grust-turso: durable group commit, parallel MVCC loads, FK-off loads, and a WAL/MVCC head-to-head

CHANGELOG.md

@@ -9,6 +9,24 @@ reconstructed from Git history, release commits, and the shipped docs.
 - Validate row-to-Arrow conversion with borrowed identity membership instead of
   constructing discarded adjacency, and copy string properties directly into
   Arrow buffers without temporary owned String clones.
+- Add automatic Cypher routing through `grust_datafusion::cypher::RoutedGraph`:
+  one graph captured as a typed index and an Arrow snapshot, one bounded-read
+  admission per query, and DataFusion execution for single-node scan plans
+  whose every physical operator charges candidate work and intermediate bytes
+  against the caller's `ReadQueryPolicy`. Relationship joins, unsupported
+  shapes, small graphs and plans over a limit stay on the reference executor;
+  explain reports the chosen route and why DataFusion was declined.
+- Add `run_prepared_read_query_indexed`, so a declined route keeps the original
+  admission and deadline instead of re-preparing the query.
+- Split large single-batch Arrow tables into contiguous zero-copy partitions.
+  Upstream round-robin repartitioning charged every queued slice the full size
+  of its shared parent buffers, which intermittently exhausted a 256 MiB pool on
+  a one-million-node scan with nothing copied.
+- Add `LadybugConfig::max_db_bytes`. Ladybug reserves its maximum database
+  size (8 TiB by default) as virtual address space per open database, so one
+  x86-64 Linux process can hold only about fifteen at once. `grust-ladybug`
+  unit tests now cap each database at 16 GiB, as Ladybug's own tests do;
+  parallel test threads previously failed intermittently with `Mmap ... failed`.
 
 ## 0.20.0 — Brine — 2026-09-14
 
@@ -21,6 +39,19 @@ reconstructed from Git history, release commits, and the shipped docs.
 - Admit snapshot relationship-ordinal payload and construction work before
   allocation, retaining reservations with emitted buffers. Combined input-policy
   capture preserves the prepared request deadline and exact serialized limits.
+- grust-turso no longer creates `<prefix>_edges_from_idx`, and bootstrap
+  drops it from existing stores. The edge primary key
+  `(from_id, label, to_id, identity_key)` leads with `from_id`, and Turso's
+  planner already answered every source lookup and traversal hop from that
+  key, so the index only cost a write per edge. Query plans and results are
+  unchanged. Other SQL dialects keep the index (`GraphSqlDialect::
+  edge_source_index`, default `true`). With the bulk_load example on a
+  4-core host, 1M edges: MVCC 12.7k -> 15.9k edges/s, WAL 43.6k -> 51.5k.
+- MVCC loads end in a `TRUNCATE` checkpoint inside `put_graph`, as WAL loads
+  already did, so a load returns durable in the database file with the MVCC
+  in-memory store drained.
+- Optional `mimalloc` feature on grust-turso: Turso's own default allocator,
+  installed as the binary's global allocator (MVCC loads about 9% faster).
 
 ## 0.19.0 — Mantis — 2026-09-14
 

Cargo.toml

@@ -39,7 +39,7 @@ categories = ["database", "data-structures"]
 [workspace.dependencies]
 async-trait = "0.1.89"
 helix-db = "=3.0.0"
-lbug = "0.20.2"
+lbug = "0.20.4"
 lancedb = "0.30.0"
 redis = "1.6.0"
 r2d2 = "0.8.10"

benchmarks/arrow-pipelines/README.md

@@ -103,3 +103,16 @@ diagnostic for physical plans and batch-versus-slice memory reporting. It
 executes the captured physical plan, retains each error/mismatch and checks
 the arithmetic count oracle. It does not report performance timings. See
 [evidence](evidence/cypher-plan-memory-c658abc) for the captured partition plans.
+
+### Warm routed-read profile
+
+`cypher_route <repeats> <nodes>...` captures one `RoutedGraph` per size and then
+times the complete bounded-read entrypoint on each forced route, alternating
+route order: admission, route planning, execution and portable output. Capture
+(index plus Arrow snapshot) is reported once per size and excluded from trials,
+so this measures reuse of warm representations, unlike `cypher_end_to_end`.
+Both routes run under the same `ReadQueryPolicy`; DataFusion additionally uses a
+1 GiB tracked pool, no spill and four target partitions. Each query's first
+result is the oracle for every later trial on either route; mismatches, errors
+and all timings are retained. Its measurements set
+`DEFAULT_MIN_DATAFUSION_NODES` for this scan workload only.

benchmarks/arrow-pipelines/evidence/cypher-route-capitola-20260914/README.md

@@ -0,0 +1,57 @@
+# Automatic routing: partition fix and warm route profile
+
+Capitola (aarch64 macOS, 10 cores, 64 GiB), stable Rust 1.97.1, run
+2026-09-14 23:19–23:22 UTC. The host was an interactive workstation during the
+run (Chrome, WindowServer, Spotlight, Docker Desktop); load averages were about
+27 at start and end, recorded in `host-start.txt` and `host-end.txt`. Exact
+binary hashes are in `hashes.txt`:
+
+- `before_cypher_end_to_end` is the retained `6544dc4` binary that produced the
+  [original failure](../cypher-conversion-6544dc4) (same SHA-256).
+- `after_cypher_end_to_end` and `cypher_route` are built from branch
+  `claude/auto-datafusion-route` (base `a134e05`) with the zero-copy partition
+  split and the router. `cypher_route` forces each route, so the threshold
+  constant later set from these results does not affect them.
+
+## Partition fix, one million nodes
+
+Ten process pairs, alternating order, each running three trials per route with
+the unchanged 256 MiB pool, spill disabled and four target partitions:
+
+| Binary | DataFusion trials | Memory exhaustion | Median DataFusion total (s) |
+| --- | ---: | ---: | ---: |
+| before | 30 | 2 | 0.798 |
+| after | 30 | 0 | 0.797 |
+
+Both before-binary failures are the retained `Memory Exhausted while SpillPool
+(DiskManager is disabled)` error; all 60 indexed trials passed. Two of thirty
+against none of thirty is not statistically decisive by itself (two-sided
+Fisher p ≈ 0.49). The structural evidence is stronger: the fixed physical plan
+contains no `RoundRobinBatch` repartition (regression
+`single_large_batches_are_split_without_round_robin`), and DataFusion 55.1
+`repartition/mod.rs` is where each queued slice was charged its shared parent
+buffers. The fix did not change measured DataFusion time.
+
+## Warm route profile
+
+`cypher_route 5 1000 3000 10000 30000 100000 300000 1000000`: one `RoutedGraph`
+per size, then the complete bounded-read entrypoint on each forced route, five
+alternating trials. All trials passed with identical results on both routes.
+Median seconds per query (q0 filtered count, q1 top-7 ids by ORDER BY, q2
+group-by with ORDER BY):
+
+| Nodes | q0 reference | q0 DataFusion | q1 reference | q1 DataFusion | q2 reference | q2 DataFusion |
+| ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| 1,000 | 0.00087 | 0.00098 | 0.00088 | 0.00055 | 0.00172 | 0.00090 |
+| 3,000 | 0.00223 | 0.00058 | 0.00227 | 0.00051 | 0.00449 | 0.00090 |
+| 10,000 | 0.00809 | 0.00072 | 0.00828 | 0.00096 | 0.01746 | 0.00106 |
+| 30,000 | 0.02698 | 0.00086 | 0.02802 | 0.00099 | 0.05764 | 0.00109 |
+| 100,000 | 0.09058 | 0.00096 | 0.09583 | 0.00109 | 0.19741 | 0.00119 |
+| 300,000 | 0.28274 | 0.00141 | 0.30139 | 0.00143 | 0.60173 | 0.00144 |
+| 1,000,000 | 0.97369 | 0.00224 | 1.07067 | 0.00304 | 1.99825 | 0.00228 |
+
+One-time capture (index plus Arrow snapshot) took 0.0012 s at 1,000 nodes,
+0.057 s at 100,000 and 1.055 s at 1,000,000, and is excluded from trials. The
+reference route was faster only for q0 at 1,000 nodes, so
+`DEFAULT_MIN_DATAFUSION_NODES` is 3,000. These are single-node scans on one
+loaded host; they do not qualify joins, other shapes, other hosts or backends.

benchmarks/arrow-pipelines/evidence/cypher-route-capitola-20260914/hashes.txt

@@ -0,0 +1,3 @@
+3f9578586e43a07c8f7c2e970ba4489ab0c8f3ece5e7542a2384d89469f54443  before_cypher_end_to_end
+d4f4f68ad822f96c8fa54792f13004d4b49d681775b59782ae4aa1aa3341fd6c  after_cypher_end_to_end
+5e262f2d5a0914099da9991efd82cd3698e5e738184ee6d5a89c5393675520aa  cypher_route

benchmarks/arrow-pipelines/evidence/cypher-route-capitola-20260914/host-end.txt

@@ -0,0 +1,2 @@
+Mon Sep 14 23:22:45 UTC 2026
+ 1:22  up 3 days,  7:24, 6 users, load averages: 26.96 25.72 24.88

benchmarks/arrow-pipelines/evidence/cypher-route-capitola-20260914/host-start.txt

@@ -0,0 +1,12 @@
+Mon Sep 14 23:19:22 UTC 2026
+ 1:19  up 3 days,  7:20, 6 users, load averages: 27.14 26.65 24.98
+10
+68719476736
+ %CPU COMMAND
+ 43.8 /Applications/Google Chrome.app/Contents/Frameworks/Google Chrome Framework.framework/Versions/152.0.7977.84/Helpe
+ 24.1 /System/Library/PrivateFrameworks/SkyLight.framework/Resources/WindowServer -daemon
+ 24.1 /System/Library/Frameworks/CoreServices.framework/Versions/A/Frameworks/Metadata.framework/Versions/A/Support/mds
+ 23.6 /Applications/Google Chrome.app/Contents/Frameworks/Google Chrome Framework.framework/Versions/152.0.7977.84/Helpe
+ 15.3 /System/Library/Frameworks/CoreServices.framework/Versions/A/Frameworks/Metadata.framework/Versions/A/Support/mds_
+ 12.0 /System/Library/Frameworks/WebKit.framework/Versions/A/XPCServices/com.apple.WebKit.WebContent.xpc/Contents/MacOS/
+ 10.2 /Applications/Docker.app/Contents/MacOS/Docker Desktop.app/Contents/Frameworks/Docker Desktop Helper (Renderer).ap

benchmarks/arrow-pipelines/evidence/cypher-route-capitola-20260914/oom-0-after.jsonl

@@ -0,0 +1,7 @@
+{"event":"configuration","source":"c658abce7ff26101f1ea75474098fac002e21de9-dirty","nodes":1000000,"repeats":3,"query":"MATCH (n:N) WHERE n.bucket = 3 RETURN count(*) AS count","deadline":null,"datafusion_working_memory_bytes":268435456,"datafusion_target_partitions":4,"datafusion_input_partitions":1,"datafusion_output_rows":1,"datafusion_output_bytes":1024,"indexed_memory_limit":null,"spill":false,"boundary":"row graph fixture excluded; cold representation per trial; preparation includes index or Arrow conversion and engine registration; query includes parsing, planning, execution and portable output; teardown excluded; no equivalent full resource envelope or automatic routing threshold claimed"}
+{"event":"trial","route":"indexed","trial":0,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.1175565,"query_seconds":1.641187583,"total_seconds":1.758744083}
+{"event":"trial","route":"datafusion","trial":0,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.787802083,"query_seconds":0.011109917,"total_seconds":0.798912}
+{"event":"trial","route":"datafusion","trial":1,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.7819865,"query_seconds":0.002232125,"total_seconds":0.7842186250000001}
+{"event":"trial","route":"indexed","trial":1,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.106516167,"query_seconds":1.526532,"total_seconds":1.633048167}
+{"event":"trial","route":"indexed","trial":2,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.117787666,"query_seconds":1.321058083,"total_seconds":1.438845749}
+{"event":"trial","route":"datafusion","trial":2,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.791805042,"query_seconds":0.00229525,"total_seconds":0.794100292}

benchmarks/arrow-pipelines/evidence/cypher-route-capitola-20260914/oom-0-before.jsonl

@@ -0,0 +1,7 @@
+{"event":"configuration","source":"6544dc410cf57addd06e2ba2057762ba713e953c","nodes":1000000,"repeats":3,"query":"MATCH (n:N) WHERE n.bucket = 3 RETURN count(*) AS count","deadline":null,"datafusion_working_memory_bytes":268435456,"datafusion_target_partitions":4,"datafusion_input_partitions":1,"datafusion_output_rows":1,"datafusion_output_bytes":1024,"indexed_memory_limit":null,"spill":false,"boundary":"row graph fixture excluded; cold representation per trial; preparation includes index or Arrow conversion and engine registration; query includes parsing, planning, execution and portable output; teardown excluded; no equivalent full resource envelope or automatic routing threshold claimed"}
+{"event":"trial","route":"indexed","trial":0,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.113686375,"query_seconds":1.9037511249999999,"total_seconds":2.0174374999999998}
+{"event":"trial","route":"datafusion","trial":0,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.8163245,"query_seconds":0.012801583,"total_seconds":0.829126083}
+{"event":"trial","route":"datafusion","trial":1,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.809607208,"query_seconds":0.002892708,"total_seconds":0.8124999159999999}
+{"event":"trial","route":"indexed","trial":1,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.116661792,"query_seconds":1.557994375,"total_seconds":1.674656167}
+{"event":"trial","route":"indexed","trial":2,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.106825584,"query_seconds":1.303186875,"total_seconds":1.410012459}
+{"event":"trial","route":"datafusion","trial":2,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.79118775,"query_seconds":0.002813458,"total_seconds":0.794001208}

benchmarks/arrow-pipelines/evidence/cypher-route-capitola-20260914/oom-1-after.jsonl

@@ -0,0 +1,7 @@
+{"event":"configuration","source":"c658abce7ff26101f1ea75474098fac002e21de9-dirty","nodes":1000000,"repeats":3,"query":"MATCH (n:N) WHERE n.bucket = 3 RETURN count(*) AS count","deadline":null,"datafusion_working_memory_bytes":268435456,"datafusion_target_partitions":4,"datafusion_input_partitions":1,"datafusion_output_rows":1,"datafusion_output_bytes":1024,"indexed_memory_limit":null,"spill":false,"boundary":"row graph fixture excluded; cold representation per trial; preparation includes index or Arrow conversion and engine registration; query includes parsing, planning, execution and portable output; teardown excluded; no equivalent full resource envelope or automatic routing threshold claimed"}
+{"event":"trial","route":"indexed","trial":0,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.119819209,"query_seconds":1.682823875,"total_seconds":1.802643084}
+{"event":"trial","route":"datafusion","trial":0,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.8161745,"query_seconds":0.003446916,"total_seconds":0.819621416}
+{"event":"trial","route":"datafusion","trial":1,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.795003042,"query_seconds":0.0020855,"total_seconds":0.797088542}
+{"event":"trial","route":"indexed","trial":1,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.108852917,"query_seconds":1.5090406669999998,"total_seconds":1.617893584}
+{"event":"trial","route":"indexed","trial":2,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.119839833,"query_seconds":1.2847146249999999,"total_seconds":1.4045544579999998}
+{"event":"trial","route":"datafusion","trial":2,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.786687583,"query_seconds":0.002204083,"total_seconds":0.7888916659999999}

benchmarks/arrow-pipelines/evidence/cypher-route-capitola-20260914/oom-1-before.jsonl

@@ -0,0 +1,7 @@
+{"event":"configuration","source":"6544dc410cf57addd06e2ba2057762ba713e953c","nodes":1000000,"repeats":3,"query":"MATCH (n:N) WHERE n.bucket = 3 RETURN count(*) AS count","deadline":null,"datafusion_working_memory_bytes":268435456,"datafusion_target_partitions":4,"datafusion_input_partitions":1,"datafusion_output_rows":1,"datafusion_output_bytes":1024,"indexed_memory_limit":null,"spill":false,"boundary":"row graph fixture excluded; cold representation per trial; preparation includes index or Arrow conversion and engine registration; query includes parsing, planning, execution and portable output; teardown excluded; no equivalent full resource envelope or automatic routing threshold claimed"}
+{"event":"trial","route":"indexed","trial":0,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.108929542,"query_seconds":1.602852959,"total_seconds":1.711782501}
+{"event":"trial","route":"datafusion","trial":0,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.790513042,"query_seconds":0.004116166,"total_seconds":0.794629208}
+{"event":"trial","route":"datafusion","trial":1,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.792127334,"query_seconds":0.0027765,"total_seconds":0.794903834}
+{"event":"trial","route":"indexed","trial":1,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.106862708,"query_seconds":1.534455042,"total_seconds":1.64131775}
+{"event":"trial","route":"indexed","trial":2,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.111730125,"query_seconds":1.302880542,"total_seconds":1.414610667}
+{"event":"trial","route":"datafusion","trial":2,"status":"pass","expected":62500,"columns":["count"],"rows":[[{"type":"int","value":62500}]],"preparation_seconds":0.791784167,"query_seconds":0.002815917,"total_seconds":0.7946000839999999}