镜像站点 · 本页由第三方 GitHub 只读镜像提供,非 GitHub 官方站点,不接受任何登录或凭据输入。前往 github.com
Skip to content

Bench: publish cold context lifecycle phase and retention evidence #661

Closed

Description

@chrisbbreuer

Parent: #479
Related: #460, #461, #480, #614

Problem

The repository has an informal human-readable gc-profile lifecycle table, while #479 requires reproducible cold-context evidence that separates creation, first source/module work, destruction, retained process memory, and exact finalization. The current table aggregates repeated in-process timings, has no versioned schema or source/options identity, and cannot prove a bounded create/destroy plateau or support exact-parent causal claims.

Scope

  • Add a versioned, machine-readable cold lifecycle schema and structural validator.
  • Measure explicit no-evaluation, first-source-evaluation, first-module-evaluation, and full-feature context scenarios without hiding context creation, builtin installation, parsing/compilation, finalization, or destruction outside their declared boundaries.
  • Retain per-sample create, configure/evaluate, destroy, total wall/CPU, peak/current/retained RSS, allocator/GC ownership, and Context.GcFinalizerStats counts.
  • Add repeated create/destroy soak checkpoints that distinguish healthy bounded allocator reuse from monotonically retained context-owned state.
  • Record exact zig-js/Zig/OS/hardware revisions, context options, source/module checksums, sample order, power/thermal availability, and every rejected/contaminated sample.
  • Integrate schema and fixture self-tests into the existing benchmark validation gate and document the publication boundary.

No-workaround rules

No pre-created contexts, hidden warmup in cold scenarios, shared mutable realm state, delayed finalizers, forced GC presented as ordinary teardown, unbounded pools/caches, omitted samples, or peak RSS substituted for post-teardown current RSS.

Acceptance

  • Machine-readable fixtures cover no-evaluation, source, module, and full-feature cold lifecycles with exact boundary/identity metadata.
  • Every accepted sample includes separated create/work/destroy timings, current/peak/retained RSS, and exact finalizer counts or an explicit unavailable reason.
  • Soak validation rejects monotonic context-owned retention and verifies callbacks/finalizers exactly once.
  • Exact-parent collection fails closed on dirty revisions, competing work, checksum/options drift, missing counters, thermal contamination, and unbalanced sample order.
  • Structural self-tests, complete benchmark validation, docs generation, offline links, and focused lifecycle tests pass; tooling/docs-only work flips 0 Test262 cases.

Activity

  1. added
    performanceMeasured runtime, memory, scaling, or build performance
    benchmarksBenchmark design, evidence, and regression gates
    runtimeJavaScript runtime and builtin implementation
    on Aug 20, 2026
  2. chrisbbreuer commented on Aug 20, 2026

    @chrisbbreuer
    MemberAuthor

    Implemented and pushed b815b3e7.

    The frozen V1 profile and V19 integration add four non-scored context_lifecycle scenarios: no evaluation, first source, first dependency-free module, and full WebAssembly feature configuration plus first class/typed-array/JSON/RegExp work. Every sample preserves create/work/destroy phase time, fresh-process wall/CPU and Darwin counters, same-domain baseline/live/peak/post-destroy RSS, four soak checkpoints, context-options identity, checksum, and all 15 GcFinalizerStats counters. Finalizer accounting starts immediately after Context creation so mid-work collection cannot disappear from the record.

    The collector requires at least eight cold iterations, exact per-iteration checksums, one lane, exact first-parent/clean revisions, alternating order, adequate CPU occupancy, no competing evidence process, measured counters, and nominal thermal state. It rejects incomplete cell-kind reconciliation and material monotonic/final retained growth beyond the frozen 8 MiB/5% bound. No contexts, warmups, collections, finalizers, or pools are hidden outside the declared boundary.

    Validation: all four real macOS scenarios returned their frozen checksums; a 16-context no-evaluation diagnostic plateau had identical 75% and 100% post-destroy checkpoints; runner build passed; representative matrix 25/25; complete benchmark-comparison-test 31/31; release compatibility 12/12; dependency audit 16 edges/61 tools; docs 50 pages and 268 offline links. Benchmark/tooling/docs only, so flips 0 Test262 cases. The local timing values are diagnostic and are not published as a speed claim. Fresh pushed CI is pending.

  3. chrisbbreuer commented on Aug 20, 2026

    @chrisbbreuer
    MemberAuthor

    Completed in b815b3e7.

    The V19 non-scored lifecycle profile now covers all four required cold scenarios with separated create/work/destroy timing, same-domain RSS checkpoints, exact checksums/options, full finalizer reconciliation, and fail-closed exact-parent validation. The local 31-part benchmark gate, release compatibility, dependency inventory, 50-page docs build, and 268-link offline check passed; benchmark/tool/docs-only scope flips 0 Test262 cases.

    Fresh pushed gates:

    The profile is already in use for #662, which was selected from a fresh current-main cold sample rather than inferred from the new schema.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    benchmarksBenchmark design, evidence, and regression gatesperformanceMeasured runtime, memory, scaling, or build performanceruntimeJavaScript runtime and builtin implementation

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions