motedb-python 0.15.1__tar.gz → 0.15.2__tar.gz

{motedb_python-0.15.1 → motedb_python-0.15.2}/.github/workflows/ci.yml

@@ -131,6 +131,7 @@ jobs:
         #     hardware-speed assumptions shared runners can't meet. Tracked in
         #     BENCHMARK.md instead.
         run: |
+          set +e
           cargo test --workspace --all-features -- \
             --test-threads=1 \
             --skip bench_ --skip stress_ --skip perf_ \
@@ -150,13 +151,29 @@ jobs:
             --skip test_scalability_latency \
             --skip test_scalability_memory \
             --skip test_durability_large_dataset_checkpoint_recovery \
-            --skip test_performance_guarantees
+            --skip test_performance_guarantees \
+            > /tmp/ci-test.log 2>&1
+          rc=$?
+          if [ $rc -ne 0 ]; then
+            # 🔑 Anonymous-readable failure surface: the trailing `Running
+            # tests/…` line before the failure block names the failing
+            # binary. (The separate diagnostics step proved unreliable — its
+            # `{ } || …` grouping suppressed set -e and silently skipped.)
+            grep -E "Running (tests|unittests)|FAILED|panicked at|^failures:" /tmp/ci-test.log \
+              | tail -8 | while IFS= read -r line; do
+                echo "::error title=integration-fail::$line"
+              done
+            echo "---- tail of test log ----"
+            tail -30 /tmp/ci-test.log
+          fi
+          exit $rc
       - name: cargo test (workspace, hardware-threshold tests skipped) [attempt 2]
         id: ci2
         if: steps.ci1.outcome == 'failure'
         env:
           CI: "1"
         run: |
+          set +e
           cargo test --workspace --all-features -- \
             --test-threads=1 \
             --skip bench_ --skip stress_ --skip perf_ \
@@ -176,7 +193,22 @@ jobs:
             --skip test_scalability_latency \
             --skip test_scalability_memory \
             --skip test_durability_large_dataset_checkpoint_recovery \
-            --skip test_performance_guarantees
+            --skip test_performance_guarantees \
+            > /tmp/ci-test.log 2>&1
+          rc=$?
+          if [ $rc -ne 0 ]; then
+            # 🔑 Anonymous-readable failure surface: the trailing `Running
+            # tests/…` line before the failure block names the failing
+            # binary. (The separate diagnostics step proved unreliable — its
+            # `{ } || …` grouping suppressed set -e and silently skipped.)
+            grep -E "Running (tests|unittests)|FAILED|panicked at|^failures:" /tmp/ci-test.log \
+              | tail -8 | while IFS= read -r line; do
+                echo "::error title=integration-fail::$line"
+              done
+            echo "---- tail of test log ----"
+            tail -30 /tmp/ci-test.log
+          fi
+          exit $rc
       - name: verdict
         if: always()
         run: |
@@ -191,8 +223,12 @@ jobs:
           # 🔑 K-debug: cargo 顶层一次跑时，crashing binary 会以 exit 101
           # 中断整个套件且不留 FAILED 行。逐 bin 运行并把每个 bin 的退出码
           # 通过 ::error:: annotations 上报（匿名可读）。
+          # 2026-10-21 修复: 曾有一行多余的 `} || echo …`（未配对大括号）
+          # 让本步骤自身以 exit 2 崩溃，逐 bin 上报从未生效过；同时对
+          # rc=101（真实测试失败，有 FAILED 行）也提取失败测试名，不再只
+          # 处理非 101 的编译错误。
           {
-            echo "### mac integration per-binary status"
+            echo "### integration per-binary status ($RUNNER_OS)"
             BINS=$(cargo test --workspace --all-features -- --list 2>/dev/null \
               | grep -E "^Running|^ " >/dev/null; cargo test --workspace \
               --all-features --no-run 2>&1 | grep -oE "Executable tests/[a-z0-9_]+\.rs" \
@@ -213,14 +249,13 @@ jobs:
                 --skip test_performance_guarantees --no-fail-fast \
                 > /tmp/one.log 2>&1
               rc=$?
-              echo "::error title=mac-integration::$t exit=$rc"
-              if [ $rc -ne 0 ] && [ $rc -ne 101 ]; then
-                grep -E "panicked at|FAILED" /tmp/one.log | head -5 \
+              echo "::error title=integration-$RUNNER_OS::$t exit=$rc"
+              if [ $rc -ne 0 ]; then
+                grep -E "^test .* FAILED|^failures:|panicked at" /tmp/one.log | head -8 \
                   | while read -r line; do echo "::error ::$line"; done
               fi
             done
           } || echo "per-binary diagnostics done" >> "$GITHUB_STEP_SUMMARY"
-          } || echo "diagnostics pass done" >> "$GITHUB_STEP_SUMMARY"
 
 
   # ── Cross-compile: aarch64-unknown-linux-gnu (build only) ────────────────

{motedb_python-0.15.1 → motedb_python-0.15.2}/.github/workflows/node-publish.yml

@@ -256,17 +256,19 @@ jobs:
         # root loader whose optionalDependencies point at them). Auth comes
         # from setup-node's .npmrc + NODE_AUTH_TOKEN. `napi pre-publish` is
         # deliberately NOT used: it assumes a GitHub release flow we don't run.
-        # Idempotent + burst-safe: skips packages already on the registry
+        # Idempotent + burst-safe: skips versions already on the registry
         # (EPUBLISHCONFLICT would otherwise abort the loop on re-runs) and
-        # paces publishes ~3 min apart with retries — npm's anti-spam
-        # heuristics reject rapid-fire new-package creation from new accounts.
+        # paces publishes with escalating retries — npm's anti-spam
+        # heuristics reject rapid-fire new-package creation from new accounts
+        # (v0.15.1 tag run: the one brand-new platform package burned all 3
+        # fast retries; a cooldown + re-dispatch is the recovery path).
         working-directory: bindings/node
         env:
           NODE_AUTH_TOKEN: ${{ secrets.NPM_TOKEN }}
         run: |
           set -e
           pub() {
-            local d="$1" name ver try
+            local d="$1" name ver try wait
             name=$(jq -r .name "$d/package.json")
             ver=$(jq -r .version "$d/package.json")
             # Compare the PINNED version, not mere existence: after a version
@@ -276,17 +278,27 @@ jobs:
             if npm view "$name@$ver" version >/dev/null 2>&1; then
               echo "  $name@$ver already on npm — skip"; return 0
             fi
-            for try in 1 2 3; do
-              echo "::group::publish $name (attempt $try)"
+            # Escalating backoff: anti-spam blocks clear after a cooldown,
+            # not immediately — fast retries just burn the attempts.
+            for try in 1 2 3 4 5; do
+              echo "::group::publish $name@$ver (attempt $try)"
               if (cd "$d" && npm publish --access public); then
-                echo "::endgroup::"; echo "  published $name"; return 0
+                echo "::endgroup::"; echo "  published $name@$ver"; return 0
               fi
               echo "::endgroup::"
-              [ "$try" = 3 ] && break
-              echo "  retry in 180s (npm anti-spam/rate heuristics)"; sleep 180
+              [ "$try" = 5 ] && break
+              case "$try" in
+                1) wait=120 ;; 2) wait=300 ;; 3) wait=600 ;; *) wait=900 ;;
+              esac
+              echo "  retry in ${wait}s (npm anti-spam/rate heuristics)"; sleep "$wait"
             done
-            echo "::error::gave up publishing $name after 3 attempts"
+            echo "::error::gave up publishing $name@$ver after 5 attempts"
             return 1
           }
-          for d in npm/*/; do pub "$d" || exit 1; sleep 90; done
+          # 主包最优先 — 它是 `npm install motedb` 的入口; 平台包是
+          # optionalDependencies, 单个被 npm 反垃圾拦截不应拖死主包
+          # (失败让 job 标红以便跟踪, 但其余包继续发)。
           pub . || exit 1
+          rc=0
+          for d in npm/*/; do pub "$d" || rc=1; sleep 120; done
+          exit $rc

{motedb_python-0.15.1 → motedb_python-0.15.2}/CHANGELOG.md

@@ -1,5 +1,53 @@
 # Changelog
 
+## [0.15.2] — 2026-10-22
+
+### 功能测试轮修复 (2026-10-22 外部功能测试 24/27 → 全绿)
+
+- **🔴 P0 Node 浮点参数静默截断**: `arr.get::<i64>` 对 JS double 不报错而是
+  向零截断(9.9→9, 1e-7→0, 123.456→123), f64 分支不可达 — 任何 `?` 传入的
+  非整数 number 全部损坏(FLOAT/DOUBLE 列通杀)。BigInt 更写 9 读 null。
+  修复 = f64 先探(napi 严格语义对非 number 干净失败)按 fract() 分流
+  Integer/Float(≤2^53), BigInt 走 get_i64 无损转换(超范围显式报错)。
+  smoke_ci.js 回归: 9.9/-2.75/1e-7/2.0/123n 逐值断言
+- **快照往还(ENOENT / already exists)**: resolve_open_path 的目录判定标记
+  还是旧布局(lsm/||wal/) — P1 惰性布局的解包快照(首写前无 wal/)判定不中,
+  落到"兄弟 .mote"规则打开不存在的路径。Python 的 import_snapshot 同中
+  (观感即"Python 没有快照 API")。修复 = 标记集扩为 meta.bin/columnar_ms/
+  lsm/wal; unpack 接受已存在的**空** dest 目录(mkdir-then-import 自然用法),
+  非空仍拒。Node openSnapshot/Python import_snapshot/Rust open_snapshot 三路
+  往返回归全绿; Python 加 export_snapshot 别名(与 Node/Rust 命名对齐,
+  原生名 export_to 保留)
+- **SHOW INDEXES [FROM t]**: 引擎语法新增(parser 原只认 SHOW TABLES), 数据
+  源为权威的 index_registry(TableSchema::indexes 不维护); 输出
+  Table/Index/Column/Type 四列
+
+### 训练生态互操作 (Phase 1 §5/§6, docs/design/2026-10-industry-parity-plan.md)
+
+- **🔑 Parquet 双向**: `EXPORT TABLE t TO 'x.parquet'` / `IMPORT PARQUET 'x'
+  INTO t` — 全类型往返(INT/FLOAT/BOOL/TEXT/TIMESTAMP/VECTOR→FixedSizeList<
+  f32>/BLOB→Binary), NULL 保留, 按列名映射(乱序/子集/缺列补 NULL), 有损方向
+  报错; arrow-array/schema/parquet 59.3.1(MSRV 1.85), feature `parquet`
+  默认开(default-edge 可关)。pyarrow 双向互认实测。Python: `export_parquet` /
+  `import_parquet`
+- **🔑 EXPORT LEROBOT**: EPISODE 表 → LeRobot v3 数据集目录(官方 writer 逐文件
+  定标), 官方 `LeRobotDataset` 加载互认实测 — MCAP→SQL→训练管线打通。Python:
+  `export_lerobot(table, dir, fps, tasks=)`
+- **修复(数据丢失级)**: 含 NULL 向量的表, 无 WHERE 全表扫描整列静默变 NULL,
+  且 compaction 将丢失永久写回 — read_vectors/read_spatial 的 7 处只前进线性
+  归并改为 `scatter_by_row_id` 精确散射; 回归测试 ×2
+- **PyTorch DataLoader**: `from motedb.torch import MoteDataset` — 列式一次
+  载入 + 逐行零拷贝视图, fork-safe; docs/19 + examples/torch_dataloader.py
+- **🔑 MCP server (§8)**: `python -m motedb.mcp <db>` — stdio JSON-RPC,
+  6 个 tools(list_tables/describe/query/vector_search/episodes/doctor);
+  引擎级只读门(Database(read_only=True), 写被引擎拒绝而非字符串嗅探),
+  多读者共存。docs/20-mcp.md; pyo3 构造器新增 read_only kwarg
+- **🔑 多引擎公开基准 (§7)**: `make bench-industry` — MoteDB vs FAISS/
+  LanceDB/sqlite-vec/DuckDB-VSS/usearch, 同语料同暴力 ground truth,
+  N/DIM 规模参数化, 方法学随数据发布(含不利行)。220K×384 实测: MoteDB
+  DiskANN recall@10 1.0000 / p50 0.57ms。docs/benchmark_report_2026-10-21.md
+- CI: 测试步骤失败时直接发逐行注解(替代从未生效的独立诊断步骤)
+
 ## [0.15.1] — 2026-10-20
 
 发布渠道与安装体验 patch(docs/design/2026-10-industry-parity-plan.md Phase 0)。