x) | 1.141 (6.39x) | 1.825 (3.99x) | 1.801 (4.05x) | 1.740 (4.19x) | 0.628 (11.61x) |
+| q9 | 0.758 | 1.728 (0.44x) | 1.342 (0.56x) | 1.888 (0.40x) | 1.158 (0.65x) | 0.235 (3.23x) | 1.348 (0.56x) | 0.815 (0.93x) |
+| q10 | 3.788 | 1.553 (2.44x) | 1.778 (2.13x) | 26.693 (0.14x) | 2.728 (1.39x) | 0.201 (18.84x) | 0.962 (3.94x) | 0.737 (5.14x) |
+| q11 | 10.282 | 7.960 (1.29x) | 4.635 (2.22x) | 3.084 (3.33x) | 2.690 (3.82x) | 6.742 (1.53x) | 2.004 (5.13x) | 1.615 (6.37x) |
+| q12 | 2.549 | 17.156 (0.15x) | 7.840 (0.33x) | 18.684 (0.14x) | 4.340 (0.59x) | 1.125 (2.27x) | 6.761 (0.38x) | 0.760 (3.35x) |
+| q13 | 6.654 | 42.661 (0.16x) | 5.837 (1.14x) | 9.616 (0.69x) | 3.382 (1.97x) | 2.522 (2.64x) | 1.099 (6.05x) | 0.801 (8.31x) |
+| q14 | 0.460 | 1.556 (0.30x) | 0.902 (0.51x) | 2.909 (0.16x) | 1.439 (0.32x) | 0.204 (2.26x) | 0.735 (0.63x) | 0.534 (0.86x) |
+| q15 | 2.369 | 2.379 (1.00x) | 1.164 (2.04x) | 3.023 (0.78x) | 1.295 (1.83x) | 0.226 (10.49x) | 0.877 (2.70x) | 0.561 (4.23x) |
+| q16 | 1.260 | 1.775 (0.71x) | 1.036 (1.22x) | 4.329 (0.29x) | 3.400 (0.37x) | 4.415 (0.29x) | 1.134 (1.11x) | 0.659 (1.91x) |
+| q17 | 3.086 | 2.616 (1.18x) | 1.738 (1.78x) | 2.301 (1.34x) | 1.365 (2.26x) | 1.199 (2.57x) | 1.599 (1.93x) | 0.908 (3.40x) |
+| q18 | 3.108 | 1.508 (2.06x) | 1.221 (2.55x) | 1.763 (1.76x) | 1.047 (2.97x) | 0.628 (4.95x) | 1.161 (2.68x) | 0.700 (4.44x) |
+| q19 | 4.856 | 11.657 (0.42x) | 3.433 (1.41x) | 19.927 (0.24x) | 4.327 (1.12x) | 2.129 (2.28x) | 1.393 (3.49x) | 0.935 (5.19x) |
+| q20 | 421.921 | 13.247 (31.85x) | 14.154 (29.81x) | 3.132 (134.71x) | 0.956 (441.15x) | 0.455 (926.79x) | 16.771 (25.16x) | 2.026 (208.27x) |
+| q21 | 0.511 | 0.441 (1.16x) | 0.500 (1.02x) | 1.459 (0.35x) | 0.662 (0.77x) | 0.346 (1.48x) | 0.846 (0.60x) | 0.778 (0.66x) |
+| q22 | 2.657 | 23.338 (0.11x) | 7.456 (0.36x) | 13.413 (0.20x) | 5.143 (0.52x) | 0.670 (3.97x) | 0.950 (2.80x) | 0.721 (3.69x) |
+| q23 | 3.133 | 1.187 (2.64x) | 0.791 (3.96x) | 2.419 (1.30x) | 1.131 (2.77x) | 0.299 (10.49x) | 1.236 (2.54x) | 0.731 (4.28x) |
+| q24 | 0.717 | 1.233 (0.58x) | 0.812 (0.88x) | 1.754 (0.41x) | 0.705 (1.02x) | 0.201 (3.56x) | 1.078 (0.67x) | 0.697 (1.03x) |
+| q25 | 2.858 | 1.453 (1.97x) | 1.269 (2.25x) | 1.468 (1.95x) | 0.850 (3.36x) | 0.175 (16.30x) | 1.207 (2.37x) | 0.813 (3.51x) |
+| q26 | 0.649 | 3.372 (0.19x) | 1.147 (0.57x) | 2.932 (0.22x) | 1.489 (0.44x) | 0.299 (2.17x) | 1.163 (0.56x) | 0.744 (0.87x) |
+| q27 | 2.219 | 17.960 (0.12x) | 3.708 (0.60x) | 24.840 (0.09x) | 2.881 (0.77x) | 0.767 (2.89x) | 1.535 (1.45x) | 0.875 (2.54x) |
+| q28 | 2.369 | 1.513 (1.57x) | 1.776 (1.33x) | 2.498 (0.95x) | 1.171 (2.02x) | 4.895 (0.48x) | 2.691 (0.88x) | 1.408 (1.68x) |
+| q29 | 2.621 | 1.022 (2.56x) | 0.718 (3.65x) | 2.963 (0.88x) | 1.231 (2.13x) | 4.379 (0.60x) | 1.061 (2.47x) | 0.691 (3.79x) |
+| q30 | 1,359.978 | 146.520 (9.28x) | 97.766 (13.91x) | 34.313 (39.63x) | 26.809 (50.73x) | timeout | 42.024 (32.36x) | 40.900 (33.25x) |
 
 ![](results/benchmark_heatmap.png)
 

diff --git a/lance_csr/README.md b/lance_csr/README.md
@@ -6,7 +6,7 @@ This directory has no queries or data of its own. `query.py` loads [`../lance_gr
 
 ## Setup
 
-lance-csr is embedded: it runs inside the Python process, with no server. This directory is its own uv project, because lance-csr 0.3.1 needs pylance 13 while the root project's lance-graph 0.5.4 uses pylance 1.0.4. Build the Lance datasets first (see [lance-graph](../lance_graph/README.md)), then install:
+lance-csr is embedded: it runs inside the Python process, with no server. This directory is its own uv project, because lance-csr 0.5.1 needs pylance 13 while the root project's lance-graph 0.5.4 uses pylance 1.0.4. Build the Lance datasets first (see [lance-graph](../lance_graph/README.md)), then install:
 
 ```sh
 uv sync --frozen
@@ -35,4 +35,4 @@ From the repository root, `./run_benchmarks.sh lance_csr` runs both modes with t
 LANCE_CSR_MODE=memory uv run --frozen pytest benchmark_query.py --benchmark-min-rounds=5 --benchmark-disable-gc
 ```
 
-The runner doesn't pass lance-csr's `timeout` argument, so `QUERY_TIMEOUT_SECONDS` doesn't apply and queries run without a limit. The shared lance-graph query module uses the deprecated `with_node_label` and `with_relationship` builder methods, so each run prints deprecation warnings.
+The runner doesn't pass lance-csr's `timeout` argument, so `QUERY_TIMEOUT_SECONDS` doesn't apply and queries run without a limit. `query.py` replaces the shared module's `build_config()`, which uses lance-graph's builder methods, with one that uses lance-csr's `node_table` and `rel_table`. It reads each relationship's endpoint node tables from its dataset name (`person_likes_post` goes from `Person` to `Post`).
diff --git a/lance_csr/pyproject.toml b/lance_csr/pyproject.toml
@@ -4,7 +4,7 @@ version = "0.1.0"
 description = "lance-csr runs of the lance-graph query suite"
 requires-python = ">=3.13"
 dependencies = [
-    "lance-csr==0.3.1",
+    "lance-csr==0.5.1",
     "polars==1.37.1",
     "pyarrow==23.0.0",
     "pytest>=9.0.2",

diff --git a/lance_csr/query.py b/lance_csr/query.py
@@ -1,7 +1,8 @@
 """Run lance_graph/query.py's 30 queries on lance-csr.
 
-The queries, graph mapping and data (lance_graph/graph_lance) are shared with lance-graph, so both
-engines answer identical Cypher over identical Lance datasets.
+The queries and data (lance_graph/graph_lance) are shared with lance-graph, so both engines answer
+identical Cypher over identical Lance datasets. The graph mapping is the same, but built with
+lance-csr's node_table and rel_table, which also name each relationship's endpoint node tables.
 
 LANCE_CSR_MODE picks how the datasets are read; both use CSR traversal (use_csr=True):
 - memory (default): every table loaded into memory as Arrow, as lance-graph's benchmark does
@@ -29,6 +30,22 @@
 _spec.loader.exec_module(q)
 
 q.CypherEngine = functools.partial(lance_csr.CypherEngine, use_csr=True)
+
+
+def build_config() -> lance_csr.GraphConfig:
+    # lance-csr's rel_table names the node tables at both ends; read them from the dataset names,
+    # such as person_likes_post for Person to Post.
+    labels = {label.lower(): label for label in q.NODE_LABELS}
+    builder = lance_csr.GraphConfig.builder()
+    for label in q.NODE_LABELS:
+        builder = builder.node_table(label, "id")
+    for rel_type, dataset_name in q.REL_DATASETS.items():
+        parts = dataset_name.split("_")
+        builder = builder.rel_table(rel_type, labels[parts[0]], labels[parts[-1]], "src", "dst")
+    return builder.build()
+
+
+q.build_config = build_config
 if MODE == "lazy":
     def load_datasets(root: Path) -> dict[str, lance.LanceDataset]:
         names = {label: label for label in q.NODE_LABELS} | q.REL_DATASETS

diff --git a/lance_csr/uv.lock b/lance_csr/uv.lock
diff --git a/results/lance-csr-0.3.1-lazy.txt → results/archived/lance-csr-0.3.1-lazy.txt b/results/lance-csr-0.3.1-lazy.txt → results/archived/lance-csr-0.3.1-lazy.txt
diff --git a/results/lance-csr-0.3.1-memory.txt → results/archived/lance-csr-0.3.1-memory.txt b/results/lance-csr-0.3.1-memory.txt → results/archived/lance-csr-0.3.1-memory.txt
diff --git a/results/benchmark_heatmap.png b/results/benchmark_heatmap.png
diff --git a/results/environment.yml b/results/environment.yml
@@ -1,4 +1,4 @@
-date: 2026-10-27
+date: 2026-10-31
 machine: Apple M5, 24 GB
 os: macOS 26.6.2
 power: AC Power