Ir al contenido

Public CI datasets are metadata-only repositories with versioned, pinned release payloads

Esta página aún no está disponible en tu idioma.

Status note (2026-11-15): proposed on 2026-11-12, accepted on 2026-11-15. By then the editions were published, pinned and verified, every workflow ran on hosted runners on dev and main, the full-corpus scale check had passed, and the open gates 3, 7 and 8 were met (see “Gate status update, 2026-11-15”). Gates 1 and 2 stay accepted by the owner’s provisional publication decision; this is not rights clearance.

Issue #1514 moves CI away from datasets mounted from private runner paths. ui-pr.yml mounts site_a through the SEED_SITE_A_PATH secret and skips its seeded UI suite when the path is absent (.github/workflows/ui-pr.yml:43, :135-:142). ui-nightly.yml mounts both sites from private paths (.github/workflows/ui-nightly.yml:228-:229). Two separate restrictions keep pull requests from running the seeded suite. Dependabot pull requests skip it because GitHub withholds the repository secret that supplies the private site_a mount path (.github/workflows/ui-pr.yml:42-:48). Fork pull requests do not run the workflow at all, because trusted self-hosted runners are not available to them (.github/workflows/ui-pr.yml:38-:40). These line references describe the workflows as of 2026-11-12, before the hosted-runner cutover recorded under “CI cutover (merged into dev)” below.

ADR 0058 makes site_a (Layer A) the public-safe tier and keeps the current site_b private. Its proposed 2026-11-11 amendment replaces site_b with a new public peer. It also says the release contract and the storage choice are separate decisions, and that it neither chooses GitHub Releases nor changes the Kaggle steps in RELEASING.md. This ADR is that separate storage decision. It does not edit ADR 0058.

The constraints measured in the v0.11.1 preflight (docs/proposals/v0.11.1-ci-datasets-preflight.md):

  • Three site_a files exceed GitHub’s 100 MiB ordinary-object limit, so the payload cannot be ordinary Git.
  • The Git LFS storage and monthly transfer allowance is shared across the organization. The preflight read the GitHub Free organization allowance as 10 GiB each. A full site_a payload fetched on every CI run would consume a material share of that.
  • GitHub’s release documentation (docs.github.com/en/repositories/releasing-projects-on-github/about-releases) requires each release asset to be under 2 GiB. It states no limit on total release size or bandwidth.

The owner chose (2026-11-11):

  • two public dataset repositories under Artist-Alley-Org, one per site, holding metadata and tooling only;
  • the binary payload published as versioned, checksum-verified GitHub Release assets;
  • no Git LFS and no binary Git history.

A public repository cannot prevent cloning, so a clone must simply contain no payload and no LFS pointer.

The owner also chose tiered testing. Full site_a (about 2,005 assets and 863 posts, after the owner-directed exclusion of the Sponza scene on 2026-11-12) and the replacement public site_b (about 503 assets and 200 posts) are for scheduled or manually dispatched scale validation. PR CI stays a small, coverage-complete site_a subset. ADR 0080 explains why: seeding the whole catalogue cost 14 minutes of a 20-minute job. The measured local site_a candidate stream is 2,664,854,328 bytes after the exclusion (2,712,587,909 before it), which is too large to download, verify and extract on every pull request. The PR subset selected by --profile ci with the overlap includes is 164 assets, 84 posts, 204 files and 728,101,752 unique bytes after the exclusion (165 assets, 87 posts, 275 files and 666,675,617 bytes before it).

Repositories hold metadata and tooling only.

  • Each public dataset repository tracks its top-level metadata files, its SHA256SUMS, its README and its .gitignore. These are ordinary Git blobs.
  • The payload directories are present only in a maintainer’s local working copy, for building releases. They are ignored by root-anchored .gitignore rules, so git add -A cannot stage them.
    • site_a payload directories: 3d/, audio/, documents/, fonts/, images/, videos/.
    • site_b payload directory: images/.
    • Both repositories use the same root layout (2026-11-13): the site root is the repository root. site_b’s generated metadata document is DATASET.md, listed in SHA256SUMS, beside a curated, unlisted README.md. Until then site_b kept its site root under site_b/.
  • Git LFS is not used, and no LFS pointer is committed.

The payload ships as versioned GitHub Release assets.

  • A release version is one tag on the dataset repository. Its assets are one manifest and the sequential parts of ONE deterministic tar.gz stream of the payload files only.

  • Every new edition uses the versioned asset layout (v2). Each asset name carries the site AND the edition tag, so downloads of two editions never overwrite each other in one folder (owner request, 2026-11-14):

    • the manifest <site>-<tag>-release.json (schema artist-alley.dataset-release/v2), which also records its edition and its own manifest name;
    • the parts <site>-<tag>-payload.tar.gz.partNNNN.

    release_dataset.py build --tag <tag> writes it. The verifier refuses a manifest whose name, recorded name, edition or part names disagree, and, given an expected tag, any other edition.

  • The published editions keep the legacy layout (v1): release.json and <site>-payload.tar.gz.partNNNN, schema artist-alley.dataset-release/v1. v2026.11.01 (both sites) and v2026.11.01-sm use it. They are not rebuilt or renamed; the tools still verify and fetch them, and build --legacy-v1 reproduces the layout byte for byte. Their names are the same in every v1 edition, which is why new editions do not use it.

  • Every part is strictly smaller than the configured limit: default 1,900,000,000 bytes, never 2 GiB or more.

  • The manifest (release.json in v1) records:

    • the schema and the site (v2 also the edition and the manifest’s own name);
    • the SHA256SUMS pin (SHA-256, bytes, entries);
    • the metadata files (path, SHA-256, bytes);
    • the payload count, bytes and sorted-list digest;
    • the archive parameters and toolchain;
    • the ordered part names, sizes and SHA-256 values, and the whole stream’s size and SHA-256;
    • a notice that packaging is not rights clearance or release approval.
  • A published version is never edited in place. A changed payload is a new version with a new tag.

  • Version tags are calendar editions, vYYYY.MM.NN (owner direction, 2026-11-13): the year and month of the edition, then a two-digit edition number that increases within that month. v2026.11.01 is the first November 2026 edition. A variant of the same numbered edition takes a lowercase variant suffix, as the PR CI subset v2026.11.01-sm of v2026.11.01 does; the tools accept only vYYYY.MM.NN with an optional -variant of lowercase letters and digits. For a v2 edition the tools also require a real month (01 to 12) and edition number (01 to 99), and a year from 0001, before any output or request, so v2026.13.01 and v2026.11.00 are refused. NN counts editions, it is not a day of the month, so a month may have more editions than days. A tag versions the DATASET and is independent of the application’s milestones (for example v0.11.1). Both dataset repositories may carry the same edition tag when they are cut together, but each tag pins only its own repository’s payload.

CI pins, downloads, verifies, then extracts.

  • A CI job that consumes a public payload is configured with four pins and one optional field:

    • the dataset repository;
    • an immutable version tag, never a mutable latest;
    • the metadata commit that tag must resolve to, checked against the remote tag and against the clone’s HEAD;
    • the expected SHA-256 of the manifest (release.json for v1, <site>-<tag>-release.json for v2), kept under the existing release_json_sha256 key for both layouts;
    • release_format: v1 when the field is absent, and required as v2 for a versioned edition.

    The tag name alone is not trusted. Release assets can be replaced under an existing tag, so the manifest pin is what binds the bytes, and the commit pin binds the metadata checkout.

    For example, the nightly pins Artist-Alley-Org/artist-alley-dataset-site-a, v2026.11.01 and release.json SHA-256 84996745.... The UI workflows take every pin from seed/contracts/ci-datasets.v1.json, on dev since PR #1533 and on main since PR #1535; all current pins are v1 editions. A v2 edition is fetched with fetch_release_dataset.py --release-format v2, which refuses a tag that is not an edition before any request and a manifest of another edition before any part. A pin declares its layout in release_format (v1, the default when the field is absent, or v2). seed/scripts/ci_pins.py validates it (and a v2 pin’s tag) before any clone or download, and scripts/ci/public-datasets.sh passes it to the fetcher as --release-format, so the CI helper consumes a v2 edition once a future pin declares it. The current pins omit the field and are unchanged.

  • The job (scripts/ci/public-datasets.sh with seed/scripts/fetch_release_dataset.py) does the following:

    1. check out the metadata repository at that tag and require it to resolve to the pinned commit;
    2. download the manifest (release.json for v1, <site>-<tag>-release.json for v2) and check it against the pinned SHA-256; for v2 it must also be a manifest of exactly that edition before any part is requested;
    3. download every listed part, and reject a missing, extra or reordered part;
    4. check every part’s size and SHA-256, and the whole stream’s SHA-256, BEFORE extraction;
    5. extract only listed regular files, into a fresh root;
    6. require the reconstructed root to pass the metadata repository’s SHA256SUMS before seeding.
  • A metadata-only clone alone cannot pass sha256sum -c SHA256SUMS until the matching payload is extracted.

What exists today, and what does not.

  • seed/scripts/release_dataset.py is a LOCAL builder and offline verifier.

    • build writes the parts and release.json.
    • verify checks a local release against a metadata-only root, extracts it into a new root and installs that root only after it passes SHA256SUMS.
    • It has no network code and knows no URLs. It is not a downloader. The CI fetch (fetch_release_dataset.py, used by the UI workflows on dev since PR #1533 and on main since PR #1535) loads it hash-pinned for its offline checks and verify.
  • On 2026-11-12 it built LOCAL CANDIDATE archives from the local metadata-only repositories, and verify reconstructed both sites from metadata-only copies:

    StatusPayload filesPartsPart bytesStream SHA-256release.json SHA-256
    site_acurrent, post-Sponza exclusion2,18021,899,999,999 and 764,854,329d99a864052651f29b42b2dd05d5b719da991a98264b93b62b6f9b8a674c3f71cc756c03880abd5729e90b9cd99262c81056965eb4e8c0b3305aed44fcd5220cd
    site_aOBSOLETE, pre-exclusion, contains Sponza: must not be uploaded2,25121,899,999,999 and 812,587,91025a50b8960089d4d9d066a4fc690eec88d73e601b970e3e51d83fc8102ff6a0e8c28c6dcccb692c0c7b3c8858d8d74eeddf419ed675ae3010ae088bf08922e76
    site_bcurrent50313,156,078bc2a609fa88afc3a5424bb54dfde54242f1ad9a94e30718e966ddb09cffb9a85cdc4b78a959461386a9053b92401c6381e86721dfd44db3b1f3e30c656e80184

    The current site_a candidate:

    • Source: site_a metadata commit 249c15fc0d2b7ea13acdbf26f4e5ff90a7796ed4: 2,005 assets, 863 posts, and 2,187 SHA256SUMS entries (2,180 payload files of 2,709,016,036 bytes).
    • Why it was built from a private staging root: the builder refuses unlisted payload files, and the 71 excluded Sponza files remain, ignored and unlisted, on local disk for reversibility.
    • Reproducibility: two independent builds with Python 3.12.3 and zlib 1.3 were byte-identical.

    The site_b candidate was rebuilt after its metadata README and SHA256SUMS line were corrected. Its payload part is byte-identical to the earlier build. After a 2026-11-12 host crash deleted an earlier temporary copy, the archive was rebuilt byte-identically from the unchanged site_b metadata repository.

    On 2026-11-13 the site_b repository was re-laid out at its root (the generated README became DATASET.md). Its rebuilt candidate’s payload part is byte-identical to the previous build (96e2841b..., 3,160,894 bytes); only its release.json changed (a4b1fa20...), because its metadata pins now name DATASET.md and the new SHA256SUMS.

    Both metadata repositories are public on GitHub since 2026-11-13 (Artist-Alley-Org/artist-alley-dataset-site-a and -site-b), metadata only. On 2026-11-13 each received ONE DRAFT prerelease GitHub Release, v2026.11.01, carrying its candidate release.json and parts:

    • site_a: release id 443965050, target 62ebfa6a..., release.json 84996745...;
    • site_b: release id 443964463, target 9fa8fb76..., release.json a4b1fa20....

    The drafts were first prepared as v0.11.1-rc1 and retagged before any publication. A draft is not public, creates no Git tag until it is published, and is not a pin. On 2026-11-13 the owner accepted the recorded source evidence for edition v2026.11.01 and decided to publish it; see “Owner publication decision” under the gates below.

    Published 2026-11-13. Both drafts were then published, unchanged, as public PRERELEASES. In each repository the tag v2026.11.01 now exists and resolves to the target commit (site_a 62ebfa6a..., site_b 9fa8fb76...); release ids, asset ids, names, sizes and digests are the ones listed above, and each main is unchanged. A fresh anonymous checkout at the tag and a fresh network download of every published asset were then verified with seed/scripts/fetch_release_dataset.py against the pinned release.json SHA-256s. Both reconstructed roots passed sha256sum -c (site_a 2,189 entries, 2,005 assets and 863 posts; site_b 514 entries, 503 assets and 200 posts). Both overlap anchors were byte-identical across the two roots under distinct asset ids, and they were the only two cross-peer byte overlaps. The non-overlap control was present only in site_b. Sponza, the Pexels videos and 2CylinderEngine were absent as files and as asset rows in the public site_a; they are named only in its replacement-provenance fields.

    The private local candidates remain retained and are not themselves pins. The pins are the PUBLISHED editions: repository, tag v2026.11.01, the commit the tag resolves to and the published release.json SHA-256, kept in one file, seed/contracts/ci-datasets.v1.json.

  • The PR subset is its own edition, v2026.11.01-sm (see “Small PR edition” below), published on 2026-11-13 as a public prerelease. aa seed-package was not used: it is rights-gated and refuses while the ledger is pending. The full edition remains the nightly’s dataset.

Test tiers.

  • PR CI keeps a small, coverage-complete site_a subset (ADR 0080).
  • The accepted nightly contract (ADR 0080, 2026-10-02 amendment) is unchanged: site_a seeded with --profile ci, and site_b as a paired peer for peer presence (.github/workflows/ui-nightly.yml:298, :306).
  • Full site_a and the replacement public site_b belong to a SEPARATE scheduled or manually dispatched scale job. That job is .github/workflows/ui-scale.yml (PRs #1548 and #1550); see “Full-corpus scale validation” below.
  • The old private site_b is excluded entirely: it is in no public repository, release asset, CI artifact or log.
  • With the hosted cutover (merged into dev and promoted to main, see below), the two UI workflows seed from the pinned public editions on GitHub-hosted runners. Dependabot and fork pull requests then run the seeded UI suite: it needs no secret and no trusted runner, only public data and the read-only token. This ADR still does not authorize running untrusted fork code on trusted runners, and the cutover does not do so.
  • On dev, every workflow now runs on GitHub-hosted runners (as of 2026-11-15, after PRs #1536 and #1539 to #1545), including ci.yml with the required Integration tests (full compose stack) check, edge.yml and release.yml. main received the same revisions through PR #1547 (merge a379e5cc, 2026-11-15), so no workflow job on main uses a self-hosted runner either. See “Runner state on dev, 2026-11-15” below.
  • The Git history stays small, and nothing in it is a binary payload. A clone is metadata. The bytes come only from a pinned, verified release.
  • A dataset change is a deliberate new version. CI moves to it only by changing its pins, and rollback means restoring the previous pins. No published version is overwritten. See “Rollback” below for the procedure, including the case where no older public edition exists.
  • Download and verification add CI time, and part splitting is part of every build. On the hosted runner, the passing feature-branch runs fetched and verified the small site_a edition in 19 s (ui-pr 43529102983) and both full editions in 1 min 21 s (ui-nightly 43529650255). On the owner’s development machine the CI helper took about 40 seconds for the full site_a and about 1 second for site_b. The workflows print free disk and per-phase time on every run, so the cost is recorded, not assumed.
  • Byte-identical rebuilds hold for one Python and zlib toolchain, which release.json records. A toolchain change may change the bytes but not the payload list.
  • GitHub’s limits can change. The 2 GiB per-asset rule is rechecked before each publication.
  • Ordinary Git. Rejected: files exceed 100 MiB, and binary history grows with every revision.
  • Git LFS. Rejected by the owner: the storage and transfer allowance is shared and would be drawn on every CI run, and LFS pointers would still put the binary history in the repository’s LFS store.
  • Kaggle. RELEASING.md uses it for the published demo dataset. It is not chosen for CI, because version-pinned anonymous downloads were not established. This ADR does not change the Kaggle procedure. Since 2026-11-15, RELEASING.md states that its Kaggle steps apply only to a release that republishes the demo seed, and that the CI editions are published and pinned separately.
  • Public object storage. It would work, but adds cost, credentials and operational burden that Releases do not.
  • Actions artifacts or cache. Their retention and eviction make them unsuitable as the only source. They remain an option for accelerating a release that is already verified.

These are the general gates for a public dataset edition. For edition v2026.11.01 the owner’s 2026-11-13 decision, recorded after the list, accepts gates 1 and 2 without completing them.

  1. site_a evidence reconciliation. The owner states that the historical site_a verification is complete. The new CI-release evidence still has to bind it to the frozen bytes:
    • the fresh structural draft ledger for the post-exclusion CI selection has 204 rows, all pending, none reviewed;
    • the private evidence crosswalk (v2) was made for the pre-exclusion 275-row draft. It lists 24 of those rows as supported candidates, 163 as needing corroboration, and 88 as conflict or hold candidates;
    • 194 of the 204 current files are byte-identical files from that 275-row set. The 10 that entered with the exclusion have manifest labels and attributions, but no row in that crosswalk, and they are pending and unreviewed in the new ledger.
  2. site_b rights review. Every file of the replacement public site_b needs a formal per-file review. All 503 private evidence rows are pending_review.
  3. Policy. The ADR 0058 amendment replacing private site_b must be accepted.
  4. Release verification. The release manifests and parts must be built from the reviewed bytes, and verified against a metadata-only clone.
  5. Publication targets. Each site needs a distinct public repository remote, an immutable tag and the uploaded release assets, all re-verified after upload.
  6. CI configuration. CI must pin the repository, the tag and the expected release.json SHA-256, with the download-verify-extract steps implemented and reviewed.
  7. First runs. An initial cold CI smoke, then the scale checks for the large tier.
  8. Rollback. The rollback procedure must be documented.

Owner publication decision for v2026.11.01 (2026-11-13). The owner reviewed the recorded source evidence and elected to publish dataset edition v2026.11.01 of both repositories as a provisional public prerelease. For this edition, gates 1 and 2 are accepted by that decision rather than completed:

  • the per-file ledger rows are NOT all reviewed (the site_a CI-selection draft is still pending, and all 503 site_b rows are pending_review);
  • the unresolved source caveats stay disclosed in the release notes, the dataset README.md/ATTRIBUTIONS.md and the records, and are not silently resolved: the 75 NPS clips are PROVISIONAL, the Factory Kit model’s source review is pending, and site_b’s comparison against Kenney’s current first-party packs was not run.

This is not legal rights clearance. Status of the other gates for this edition:

  • Gate 4 is met by the verified rev115 (site_a) and rev125 (site_b) candidates.
  • Gate 5 is met: both public repositories carry a published v2026.11.01 prerelease whose tag resolves to the pinned commit, and a fresh public download of every asset was verified against the pinned release.json SHA-256s (see “Published 2026-11-13” above).
  • Gate 6 is met on dev: the CI wiring below (PR #1533) was squash-merged into dev on 2026-11-14 as 7b5eaa0e, whose tree is identical to the tested head e1b5d950. PR #1535 promoted it to main on 2026-11-14 (merge 2fff3580), so the scheduled nightly now runs the hosted workflows.
  • Gate 7’s cold hosted smoke passed before the merge (see below); its large-tier scale checks are still open.
  • Gates 3 and 8 remain open.

Gate status update, 2026-11-15. The three open gates above are now met:

  • Gate 3. ADR 0058’s amendment is accepted (“Amendment accepted, 2026-11-15”).
  • Gate 7. The large-tier scale checks passed on hosted runners, both on the feature branch and from main; see “Full-corpus scale validation” below. The first SCHEDULED run had not yet been observed when this was written.
  • Gate 8. The rollback procedure is documented under “Rollback” below.

Gates 1 and 2 remain accepted by the owner’s decision, not completed.

CI cutover (merged into dev). The owner directed that the dataset CI run on public GitHub-hosted runners rather than the owner’s self-hosted machines. The change:

  • ui-pr.yml and ui-nightly.yml run on ubuntu-24.04 (4 CPU, 16 GiB RAM, 14 GiB SSD). They no longer use SEED_SITE_A_PATH, SEED_SITE_B_PATH, /datasets, /runner-tls or a docker network connect of a runner container. The seed-preflight job that skipped the suite without the secret is removed; the job and check names are kept.

  • scripts/ci/public-datasets.sh reads seed/contracts/ci-datasets.v1.json, checks that the tag resolves to the pinned commit remotely and in a fresh anonymous metadata checkout, runs seed/scripts/fetch_release_dataset.py with the pinned release.json SHA-256 into a new root, makes the tree readable by the containers’ nonroot user (mode bits only) and re-runs sha256sum -c on the tree that is mounted. Any failure stops the job.

  • infra/docker/ci/docker-compose.hosted.yml publishes app, nginx and nginx-b on loopback at their in-network ports, and the VM’s /etc/hosts maps their names to 127.0.0.1, so each URL means the same thing from the runner and from the containers. The apps still federate over Docker DNS at the same peer URLs.

  • The nightly’s studio-b.local certificate comes from a per-run throwaway CA (scripts/ci/ephemeral-studio-b-tls.sh) whose key is destroyed after signing.

  • The seed contracts are unchanged: PR site_a --profile ci --fixtures; nightly site_a --profile ci --fixtures and site_b --limit-per-extension 3 --fixtures, now against the public replacement site_b.

  • ui-pr fetches the small edition (pins key site_a_sm, tag v2026.11.01-sm, passed to the fetch tool as dataset site site_a); ui-nightly keeps the full site_a and site_b v2026.11.01 editions. Pin validation lives in seed/scripts/ci_pins.py: full editions accept only vYYYY.MM.NN, and site_a_sm only vYYYY.MM.NN-sm; release_format is v1 (the default when absent) or v2.

  • Disk is the binding constraint on the advertised standard runner (14 GB). The jobs are ordered by disk phase: build the image(s), prune the job’s own build cache, then fetch, seed and release the dataset. A phase model (seed/scripts/ci_capacity.py, inputs in the pins file’s capacity_model) takes the peak per phase rather than one sum, and two checks gate on free space MEASURED on the runner: a whole-job preflight before the build and a binding check after the build. Neither deletes anything to make room. The build inputs were measured on a clean overlay2 daemon (cold builds, du of the Docker root before and after the cache prune):

    profilebuild peakkept after prunewhole-job needneed after build
    ui-pr (small edition)5,702,983,5241,758,671,9058,176,725,3485,684,354,560
    ui-nightly8,679,315,1872,981,508,04712,913,080,7358,431,398,082

    Both fit 14,000,000,000 bytes; the nightly has about 1.1 GB of headroom. With the full edition, ui-pr needed 9,956,057,251 and 6,797,385,346. The earlier staged arithmetic (every download and extracted byte plus an 8 or 10 GiB Docker floor, all at once) needed 14,607,515,312 and 16,761,559,854 bytes and could not pass. The seed growth, postgres, Playwright and margin terms are partly allowances, and the hosted Docker storage layout may differ from the measured one. Both capacity checks passed on the hosted runs; the allowances themselves have not been re-measured.

Small PR edition v2026.11.01-sm (published 2026-11-13, public prerelease). By owner direction, PR CI uses a smaller site_a edition; the full edition stays the nightly’s. It is a byte-identical subset of v2026.11.01: the rows and files aa seed --profile ci selects (application 2421d6d8), the two declared overlap anchors and the selected rows’ licence notice.

  • 165 of 2,005 assets, 91 of 863 posts, 9 of 92 videos, 207 payload files (665,709,517 bytes); 131 of 131 coverage dimensions; both anchors; every included row byte-identical to v2026.11.01.
  • Its descriptive files describe the subset: groups.csv asset_count and metadata.csv group_size are the retained members; ATTRIBUTIONS.md summarises the included rows’ licence and attribution values and keeps the full edition’s text as labelled background; dataset-metadata.json describes the subset; EDITION.json records the source pins, the selection and every change. The rights caveats are carried unchanged.
  • Release (built with release_dataset.py): public prerelease github.com/Artist-Alley-Org/artist-alley-dataset-site-a/releases/tag/v2026.11.01-sm, release id 444795243. Its release.json has SHA-256 c6838e5020b1bc56f048c0602c2021c1c840e3e3ddf236827eb7358b521a503f, and its one part is 658,160,717 bytes, SHA-256 139292dcd564bff3b513356087acf9992544b0e4b9a3a5e8f8f412d856f31f88.
  • Metadata: commit 3f078b1500ce066d5df04adeba11d38fc8ad7277 on the public branch edition/v2026.11.01-sm of the site_a repository (parent 62ebfa6a, 11 metadata files, no payload), with the public lightweight tag v2026.11.01-sm on it. main and v2026.11.01 are unchanged.
  • Verified before publication: the tagged metadata plus the release rebuild a root that passes all 215 checksums; the CI selection on it equals the full-edition coverage selection; aa seed-overlap passes against the public site_b; an isolated aa seed --profile ci --fixtures covered 131 of 131.
  • Verified after publication: an anonymous fetch of the public edition through scripts/ci/public-datasets.sh fetch site_a_sm (tag check, metadata checkout at the pinned commit, download and verification against the pinned release.json) passed. The rebuilt root passed all 215 checksums, held 165 assets, 91 posts, 9 videos and both overlap anchors, and covered 131 of 131 CI dimensions; the downloaded assets’ digests equal the published ones. The temporary copy was deleted afterwards.

The CI cutover (PR #1533) was squash-merged into dev on 2026-11-14 as 7b5eaa0e, whose tree is identical to the tested head e1b5d950. Before the merge, at e1b5d950, both UI workflows ran on GitHub-hosted runners and passed, which is gate 7’s cold hosted smoke:

  • ui-pr run 43529102983 (pull request) fetched and verified the small edition, seeded, drained the preview queue (112 gated jobs, cleared in about 13 min 44 s under the 1,500 s cap) and passed all 556 Playwright tests (0 failed, skipped, not run or flaky).
  • The ui-nightly dispatch 43529650255 fetched and verified both full editions, passed 8 shell scenarios (3 outline scenarios skip by design), drained in about 10 min 8 s and passed all 560 Playwright tests.
  • The other PR checks passed too, including the required self-hosted integration tests.

Getting there took two corrections. The first hosted runs (43520041934 and 43520575280, on 232cbb44) fetched, verified and seeded, then hit the original 480 s drain cap with 81 of 112 jobs outstanding. With the cap raised to 1,500 s and the hosted browser cache given its own key, run 43523595164 (on 9bede5a4) cleared the drain but failed one layout test by 8.65625 px: a slot switch in the collection cover editor resized the dialog when one slot’s hint wrapped on the hosted font. e1b5d950 fixed that in the component and kept the test’s strict assertion.

Promotion and runner split on 2026-11-14. PR #1535 promoted dev to main (merge 2fff35808ac124d94c8cfa1693d8ebbf6023a3d5), so the scheduled nightly, which runs main, now uses the hosted UI workflows. Before the merge, the hosted nightly passed on the exact promoted tree (dispatch 43610213570 on dev ce34d692). The public dataset release descriptions and README files now state these CI uses.

At that date ci.yml’s six jobs, release.yml and edge.yml still ran on the self-hosted pool, and PR #1536 (merged into dev as d2650d74) had moved site-rebuild-trigger.yml’s dispatch job to a GitHub-hosted runner.

Runner state on dev, 2026-11-15. PRs #1539 to #1545 moved the rest: the four light ci.yml jobs (#1539), the production image build check (#1540), the Edge image job (#1541), Integration (#1542, 13 min 17 s on the hosted PR run 43649134640), both release.yml jobs (#1543), actions/cache v5 (#1544) and the Node 24 Docker actions (#1545). On dev at a3c9611f no workflow job uses a self-hosted runner, and the CI, Edge, Seed guard and UI push runs there passed with no Node 20 annotations. A no-push hosted rehearsal of the release image build passed (run 43654921519, build step 5 min 44 s), but the release publishing path (registry push, cosign signing, GitHub Release creation) has not yet run on a hosted runner; the first real tag is its first live test. On that date main (2fff3580) still had the earlier self-hosted revisions until a later dev to main promotion. That promotion was PR #1547 (merge a379e5cc, 2026-11-15). Since then no workflow job on main uses a self-hosted runner, and the first application tag cut from main after it is the first live test of the hosted release path.

The next dataset edition uses the versioned (v2) asset layout, and its pin declares release_format: v2. Gate 3 (the ADR 0058 policy amendment), gate 8 (rollback) and gate 7’s large-tier scale checks stay open and are tracked separately; this cutover does not resolve or waive them. All three were met on 2026-11-15; see the gate status update above.

.github/workflows/ui-scale.yml (PR #1548; daily schedule added in PR #1550, promoted to main in PRs #1549 and #1551) checks paging and endless scroll on each whole public edition. It leaves ui-pr, ui-nightly, their pins and the fixture principals untouched.

How it works.

  • One job per edition. A matrix of site_a and site_b, each on its own disposable single-stack ubuntu-24.04 VM.
  • Fetch and seed. Each job fetches and verifies its pinned edition with scripts/ci/public-datasets.sh. It then seeds ALL of it, with the default profile and --previews=false, and without --fixtures.
  • Expected sets. scripts/ci/scale_expected.py records them straight from the database, in read-only transactions, and fails the job if any of these disagree: the seed report, the edition’s MANIFEST.json, preview jobs that should not exist, or the seeded principals.
  • The scale Playwright project (scripts/dogfood/run-ui.sh scale, retries 0, one worker):
    • walks /api/v1/assets and /api/v1/posts to their end at page sizes 36 and 200, as the admin (equal to the database) and as seeded users (chen.wei on site_a; chen.wei and anya.volkov on site_b);
    • scrolls the home feed to its end marker in a browser, and compares the pages it loaded with an API replay of its own first query.
  • Triggers. workflow_dispatch, and a nominal daily schedule at 17 2 * * * (02:17 UTC). Both run only from main. GitHub may start a scheduled run late, or drop it under load.
  • Disk. The site_a profile’s modelled whole-job need is 13.55 GB, against the 14 GB advertised runner. That is a fail-closed risk, not a guarantee: measured free space decides each run.

Runs.

runtriggerref / headsite_a jobsite_b job
43687177547 attempt 1push (temporary branch-scoped trigger, since removed)feat/1514-ui-scale @ c2d02cf6132729983049 success132729982839 success
43687177547 attempt 2rerunsame132731850375 success132731850584 success
43696272244workflow_dispatchmain @ 9f430645132757387430 success132757387103 success

All three runs recorded identical counts:

  • site_a:
    • 2,005 assets and 900 posts (the 863 dataset posts plus the collection-post backfill), with 0 per-asset preview jobs;
    • chen.wei can list 1,993 assets: the 2,005 live assets minus the 12 mature rows that user does not own, which are the only rows the listing excludes. 1,111 of those 1,993 rows come back as restricted placeholders that keep their id, counted at page size 36, so they are listed but their fields are withheld;
    • chen.wei sees 598 of the 900 posts, with 24 short non-final post pages at page size 36;
    • the browser loaded 25 feed pages to the end marker.
  • site_b: 503 assets and 203 posts. Both principals see exactly the admin set; the browser loaded 6 feed pages.
  • Disk: measured peak growth on the site_a job was about 7.8 GB, on runners with about 91 GB free at the start.

The first scheduled run had not been observed when this was written. Its first nominal slot after the schedule reached main (d812efd2) was 2026-11-16T02:17Z. A scheduled result is recorded when one exists; the runs above are not a substitute for it.

A published edition is never changed: its tag, its release and its assets stay as they are. Replacing an asset would not help anyway, because the pinned manifest SHA-256 would refuse the new bytes. Rollback therefore moves CI, never the data.

A bad edition, or a pin or workflow regression.

  1. Revert through a reviewed PR. Restore the last known-good entry in seed/contracts/ci-datasets.v1.json: repository, tag, commit, release_json_sha256, release_format and capacity. Also revert the workflow or helper change if that is the cause. The PR goes through dev, then main, like any change. The scheduled ui-nightly and ui-scale runs read main.
  2. Re-verify. The PR’s own CI fetches the restored edition and verifies it again. That means scripts/ci/public-datasets.sh and seed/scripts/fetch_release_dataset.py against the pinned manifest, then sha256sum -c on the extracted tree.
  3. Then, separately, a corrected edition. A fixed dataset is a NEW edition (for example v2026.11.02, in the v2 layout). It passes the gates above and is pinned by its own reviewed PR.

When there is no older public edition (as now: v2026.11.01 and v2026.11.01-sm are the first):

  • A pin regression is reverted to the pins file’s previous commit.
  • If the edition itself is bad, there is nothing earlier to pin. The affected dataset jobs (ui-pr, ui-nightly, ui-scale) FAIL CLOSED and stay red until a corrected edition is published and pinned.
  • They are NOT pointed at the former private site_b, a private path, a cache or an unverified download.
  • These dataset jobs are not the required checks: main requires Integration and the Seed guard. So a red dataset job does not block other merges, but it stays visible on every run.

Release rule. For v0.11.1 and every later application release, a red ui-pr, ui-nightly or ui-scale dataset job on the release line STOPS the release until it is resolved.

  • Resolved means a green run on the release candidate’s main SHA, or a documented root cause fixed through the reviewed procedure above.
  • Not a waiver: that branch protection does not mark these jobs as required is a technicality, not a waiver. A red dataset job is never released over, re-pinned around, or treated as passed.

Never:

  • edit or delete a published tag or asset to “fix” CI;
  • skip or soften verification, or mark a failed fetch as passed;
  • substitute unpinned bytes.

Withdrawing a published edition from public view, for example for a rights reason, is an owner decision outside this procedure. CI moves off it through the pins first.

A local candidate archive by itself authorizes no upload; for v2026.11.01 the owner’s decision does. The obsolete pre-exclusion site_a candidate must never be uploaded. Nothing in this ADR calls the current sources rights-cleared.

  • ADR 0058 (demo seed dataset tiers; site_b replacement amendment, accepted 2026-11-15)
  • ADR 0080 (coverage-selected CI fixture; nightly two-contract amendment)
  • ADR 0097, ADR 0098 (archive outputs, deterministic assembly)
  • docs/proposals/v0.11.1-ci-datasets-preflight.md
  • seed/scripts/release_dataset.py, seed/scripts/test_release_dataset.py
  • seed/scripts/fetch_release_dataset.py, seed/scripts/test_fetch_release_dataset.py
  • seed/contracts/ci-datasets.v1.json, scripts/ci/public-datasets.sh, scripts/ci/ephemeral-studio-b-tls.sh, infra/docker/ci/docker-compose.hosted.yml
  • .github/workflows/ui-pr.yml, .github/workflows/ui-nightly.yml, .github/workflows/ui-scale.yml, scripts/ci/scale_expected.py
  • GitHub, About releases: docs.github.com/en/repositories/releasing-projects-on-github/about-releases