Skip to content

RELEASE.md

Latest commit

 

History

History
87 lines (69 loc) · 3.84 KB

RELEASE.md

File metadata and controls

87 lines (69 loc) · 3.84 KB
Raw

TRACE-Rec v2.2.0 Release Evidence

Historical evidence: this document describes the verified v2.2.0 Docker and deployment artifacts. The sanitized public source release is v2.3.0; it does not claim that these Docker checks were rerun for the new tag. See PUBLIC_RELEASE.md.

Post-v2.2 Retrieval hardening candidate

The historical post-v2.2 candidate centralizes the Retrieval cohort key correction, adds a pre-training cohort preflight, preserves the two failed KuaiRec Retrieval evidence files byte-for-byte, and validates the repair using source-only and synthetic inputs. That historical post-v2.2 candidate passed 263/263 unit tests at its own commit; it is not the current-source count.

Test evidence is intentionally time-scoped: the v2.2 Docker release ran 249/249 tests, this post-release candidate ran 263/263, the later historical source-only snapshot ran 348/348, and the current source tree runs 365/365. These environments and revisions must not be collapsed into one release claim.

This candidate creates no new Test metric and does not turn the historical failed audit into a completed audit. The wheel, deployment archives and Docker image documented below remain verified v2.2.0 artifacts; they must not be renamed or presented as v2.3.0 evidence without a separate version-consistent build and deployment verification.

v2.1.0 is the immutable Git tag for the audited KuaiRand External Final. v2.2.0 is the later presentation and deployment release; it does not rewrite or rerun that Final.

Python Wheel

  • File: dist/shortrec-2.2.0-py3-none-any.whl
  • Size: 209,843 bytes
  • SHA-256: 37757b07eec218e17bd7801432b018e2793a4dae8c929e5a7a2959e8e44d06eb
  • Python: 3.10-3.12
python -m pip install dist\shortrec-2.2.0-py3-none-any.whl

An isolated --no-deps --target install imported shortrec.__version__ == "2.2.0". The wheel contains the shortrec-serve entry point and the complete HTML/CSS/JavaScript workbench. Runtime configs, arrays and checkpoints remain external assets.

Deployment Assets

  • Windows ZIP: dist/trace-rec-assets-v2.2.0.zip
    • 84,468,684 bytes
    • SHA-256 d5d4bdc9f7fd327bd2efe986774929cd88566baa91fc94cd321369d9f8a39267
  • Linux TAR.GZ: dist/trace-rec-assets-v2.2.0.tar.gz
    • 80,950,739 bytes
    • SHA-256 ce745d8973c9ed82f05e2590e171f2dd434ae306cb7302c4a7e614a3414c88c1

Both archives contain 24 runtime/presentation assets and 8 deployment control files. All 32 entries passed archive verification and safe extraction into an empty directory. Raw holdout rows, External Final checkpoints and training candidate checkpoints are excluded.

Docker E2E

  • Image: trace-rec:v2.2.0
  • Digest: sha256:3869856fb8a127b393c623c4b258b2f6f83c5fd89c7f1d4c22610ae165b53b0e
  • Image size: 406,642,285 bytes
  • Build: full committed Dockerfile build from python:3.11-slim
  • Runtime: CPU PyTorch 2.6.0, FAISS 1.14.3, UID 10001 non-root
  • Complete regression suite: 249/249 PASS
  • Release Compose from a restored ZIP directory: healthy
  • Real request: 20 recommendations, formal MMoE ranking applied
  • First request: 20 recommendations, cache miss, 1,484.41 ms service latency in this local run
  • Repeated request: 20 recommendations, cache hit, 0.21 ms service latency in this local run
  • External Final page: audit complete, hashes verified, not serving-eligible
  • Desktop and 390px mobile Chrome rendering: passed

Machine-readable evidence is in deployment_v220_verification.json. Cross-machine commands are in DEPLOY_OTHER_MACHINE.md.

Boundary

This is a Windows + WSL2, single-node deployment verification. It does not establish a production SLA, high availability, online A/B lift or SOTA. The default serving model remains the explicit KuaiRec Dev artifact; KuaiRand is displayed only as read-only external evaluation evidence.