Skip to content

CAVS v1.1.0 — Practical Patch Policy Benchmark

Choose a tag to compare

View all tags
@github-actions github-actions released this 08 Jul 14:39
v1.1.0
3d367f6

The practical patch policy benchmark release.

CAVS now benchmarks against practical pairwise patch policies — adjacent-only updates, skip-heavy and warm-cache scenarios, base-version hubs and storage-budgeted hot pairs — instead of relying on the theoretical all-pairs O(N²) argument. The all-pairs graph stays only as the labeled theoretical one-hop baseline; every pairwise number is a real diff, applied and byte-verified.

Added

  • cavs bench patch-policy: policy-level comparison with real per-engine measurements (cavsplan built in; bsdiff, xdelta3, butler-offline when installed — missing tools skip with a recorded reason, never fatal).
  • Policies: adjacent, sparse dyadic ladder (--ladder-mode aligned|dense), base hub (--base-policy first|middle|latest-major|fixed:<id>|auto), hot pairs with storage budgets (--hot-pairs latest:K|traffic-top:K, --patch-storage-budget 1GiB|2x-latest-build), all-pairs one-hop baseline, and the CAVS content-addressed route.
  • Traffic models (adjacent-heavy, skip-heavy, live-service-weekly, major-release, random, custom:file.toml) with deterministic weighted expansion; client states cold-cache-with-previous-install and warm-cache.
  • cavs patch-policy graph|simulate|explain (replay another traffic model with no re-diffing) and cavs bench gen-stream.
  • Reports: summary.md/json (avg/P95/P99, steps, coverage), per-edge and per-query CSVs, storage/traffic/apply-chain reports, replayable patch_graph.json, plus a results README.md.
  • Route planner patch-chain risk penalty so multi-step chains are never chosen over a one-step route for a few KiB.
  • New docs: PATCH_POLICY_BENCHMARK, PRACTICAL_PAIRWISE_DIFFS, PATCH_GRAPH_POLICIES, TRAFFIC_MODELS, STORAGE_BUDGET_POLICIES.

Changed

  • README, BENCHMARKS and comparison docs no longer imply that all pairwise diff systems require O(N²) patches: O(N²) now refers specifically to the all-pairs one-hop baseline, and practical pairwise policies are first-class benchmark baselines.

Measured 10-version results (adjacent-heavy, skip-heavy, warm-cache) are in docs/results/v1.1.0/patch-policy/ and docs/BENCHMARKS.md.

Assets 7
Loading