Repository navigation
perf: reuse content-verified dataset validation in process - #230
Conversation
|
最终冻结树 793f37905ee809eba82ffe2acde3c9624a0f306a(当前head56eb261b)的本地验收完成:
这些是同进程固定云端夹具的测量,不是所有机器的延时保证,也不优化新CLI进程冷启动。#219保留,未改变完整性/原子快照边界。当前exact-head远端CI另行核验。 |
|
当前精确 head 56eb261 的完整 CI 已全部成功:https://github.com/codeasier/pt-snap-cli/actions/runs/38062097375 。Python3.10–3.13、lint/format/type和snapshot provenance全部通过。PR保持草稿,未合入;#219剩余边界如前述。 |
|
Merged by squash as bd74695 after final verification. The merged tree is exactly 793f37905ee809eba82ffe2acde3c9624a0f306a. Final results supersede the earlier pending status in the PR body: full suite 2142 passed / 1 opt-in skip, 92.78% coverage; installed default-scale case passed separately on the same frozen wheel; lint, format, typecheck (0 errors), wheel audit and clean-wheel acceptance passed. Exact-head CI 38062097375 passed all six jobs. Main push CI: https://github.com/codeasier/pt-snap-cli/actions/runs/38065366192 (in progress). This remains a partial improvement for #219; that issue stays open. Default warm checks still hash all bytes and shards, and new CLI processes remain cold. |
范围
Related to #219;本 PR 不关闭 #219。
默认同进程 dataset inspect 使用有界的 content-verified 验证结果缓存。每次命中前仍完整哈希 manifest/全部分片并核对路径、成员、sidecar 和前后 generation,只有与曾通过完整验证的 fingerprint 相同才省略重复 schema/quick_check/逐行验证。缓存不保留数据库连接或行级集合;返回副本,失败/取消/变更清除条目。
当前验证
最终 tree 793f37905ee809eba82ffe2acde3c9624a0f306a 已冻结。Ruff/Black、typecheck(0 errors)、build/wheel audit、clean-wheel acceptance 通过;最新非editable wheel中87项cache/immutable/report测试通过。独立复核较早同逻辑候选339项通过,并发现、验证修复同名manifest.json目录的取消清理键问题;最终微调仅显式丢弃5处pop返回值。
最终完整覆盖率与同树性能复测仍在运行,远端CI待运行,暂不宣称最终全套通过。性能记录会明确cold首次多一次hash成本、warm完整扫描字节量,以及同进程而非跨CLI的适用边界。