rivet-dev · NathanFlurry · Mar 24, 2026
diff --git a/.agent/notes/bench-report.md b/.agent/notes/bench-report.md
@@ -0,0 +1,170 @@
+# SQLite Benchmark Report
+
+Generated: 2026-04-02 21:48:19 UTC
+Engine: http://127.0.0.1:6420
+Mode: quick
+
+## Results
+
+| Benchmark | Native (ms) | WASM (ms) | Speedup | Native/op | WASM/op |
+|-----------|------------:|----------:|--------:|----------:|--------:|
+| Migration (50 tables + indexes) | 21.4 | 89.6 | **4.2x** | 0.428ms/op | 1.793ms/op |
+| Migration TX (50 tables + indexes) | 10.5 | 15.5 | **1.5x** | 0.209ms/op | 0.310ms/op |
+| Insert single x1 | 2.9 | 6.3 | **2.2x** | 2.883ms/op | 6.268ms/op |
+| Insert single x10 | 21.1 | 33.4 | **1.6x** | 2.110ms/op | 3.341ms/op |
+| Insert single x100 | 129.2 | 158.0 | **1.2x** | 1.292ms/op | 1.580ms/op |
+| Insert single x1000 | 665.0 | 1024.4 | **1.5x** | 0.665ms/op | 1.024ms/op |
+| Insert single x10000 | 7525.5 | 21641.3 | **2.9x** | 0.753ms/op | 2.164ms/op |
+| Insert batch x1 | 38.1 | 4.6 | 0.1x | 38.078ms/op | 4.565ms/op |
+| Insert batch x10 | 3.6 | 5.6 | **1.6x** | 0.359ms/op | 0.564ms/op |
+| Insert batch x100 | 7.0 | 10.4 | **1.5x** | 0.070ms/op | 0.104ms/op |
+| Insert batch x1000 | 72.2 | 90.9 | **1.3x** | 0.072ms/op | 0.091ms/op |
+| Insert batch x10000 | 768.9 | 959.7 | **1.2x** | 0.077ms/op | 0.096ms/op |
+| Insert TX x1 | 5.8 | 3.6 | 0.6x | 5.769ms/op | 3.572ms/op |
+| Insert TX x10 | 1.7 | 1.2 | 0.7x | 0.171ms/op | 0.119ms/op |
+| Insert TX x100 | 10.7 | 7.9 | 0.7x | 0.107ms/op | 0.079ms/op |
+| Insert TX x1000 | 29.3 | 29.5 | ~1.0x | 0.029ms/op | 0.029ms/op |
+| Insert TX x10000 | 560.0 | 442.0 | 0.8x | 0.056ms/op | 0.044ms/op |
+| Point read x1 | 0.1 | 0.4 | **2.9x** | 0.127ms/op | 0.375ms/op |
+| Point read x10 | 0.2 | 0.4 | **1.6x** | 0.023ms/op | 0.035ms/op |
+| Point read x100 | 4.5 | 1.7 | 0.4x | 0.045ms/op | 0.017ms/op |
+| Point read x1000 | 21.4 | 14.1 | 0.7x | 0.021ms/op | 0.014ms/op |
+| Point read x10000 | 196.3 | 122.3 | 0.6x | 0.020ms/op | 0.012ms/op |
+| Full scan (500 rows) | 0.5 | 1.2 | **2.6x** | - | - |
+| Range scan indexed (undefined rows) | 0.4 | 0.4 | **1.1x** | - | - |
+| Range scan unindexed (undefined rows) | 0.7 | 0.5 | 0.8x | - | - |
+| Large payload insert (4KB x 100) | 47.3 | 25.4 | 0.5x | 0.473ms/op | 0.254ms/op |
+| Large payload read (4KB x 100) | 24.5 | 4.1 | 0.2x | - | - |
+| Large payload insert (32KB x 20) | 146.3 | 307.5 | **2.1x** | 7.315ms/op | 15.376ms/op |
+| Large payload read (32KB x 20) | 42.6 | 4.4 | 0.1x | - | - |
+| Complex: join (200 rows) | 0.3 | 0.9 | **3.1x** | - | - |
+| Complex: aggregation (10 rows) | 0.3 | 1.3 | **4.2x** | - | - |
+| Complex: cte_window (50 rows) | 0.4 | 2.2 | **5.3x** | - | - |
+| Complex: subquery (100 rows) | 0.2 | 0.6 | **3.4x** | - | - |
+| Bulk update (~250 rows) | 1.0 | 1.9 | **2.0x** | - | - |
+| Bulk delete (~250 rows) | 3.5 | 3.0 | 0.9x | - | - |
+| VACUUM after delete | 5.5 | 6.8 | **1.2x** | - | - |
+| Mixed OLTP x1 (1R/0W) | 0.1 | 0.1 | ~1.0x | 0.111ms/op | 0.115ms/op |
+| Mixed OLTP x10 (8R/2W) | 1.9 | - | - | 0.189ms/op | - |
+| Mixed OLTP x100 (75R/25W) | 48.6 | - | - | 0.486ms/op | - |
+| Mixed OLTP x1000 (698R/302W) | 309.9 | - | - | 0.310ms/op | - |
+| Mixed OLTP x10000 (7016R/2984W) | 2693.0 | - | - | 0.269ms/op | - |
+| Hot row updates x1 | 0.7 | 3.5 | **5.1x** | 0.687ms/op | 3.477ms/op |
+| Hot row updates x10 | 3.9 | 27.0 | **6.9x** | 0.392ms/op | 2.705ms/op |
+| Hot row updates x100 | 33.8 | 94.1 | **2.8x** | 0.338ms/op | 0.941ms/op |
+| Hot row updates x1000 | 621.3 | 1645.7 | **2.6x** | 0.621ms/op | 1.646ms/op |
+| Hot row updates x10000 | 6628.4 | 18517.6 | **2.8x** | 0.663ms/op | 1.852ms/op |
+| JSON insert x100 | 17.1 | 9.0 | 0.5x | 0.171ms/op | 0.090ms/op |
+| JSON extract query (58 rows) | 0.4 | 5.0 | **13.5x** | - | - |
+| JSON each aggregation (5 groups) | 0.3 | 4.5 | **13.3x** | - | - |
+| FTS5 insert x100 | 6.6 | - | - | 0.066ms/op | - |
+| FTS5 search (50 hits) | 0.5 | - | - | - | - |
+| FTS5 prefix search (50 hits) | 0.5 | - | - | - | - |
+| Growth @500 rows: insert batch | 24.2 | 58.1 | **2.4x** | 0.048ms/op | 0.116ms/op |
+| Growth @500 rows: 100 point reads | 2.1 | 6.0 | **2.8x** | 0.021ms/op | 0.060ms/op |
+| Growth @1000 rows: insert batch | 9.7 | 11.1 | **1.1x** | 0.019ms/op | 0.022ms/op |
+| Growth @1000 rows: 100 point reads | 1.7 | 1.3 | 0.8x | 0.017ms/op | 0.013ms/op |
+| Growth @1500 rows: insert batch | 10.1 | 8.8 | 0.9x | 0.020ms/op | 0.018ms/op |
+| Growth @1500 rows: 100 point reads | 1.5 | 1.8 | **1.2x** | 0.015ms/op | 0.018ms/op |
+| Growth @2000 rows: insert batch | 8.7 | 9.3 | ~1.0x | 0.017ms/op | 0.019ms/op |
+| Growth @2000 rows: 100 point reads | 1.5 | 1.8 | **1.2x** | 0.015ms/op | 0.018ms/op |
+| Concurrent 5 actors total wall time | 341.0 | - | - | 500 total rows | - |
+| Concurrent 5 actors avg per-actor | 12.5 | - | - | 0.125ms/op | - |
+| [baseline] Migration (50 tables) | 7.3 | 2.2 | 0.3x | 0.145ms/op | 0.043ms/op |
+| [baseline] Migration TX (50 tables) | 3.3 | 2.0 | 0.6x | 0.067ms/op | 0.041ms/op |
+| [baseline] Insert single-row x100 | 1.9 | 2.0 | ~1.0x | 0.019ms/op | 0.020ms/op |
+| [baseline] Insert TX x100 | 0.2 | 0.3 | **1.3x** | 0.002ms/op | 0.003ms/op |
+| [baseline] Point read x100 | 0.2 | 0.3 | **1.3x** | 0.002ms/op | 0.003ms/op |
+| [baseline] Full scan (200 rows) | 0.1 | 0.1 | **1.3x** | - | - |
+| [baseline] Hot row updates x100 | 7.3 | 6.3 | 0.9x | 0.073ms/op | 0.063ms/op |
+| [baseline] Mixed OLTP x100 (66R/34W) | 0.6 | - | - | 0.006ms/op | - |
+| Mixed OLTP x10 (6R/4W) | - | 3.7 | - | - | 0.366ms/op |
+| Mixed OLTP x100 (73R/27W) | - | 59.3 | - | - | 0.593ms/op |
+| Mixed OLTP x1000 (693R/307W) | - | 798.3 | - | - | 0.798ms/op |
+| Mixed OLTP x10000 (7053R/2947W) | - | 6302.8 | - | - | 0.630ms/op |
+| Concurrent x1 wall | - | 118.7 | - | - | 100 total rows |
+| Concurrent x1 avg/actor | - | 7.9 | - | - | 0.079ms/op |
+| Concurrent x1 throughput | - | 842.7 | - | - | rows/sec |
+| Concurrent x5 wall | - | 575.3 | - | - | 500 total rows |
+| Concurrent x5 avg/actor | - | 80.8 | - | - | 0.808ms/op |
+| Concurrent x5 throughput | - | 869.1 | - | - | rows/sec |
+| Concurrent x10 wall | - | 1231.3 | - | - | 1000 total rows |
+| Concurrent x10 avg/actor | - | 350.4 | - | - | 3.504ms/op |
+| Concurrent x10 throughput | - | 812.2 | - | - | rows/sec |
+| Concurrent x50 wall | - | 4907.3 | - | - | 5000 total rows |
+| Concurrent x50 avg/actor | - | 1432.0 | - | - | 14.320ms/op |
+| Concurrent x50 throughput | - | 1018.9 | - | - | rows/sec |
+| Concurrent x100 wall | - | 10907.3 | - | - | 10000 total rows |
+| Concurrent x100 avg/actor | - | 2006.7 | - | - | 20.067ms/op |
+| Concurrent x100 throughput | - | 916.8 | - | - | rows/sec |
+| [baseline] Mixed OLTP x100 (64R/36W) | - | 0.6 | - | - | 0.006ms/op |
+
+### Insert single (scale sweep)
+
+| N | Native (ms) | Native/op | WASM (ms) | WASM/op | Speedup |
+|--:|------------:|----------:|----------:|--------:|--------:|
+| 1 | 2.9 | 2.883ms/op | 6.3 | 6.268ms/op | **2.2x** |
+| 10 | 21.1 | 2.110ms/op | 33.4 | 3.341ms/op | **1.6x** |
+| 100 | 129.2 | 1.292ms/op | 158.0 | 1.580ms/op | **1.2x** |
+| 1000 | 665.0 | 0.665ms/op | 1024.4 | 1.024ms/op | **1.5x** |
+| 10000 | 7525.5 | 0.753ms/op | 21641.3 | 2.164ms/op | **2.9x** |
+
+### Insert batch (scale sweep)
+
+| N | Native (ms) | Native/op | WASM (ms) | WASM/op | Speedup |
+|--:|------------:|----------:|----------:|--------:|--------:|
+| 1 | 38.1 | 38.078ms/op | 4.6 | 4.565ms/op | 0.1x |
+| 10 | 3.6 | 0.359ms/op | 5.6 | 0.564ms/op | **1.6x** |
+| 100 | 7.0 | 0.070ms/op | 10.4 | 0.104ms/op | **1.5x** |
+| 1000 | 72.2 | 0.072ms/op | 90.9 | 0.091ms/op | **1.3x** |
+| 10000 | 768.9 | 0.077ms/op | 959.7 | 0.096ms/op | **1.2x** |
+
+### Insert TX (scale sweep)
+
+| N | Native (ms) | Native/op | WASM (ms) | WASM/op | Speedup |
+|--:|------------:|----------:|----------:|--------:|--------:|
+| 1 | 5.8 | 5.769ms/op | 3.6 | 3.572ms/op | 0.6x |
+| 10 | 1.7 | 0.171ms/op | 1.2 | 0.119ms/op | 0.7x |
+| 100 | 10.7 | 0.107ms/op | 7.9 | 0.079ms/op | 0.7x |
+| 1000 | 29.3 | 0.029ms/op | 29.5 | 0.029ms/op | ~1.0x |
+| 10000 | 560.0 | 0.056ms/op | 442.0 | 0.044ms/op | 0.8x |
+
+### Point read (scale sweep)
+
+| N | Native (ms) | Native/op | WASM (ms) | WASM/op | Speedup |
+|--:|------------:|----------:|----------:|--------:|--------:|
+| 1 | 0.1 | 0.127ms/op | 0.4 | 0.375ms/op | **2.9x** |
+| 10 | 0.2 | 0.023ms/op | 0.4 | 0.035ms/op | **1.6x** |
+| 100 | 4.5 | 0.045ms/op | 1.7 | 0.017ms/op | 0.4x |
+| 1000 | 21.4 | 0.021ms/op | 14.1 | 0.014ms/op | 0.7x |
+| 10000 | 196.3 | 0.020ms/op | 122.3 | 0.012ms/op | 0.6x |
+
+### Mixed OLTP (scale sweep)
+
+| N | Native (ms) | Native/op | WASM (ms) | WASM/op | Speedup |
+|--:|------------:|----------:|----------:|--------:|--------:|
+| 1 | 0.1 | 0.111ms/op | 0.1 | 0.115ms/op | ~1.0x |
+| 10 | 1.9 | 0.189ms/op | - | - | - |
+| 100 | 48.6 | 0.486ms/op | - | - | - |
+| 1000 | 309.9 | 0.310ms/op | - | - | - |
+| 10000 | 2693.0 | 0.269ms/op | - | - | - |
+| 10 | - | - | 3.7 | 0.366ms/op | - |
+| 100 | - | - | 59.3 | 0.593ms/op | - |
+| 1000 | - | - | 798.3 | 0.798ms/op | - |
+| 10000 | - | - | 6302.8 | 0.630ms/op | - |
+
+### Hot row updates (scale sweep)
+
+| N | Native (ms) | Native/op | WASM (ms) | WASM/op | Speedup |
+|--:|------------:|----------:|----------:|--------:|--------:|
+| 1 | 0.7 | 0.687ms/op | 3.5 | 3.477ms/op | **5.1x** |
+| 10 | 3.9 | 0.392ms/op | 27.0 | 2.705ms/op | **6.9x** |
+| 100 | 33.8 | 0.338ms/op | 94.1 | 0.941ms/op | **2.8x** |
+| 1000 | 621.3 | 0.621ms/op | 1645.7 | 1.646ms/op | **2.6x** |
+| 10000 | 6628.4 | 0.663ms/op | 18517.6 | 1.852ms/op | **2.8x** |
+
+## Totals
+
+- **Native total**: 21.2s
+- **WASM total**: 78.7s
+- **Overall speedup**: 3.7x
diff --git a/.agent/notes/native-sqlite-kv-channel-bench-results.md b/.agent/notes/native-sqlite-kv-channel-bench-results.md
@@ -0,0 +1,154 @@
+# Native SQLite KV Channel Benchmark Results
+
+## 2026-04-01: Local Manager (file-system driver)
+
+Both sides use `driver-file-system.test.ts` with `Actor Database` filter.
+KV ops go through `FileSystemManagerDriver` → `FileSystemGlobalState` → in-memory SQLite via `better-sqlite3`.
+No Rust engine involved.
+
+**Before**: main branch (`2e29ab3f8`), WASM VFS only
+**After**: `native-sqlite-kv-channel` branch (`4c5ec7a78` + fixes), native SQLite via KV channel WebSocket
+
+### Results (http/bare encoding)
+
+| Test | WASM (ms) | Native (ms) | Speedup |
+|------|-----------|-------------|---------|
+| bootstraps schema on startup (raw) | 881 | 225 | **3.9x** |
+| bootstraps schema on startup (drizzle) | 858 | 324 | **2.6x** |
+| CRUD + multi-statement exec (drizzle) | 771 | 858 | 0.9x |
+| handles transactions (drizzle) | 2383 | 653 | **3.6x** |
+| persists across sleep/wake (drizzle) | 3071 | 1613 | **1.9x** |
+| onDisconnect DB writes (drizzle) | 2819 | 1254 | **2.2x** |
+| high-volume inserts (drizzle) | 1862 | 343 | **5.4x** |
+| shrink+regrow with vacuum (drizzle) | 868 | 883 | ~1.0x |
+| repeated updates same row (drizzle) | 4115 | 4492 | ~0.9x |
+| integrity checks mixed workload (drizzle) | 3778 | 2211 | **1.7x** |
+
+### Summary
+
+- 2-5x speedups on most DB operations
+- Biggest gains: high-volume inserts (5.4x), transactions (3.6x), startup (3.9x)
+- Roughly equivalent: vacuum, repeated single-row updates (dominated by I/O, not VFS overhead)
+- Slight regression on simple CRUD (KV channel round-trip overhead for small ops)
+
+### Notes
+
+- Native raw DB tests fail due to singleton KV channel lifecycle issues in the test framework (each test creates a new server on different port, singleton channel can only point at one)
+- Drizzle tests pass because they use the same native SQLite path
+- The "native" speedup is native SQLite (napi-rs) vs WASM SQLite (`@rivetkit/sqlite`), both storing pages in the same local KV store
+
+---
+
+## 2026-04-01: Real Rust Engine (native SQLite via KV channel)
+
+Runner connected to local Rust engine on 127.0.0.1:6420 (RocksDB backend).
+Native SQLite addon connects to engine's `/kv/connect` endpoint via WebSocket.
+KV ops go through guard -> pegboard-kv-channel -> universaldb (RocksDB).
+
+**Branch**: `native-sqlite-kv-channel` (`4c5ec7a78` + fixes)
+**Config**: `--quick` mode (100-500 rows), `runnerKind: "normal"`, `runEngine: false`
+
+### Results (kitchen-sink/scripts/bench-sqlite.ts)
+
+| Benchmark | Time (ms) | Per-Op |
+|-----------|-----------|--------|
+| Migration (50 tables + indexes) | 2192.2 | 43.8ms/op |
+| Insert single-row x100 | 4614.2 | 46.1ms/op |
+| Insert batch x100 (batch=50) | 92.3 | 0.9ms/op |
+| Insert transaction x100 | 49.9 | 0.5ms/op |
+| Point read x100 | 302.8 | 3.0ms/op |
+| Full scan (500 rows) | 3.7 | - |
+| Range scan indexed | 3.0 | - |
+| Range scan unindexed | 3.2 | - |
+| Large payload insert (4KB x 100) | 191.9 | 1.9ms/op |
+| Large payload read (4KB x 100) | 120.1 | - |
+| Large payload insert (32KB x 20) | 246.5 | 12.3ms/op |
+| Large payload read (32KB x 20) | 195.6 | - |
+| Complex: join (200 rows) | 3.4 | - |
+| Complex: aggregation (10 rows) | 3.8 | - |
+| Complex: cte_window (50 rows) | 4.8 | - |
+| Complex: subquery (100 rows) | 3.8 | - |
+| Bulk update (~250 rows) | 66.3 | - |
+| Bulk delete (~250 rows) | 93.7 | - |
+| VACUUM after delete | 80.4 | - |
+| Mixed OLTP x100 (76R/24W) | 1638.1 | 16.4ms/op |
+| Hot row updates x100 | 3594.2 | 35.9ms/op |
+| JSON insert x100 | 38.4 | 0.4ms/op |
+| JSON extract query (58 rows) | 3.1 | - |
+| JSON each aggregation (5 groups) | 3.4 | - |
+| FTS5 insert x100 | 69.4 | 0.7ms/op |
+| FTS5 search (50 hits) | 3.1 | - |
+| FTS5 prefix search (50 hits) | 3.7 | - |
+| Growth @2000 rows: insert batch | 141.1 | 0.3ms/op |
+| Growth @2000 rows: 100 point reads | 317.7 | 3.2ms/op |
+| Concurrent 5 actors total wall | 880.7 | 500 total rows |
+| Concurrent 5 actors avg per-actor | 114.3 | 1.1ms/op |
+
+**Total benchmark time: 16,345ms**
+
+### Observations
+
+- Single-row inserts are expensive (46ms/op) due to per-op KV channel round-trip through WebSocket to engine
+- Batch inserts are ~50x faster (0.9ms/op) because they amortize round-trip cost
+- Transactions are very fast (0.5ms/op) because SQLite journals locally and only flushes KV at commit
+- Read-heavy queries (scans, joins, aggregations) are <5ms since pages are cached locally in the VFS
+- Point reads cost ~3ms/op due to KV channel round-trip for uncached pages
+- FTS5 works (native SQLite has the extension compiled in)
+- 5-actor concurrency shows good parallelism (880ms wall time for 500 total rows vs 4614ms for 100 serial inserts)
+
+### WASM vs Native comparison (both against real Rust engine)
+
+| Benchmark | WASM (ms) | Native (ms) | Speedup |
+|-----------|-----------|-------------|---------|
+| Migration (50 tables) | 8,438 | 2,192 | **3.8x** |
+| Insert single-row x100 | 9,737 | 4,614 | **2.1x** |
+| Insert batch x100 | 210 | 92 | **2.3x** |
+| Insert transaction x100 | 107 | 50 | **2.1x** |
+| Point read x100 | 703 | 303 | **2.3x** |
+| Full scan (500 rows) | 11 | 4 | **2.8x** |
+| Range scan indexed | 7 | 3 | **2.3x** |
+| Range scan unindexed | 9 | 3 | **3.0x** |
+| Large payload insert (4KB x100) | 383 | 192 | **2.0x** |
+| Large payload read (4KB x100) | 9 | 120 | 0.1x (regression) |
+| Large payload insert (32KB x20) | 471 | 247 | **1.9x** |
+| Large payload read (32KB x20) | 9 | 196 | 0.05x (regression) |
+| Complex: join (200 rows) | 7 | 3 | **2.3x** |
+| Complex: aggregation | 7 | 4 | 1.8x |
+| Complex: cte_window | 9 | 5 | 1.8x |
+| Bulk update (~250 rows) | 124 | 66 | **1.9x** |
+| Bulk delete (~250 rows) | 175 | 94 | **1.9x** |
+| VACUUM | 157 | 80 | **2.0x** |
+| Mixed OLTP x100 | 4,288 | 1,638 | **2.6x** |
+| Hot row updates x100 | 7,093 | 3,594 | **2.0x** |
+| JSON insert x100 | 86 | 38 | **2.3x** |
+| FTS5 insert x100 | N/A | 69 | native only |
+| Growth @2000: insert batch | 251 | 141 | **1.8x** |
+| Growth @2000: point reads | 679 | 318 | **2.1x** |
+| Concurrent 5 actors | 1,101 | 881 | **1.2x** |
+| **Total** | **36,979** | **16,345** | **2.3x** |
+
+Large payload reads regress because native SQLite does a full KV round-trip per page read while WASM caches pages in-process. This needs investigation.
+
+### Three-way comparison: Baseline vs Native vs WASM (engine v2.2.0, debug build, RocksDB)
+
+| Benchmark | Baseline (raw) | Native KV | WASM VFS | Native vs WASM | vs Baseline |
+|-----------|---------------:|-----------:|---------:|---------------:|------------:|
+| Migration (50 tables) | 2.1ms | 2,256ms | 8,438ms | **3.7x** | 1,074x |
+| Migration TX (50 tables) | 2.0ms | **136ms** | N/A | N/A | 68x |
+| Insert single-row x100 | 1.5ms | 4,719ms | 9,737ms | **2.1x** | 3,146x |
+| Insert batch x100 | N/A | 95ms | 210ms | **2.2x** | N/A |
+| Insert TX x100 | 0.2ms | 51ms | 107ms | **2.1x** | 256x |
+| Point read x100 | 0.2ms | 310ms | 703ms | **2.3x** | 1,550x |
+| Full scan (500 rows) | 0.1ms | 3.7ms | 11ms | **2.8x** | 37x |
+| Hot row updates x100 | 5.8ms | 3,633ms | 7,093ms | **2.0x** | 626x |
+| Mixed OLTP x100 | 0.5ms | 2,446ms | 4,288ms | **1.8x** | 4,892x |
+
+Key finding: wrapping migration in BEGIN/COMMIT gives **16.6x speedup** (136ms vs 2,256ms).
+The baseline shows 68-5000x overhead from KV channel round-trips per write.
+Batch/TX amortize this cost effectively.
+
+### Fixes required to run against engine
+
+1. Native addon must request `Sec-WebSocket-Protocol: rivet` subprotocol (guard adds it unconditionally to responses)
+2. Native addon needs `connected_notify` to wait for initial WebSocket connection before sending KV ops
+3. URL fix: TS side should not append `/kv/connect` since Rust `build_ws_url` already does
diff --git a/.cargo/config.toml b/.cargo/config.toml
@@ -1,3 +1,5 @@
 [build]
 rustflags = ["--cfg", "tokio_unstable"]
 
+[env]
+LIBSQLITE3_FLAGS = "SQLITE_ENABLE_BATCH_ATOMIC_WRITE"