docs/DOWNLOADER_SPEC.md

VfxDB Dataset Downloader Specification

This document is the normative product contract for the VfxDB downloader. It describes what a user asks for, which local files are authoritative, how whole archives are selected, and what must be true before a download is considered successful.

The words must, must not, should, and may are requirements in this document. The implementation must not preserve an older behavior when it conflicts with this specification.

1. Product model

The downloader has two deliberately separate jobs:

1. Install the complete local JSON control data needed to understand VfxDB.
2. Optionally download VDB data using one of three data-selection modes.

Every valid invocation must finish the first job before planning or downloading any VDB archive. JSON control data is mandatory; it is not an optional add-on and there is no --metadata switch.

The only user-facing data unit is a complete tar archive. The downloader must not select individual VDB members from a tar, must not download a tar in order to inspect which sequence it contains, and must not search every remote tar for a sequence. A local category index maps a sequence directly to its tar.

2. Mandatory local JSON control data

For a valid invocation, the downloader must first pin one immutable Hugging Face dataset revision and install all of the following from that same revision:

• Every top-level <Category>/category_index.json in the dataset repository.
• meta/vfxdb_meta.tar.zst, fully unpacked so that every published <Category>/index/*.json is present locally, except JSON files deliberately removed by the IO-bad rule in section 8.
• meta/io_bad_vdbs.jsonl and meta/io_bad_vdbs.meta.json.

Category names are discovered from repository paths that exactly match <Category>/category_index.json. dataset_index.json is not an input to discovery, planning, selection, or installation.

All category indexes must be downloaded, including categories not named in a later category-specific request. Each downloaded category_index.json must retain every upstream row and every upstream field. The only permitted local addition is deleted_bad_io_sample on an IO-bad sample row, as defined in section 8. The downloader must not generate a replacement category index, reduce its sample list, recompute its counts, or remove fields.

After these files are local, all data planning must read the local <Category>/category_index.json files and local IO-bad list. A remote index, dataset_index.json, remote archive scan, or previously generated private index must not silently replace them.

For each sample row, the first component of vdb_path identifies its sequence and therefore its archive directly:

<Category>/<sequence>/<file>.vdb
        -> archives/<Category>/<sequence>.tar

Distinct sequence components are distinct tar units. Within a category, the fixed tar order is the order in which each distinct sequence first appears in that category's local category_index.json. Repeated runs against the same revision therefore produce the same selection. Categories are processed in lexicographic category-name order when an algorithm needs a stable order among categories.

3. Command-line interface

The public command is:

python tools/download_extract_data.py [DATA_ROOT] [OPTIONS]

DATA_ROOT defaults to data/vdbset.

There are exactly three mutually exclusive data-selection modes:

3.1 Preset mode

python tools/download_extract_data.py /data/vfxdb --preset smoke
python tools/download_extract_data.py /data/vfxdb --preset medium
python tools/download_extract_data.py /data/vfxdb --preset full

--preset must not be combined with --percentage, --category, or --max-samples.

3.2 All-category percentage mode

python tools/download_extract_data.py /data/vfxdb --percentage 10

One percentage applies to the complete set of categories. Percentage mode does not accept --category or --max-samples. The percentage must be greater than zero and at most 100.

3.3 Category and maximum-sample mode

python tools/download_extract_data.py /data/vfxdb \
  --category CloudWave \
  --max-samples 1000

python tools/download_extract_data.py /data/vfxdb \
  --category CloudWave \
  --category SurfaceFire \
  --max-samples 1000

--category may be repeated. The same positive --max-samples value applies separately to every requested category. This mode requires both arguments; a category without a maximum, or a maximum without a category, is an error.

3.4 Interactive terminal interface

python tools/download_extract_data.py /data/vfxdb --tui

--tui is an interface, not another selection mode. It must reject --preset, --percentage, --category, and --max-samples; destination, revision, cache, and the initial IO-bad preference may be provided as defaults. The ordinary bare command must never enter the TUI automatically.

The interactive order is normative:

1. Confirm the destination and whether to start mandatory-file preparation. Revision, cache, and the explicit diagnostic IO-bad override are grouped under advanced settings; the ordinary path does not require users to learn the internal bad-sample filtering mechanism.
2. Prepare and validate every category index, both named IO-bad control files, and every indexed <Category>/index/*.json exactly as the plain CLI does.
3. Present all six choices using the installed local category indexes: required JSON only, Smoke, Medium, Full, global percentage, or selected categories with one shared maximum sample count.
4. Generate the exact DownloadPlan from those local indexes.
5. Show revision, destination, cache, exact tar count, usable sample counts, per-category allocation, whole-tar target overshoot, conservative network and installed-volume upper bounds, and free space on the destination and cache filesystems. An explicit IO-bad override must be shown as a warning; default filtering remains transparent.
6. Require an explicit three-way decision before the first VDB tar download: download, return to selection, or quit. Returning to selection must not repeat mandatory JSON preparation. Full mode requires the exact text FULL; invalid input is retried rather than treated as an implicit quit.

Rejecting the preparation prompt performs no Hub access and no destination write. Quitting from selection or the exact VDB plan returns success, keeps all mandatory JSON, and downloads no VDB tar. Ctrl-C returns 130 after restoring the terminal; completed cache objects and installed tars remain reusable. Other errors return 1 and must identify the stage, current archive/path, concrete cause, and rerun action without printing a traceback.

The Rich view must not change downloader semantics. It uses core events and Hugging Face's own byte progress where supported; older Hub versions suppress their fallback bar only for the active call and use a stage indicator without replacing the HF transfer implementation. It must not leave Hub progress globally disabled after returning. Overall tar progress, current-file bytes, and per-sample JSON work are distinct units and must never overwrite one another. Non-TTY --tui exits 2 before Hub or filesystem access. Below 60 columns the view reduces to single-column numeric status; NO_COLOR and narrow terminals retain the same labels and decisions without depending on color or Unicode for meaning. A TTY reporting TERM=dumb uses periodic plain progress lines instead of cursor-addressed Live rendering.

The following infrastructure options may be combined with any valid mode or with a bare invocation:

• --include-bad: retain selected IO-bad files after extraction.
• --revision REVISION: select the Hugging Face revision to pin.
• --cache-dir PATH: select the Hugging Face Hub cache location. Complete Hub objects and completed local tar installations are reused on rerun.

These are not additional data-selection modes. In particular, --include-bad must not change archive selection, requested quotas, ordering, or normal-sample counts.

Old selection controls such as --all, --limit, and --metadata are not part of this interface and must be rejected rather than assigned legacy semantics.

4. Bare invocation

A command with no data-selection option is valid:

python tools/download_extract_data.py /data/vfxdb

It must:

1. Install and validate every category_index.json.
2. Download and fully unpack meta/vfxdb_meta.tar.zst.
3. Install and validate the IO-bad list and its integrity JSON.
4. Apply the section 8 IO-bad rule to the installed JSON files and category index annotations.
5. Download no archives/<Category>/<sequence>.tar files.
6. Clearly state that no VDB data was downloaded.
7. Print concise examples for Smoke, Medium, Full, all-category percentage, and category plus maximum-sample modes.

This is the supported way to prepare or refresh the local control data without downloading VDB data. It must not silently choose a starter dataset.

5. Presets

5.1 Smoke

Smoke selects the first 2 tar units from every category, using each category's fixed tar order.

• A category with fewer than 2 tars contributes all of its tars.
• Its shortfall is not redistributed to other categories.
• Selection and download remain whole-tar operations.

Formally, for category c with ordered tar list T_c, Smoke selects T_c[:2].

5.2 Medium

Let A be the number of distinct tar units across all categories. Medium's global target is:

ceil(A * 0.20)

Tars are assigned by balanced rounds:

1. In a round, take the next tar from every category that still has one.
2. Stop immediately when the global target is reached.
3. Once a category is exhausted, skip it in later rounds.
4. Continue filling from categories with remaining tars until the exact target number of tars has been selected.

Thus categories receive the same number of tars while possible, and categories with more tars supply the remainder only after shorter categories are exhausted. The selected tar count is exactly the target; archive boundaries do not require further rounding because the target is already a tar count.

5.3 Full

Full selects every tar unit in every category.

6. All-category percentage mode

For requested percentage P, let A be the total number of distinct tar units across all categories. The global target is:

ceil(A * P / 100)

Selection uses the same balanced-round algorithm as Medium. A single P always applies to all categories together; per-category percentages are not supported. A request for 100 percent is exactly equivalent to Full, including the deterministic category-major download order.

The console summary must report the requested percentage, total available tar count, selected tar count, and per-category selected tar count before data download begins.

7. Category and maximum-sample mode

Every row of a local category_index.json represents one sample. Samples known to be IO-bad do not count toward --max-samples, regardless of whether --include-bad is present.

For each requested category independently:

1. Walk complete tars in that category's fixed order.
2. Add the number of non-IO-bad sample rows belonging to each tar.
3. Select the complete tar and continue until the cumulative normal-sample count is greater than or equal to --max-samples.
4. If the requested maximum is greater than the category's total normal-sample count, select the category in full.

The result deliberately rounds upward at the tar boundary. For example, if selected tars contain 900 normal samples and the next complete tar raises the count to 1,100, a request for 1,000 selects that next tar and yields 1,100 normal samples. No tar is partially extracted to enforce the numeric maximum.

Tars containing zero normal samples may occur. Walking the fixed order still selects such a tar before continuing; its zero contribution must not terminate the loop or change the order.

7.1 EnvironmentalFog

EnvironmentalFog is single-frame data. Each VDB row in its local category_index.json counts as one independent sample; the downloader must not invent a previous frame, temporal relationship, or synthetic sequence. Its storage and download unit remains the complete tar, exactly like every other category:

• Smoke counts its tar units, not individual frames.
• Medium and percentage mode include its tar units in balanced allocation.
• Category maximum mode accumulates its independent non-IO-bad VDB rows per complete tar and rounds upward at the tar boundary.

8. IO-bad behavior

The IO-bad list is a mandatory lower-level rule, not a fourth selection mode. Normal user-visible quotas and sample counts exclude IO-bad rows. The archive plan must be identical with and without --include-bad; the flag controls only whether bad files found inside already-selected tars are retained.

The cleanup occurs after complete downloads and extraction:

8.1 Default behavior

For every row identified by meta/io_bad_vdbs.jsonl, the downloader must:

1. Remove the installed VDB if it exists.

2. Remove the row's meta_path file under <Category>/index/ if it exists.

3. Preserve the row and all of its original fields in <Category>/category_index.json.

4. Add or set:

   "deleted_bad_io_sample": true
   

The deletion and annotation are idempotent. A missing bad file is already in the desired state and is not an error. A successfully finalized default data root exposes neither the bad VDB nor its corresponding per-sample JSON.

8.2 Explicit --include-bad

For the same IO-bad rows, the downloader must retain or restore files from the mandatory JSON archive and from any selected VDB tars, while preserving the same tar plan and normal-sample counts. It must add or set:

"deleted_bad_io_sample": false

The annotation is required even when bad files are explicitly retained. It is the only behavior change in a category index between the two policies. A bad row whose tar was not selected remains in the category index with false, but no unselected VDB is downloaded merely because --include-bad is present.

Switching policies on a later run must converge to the newly requested state: default mode removes previously retained bad files and changes annotations to true; --include-bad restores applicable JSON files, retains bad members of newly extracted selected tars, and changes annotations to false.

Archive validation manifests are internal validation inputs. They must not be installed as an alternative user-facing index that makes a deleted bad sample appear available.

9. Whole-tar download and extraction

For every selected tar, the downloader must:

1. Construct its exact remote path directly from the category and sequence learned from the local category index.
2. Ask huggingface_hub to download that one complete file at the pinned revision.
3. Reuse the standard Hugging Face cache. The normal script path uses bounded HTTP reads and retries so a stalled transfer returns control to the retry loop instead of holding the destination lock forever.
4. Validate the complete tar and its published sequence manifest before making extracted VDB files visible in the final destination.
5. Extract every VDB member in the tar as a unit; member-level selection is forbidden.
6. Apply the IO-bad cleanup policy only after the complete tar is installed.

The complete tar object must always match its published Hugging Face content identity before extraction. Legacy sequence manifests across multiple categories contain empty or stale per-member sha256 values, so those fields are advisory rather than an integrity root. The downloader instead requires the verified outer tar, exact manifest membership, canonical paths, and exact agreement between indexed and tar-member byte sizes. A nonempty member digest must still be syntactically valid.

If several requested samples map to the same tar, the tar is planned, downloaded, verified, and extracted once. Existing verified tars and files are reused on rerun. An interrupted run must be restartable without redownloading a valid complete Hugging Face object or treating a partial extraction as complete. The currently transferring tar may restart from zero when the installed huggingface_hub version does not preserve partial objects across processes; every earlier completed tar remains reusable.

The metadata archive is also downloaded through huggingface_hub and its cache. Safe extraction must reject absolute paths, .. traversal, links, devices, duplicate conflicting members, malformed JSON, and content that does not match published integrity information.

The downloader must read meta/vfxdb_meta.tar.zst as one zstd frame using the Python zstandard package directly into streaming tar extraction. It must not start an external zstd process or write an intermediate uncompressed tar. After tar end markers are reached, it must drain and validate the compressed stream through exact frame EOF so truncation, trailing data, and checksum failures remain fatal.

10. Local layout

After a normal download, the relevant public layout is:

DATA_ROOT/
├── <Category>/
│   ├── category_index.json
│   ├── index/
│   │   └── <sample>.json
│   └── <sequence>/
│       └── <sample>.vdb
└── meta/
    ├── io_bad_vdbs.jsonl
    └── io_bad_vdbs.meta.json

Every category index is the complete upstream index with only the IO-bad annotation added. It may therefore reference VDBs not selected by the current preset or quota; downstream consumers determine local availability by checking whether the referenced VDB and per-sample JSON exist. The downloader must not crop the index to hide unselected normal data.

11. Status and console behavior

Before downloading data tars, the console must make the request legible without requiring knowledge of repository internals. It must show at least:

• Pinned revision.
• Destination and Hugging Face cache location.
• Selected mode and its value.
• Total and selected tar counts, globally and by category.
• Normal-sample counts for category maximum mode.
• Estimated remote bytes when the repository exposes sizes.
• Whether selected IO-bad files will be removed or retained.

Progress must distinguish mandatory JSON preparation, tar download, tar verification, extraction, IO-bad cleanup, and completion. A cache hit must be reported as reuse, not as a fresh network download. The final message must report actual installed tar and normal-sample counts and explicitly mention any upward rounding caused by a complete tar.

A bare invocation must end successfully after mandatory JSON preparation and must prominently say that it downloaded no VDB data. It must print examples of all three data-selection modes rather than silently doing nothing.

12. State, cache, and reruns

• All files for a destination must come from one pinned immutable Hugging Face commit. A branch or tag supplied by the user is resolved before installation.
• The implementation may keep minimal internal state needed for revision pinning, locking, atomic installation, and resumption. That state must not become another source of dataset membership.
• Hugging Face's standard cache is authoritative for downloaded blobs. The implementation should not build a competing content cache.
• A second identical run must verify and reuse complete local work.
• A nonempty destination without this downloader's revision state must be rejected. Silently adopting files from an old script or manual copy could mix dataset revisions; migration, if added later, must be explicit and verified.
• Ctrl-C, a process crash, or a transient network failure must leave a state from which rerunning the same command safely continues.
• Concurrent writers to the same destination must be prevented by a lock; readers must never observe a partially committed tar extraction or partially rewritten category index.
• Publishing category indexes and IO-bad controls must use a durable transaction. A failure rolls every already-published control back, and a later invocation recovers a transaction interrupted by process death.
• Cache and extraction space requirements on the same filesystem must be aggregated before data transfer. Space reserved for control-file backups and atomic publication must also be checked before publication begins.
• IO-bad cleanup and index annotation must be repeatable and must reconcile a change between default and --include-bad policy.
• Because one IO-bad policy transition spans many files and category indexes, the downloader must keep a durable internal transition marker until both controls and files reach the requested policy. A data root with that marker is not a completed dataset and downstream consumers must treat it as unavailable. A failed or interrupted transition is completed by rerunning the downloader; no mixed IO-bad policy may be exposed as successful state.

13. Failure semantics

Command-line mode conflicts and invalid values must fail before network or filesystem work. For a syntactically valid request, mandatory JSON preparation must complete before any data tar begins downloading.

The downloader must exit nonzero with a concrete path and cause when any of the following occurs:

• A category index is missing, malformed, internally inconsistent, or contains an unsafe path.
• The IO-bad list fails its integrity JSON.
• meta/vfxdb_meta.tar.zst is missing, corrupt, unsafe, or incomplete.
• Any sample row lacks a usable meta_path, or its referenced JSON is absent from the full metadata archive. This includes IO-bad rows because --include-bad must be able to restore their published JSON before data-tar selection begins.
• A requested category does not exist locally.
• A selected tar is missing, corrupt, unsafe, or does not match its sequence manifest.
• Available disk space or inodes are insufficient for a safe installation.
• The destination is already pinned to an incompatible revision and cannot be reconciled without mixing revisions.
• IO-bad cleanup or atomic category-index annotation cannot finish.

No later data tar may be attempted after a mandatory-control failure. A failed tar must not leave any of its VDB members marked complete. Cached complete downloads and diagnostic artifacts should be retained so a rerun can resume or so the failure can be investigated.

14. Required tests and acceptance criteria

Tests must be deterministic, must not depend on the live Hugging Face dataset, and must exercise the public CLI as well as selection helpers. A small fixture repository must include multiple uneven categories, a short category, IO-bad rows, a bad-only tar, EnvironmentalFog frames, valid archive manifests, and a full metadata archive.

14.1 CLI contract

• Bare invocation installs all mandatory JSON and requests zero data tars.
• Bare output names Smoke, Medium, Full, percentage, and category maximum usage.
• Each valid mode parses with infrastructure options.
• Preset plus percentage/category/maximum is rejected.
• Percentage plus category/maximum is rejected.
• Category without maximum and maximum without category are rejected.
• Zero, negative, and over-100 percentages are rejected.
• Nonpositive maximum-sample values and unknown categories are rejected.
• Removed legacy selectors are rejected.

14.2 Mandatory JSON ordering

• All category indexes, all published <Category>/index/*.json files, and both IO-bad files are local and validated before the first data-tar request.
• All category indexes are installed even for category-specific mode.
• Planning reads local category indexes after installation.
• No selection code reads dataset_index.json.
• Missing referenced JSON for a normal row fails before any data-tar request.

14.3 Fixed ordering and presets

• Tar order follows sequence first appearance in each local category index, including nonnumeric sequence names.
• Smoke selects 2 tars per category, selects a short category in full, and does not redistribute its shortfall.
• Medium selects exactly ceil(total_tars * 0.20).
• Medium gives categories equal tar counts while possible and then fills only from categories with remaining tars.
• Full selects every tar exactly once.
• Repeated planning against identical inputs yields byte-for-byte identical plans.

14.4 Percentage selection

• Several percentages, including fractional percentages and 100, select exactly ceil(total_tars * P / 100) tars.
• Uneven categories are balanced round by round and exhausted categories are skipped.
• Percentage cannot be scoped to named categories.

14.5 Category maximum and EnvironmentalFog

• The same maximum is applied independently to every repeated category.
• The last complete tar is retained when it pushes the sample count above the maximum.
• A maximum above a category's total selects the category in full.
• IO-bad rows do not contribute to the cumulative normal-sample count.
• A zero-normal-sample tar does not break accumulation or reorder later tars.
• Every EnvironmentalFog VDB row counts once; no temporal grouping is invented.
• EnvironmentalFog still downloads and extracts whole tars.

14.6 IO-bad behavior

• Default and --include-bad produce exactly the same tar plan and normal quota counts.
• Default removes both the bad VDB and its referenced per-sample JSON and sets deleted_bad_io_sample to true.
• --include-bad retains applicable files and sets the same key to false.
• All original index rows, row order, top-level fields, and row fields remain unchanged apart from that one annotation.
• Cleanup is idempotent when files are already absent.
• Rerunning in the opposite policy reconciles files and annotations.
• Unselected bad VDBs are not fetched merely because --include-bad is set.

14.7 Archive, cache, resume, and safety

• Every selected tar is requested once and every member is extracted before IO-bad cleanup; no selected-member extraction path exists.
• Archive paths are derived directly from local indexes; the implementation never scans all remote tars looking for sequences.
• A cache hit avoids a fresh network transfer.
• After an interrupted run, every complete cached tar and every complete local installation is reused; an incomplete current transfer may restart according to the installed Hugging Face Hub version. Partial extraction is never exposed as complete.
• A corrupted cached object is detected and recovered through a fresh Hugging Face download when possible.
• Unsafe tar members, manifest mismatches, truncated archives, disk-full failures, and concurrent destination writers fail without exposing partial successful state.

14.8 Release gate

The implementation is acceptable only when:

1. The complete automated downloader test suite passes.
2. A clean fixture run passes for bare, Smoke, Medium, Full, percentage, and repeated-category maximum modes.
3. Default and --include-bad filesystem trees match the expected policy.
4. An interruption/rerun integration test proves completed-tar cache reuse and atomic local installation.
5. Every installed category-index meta_path reference resolves to the exact validated per-sample JSON object published by the same dataset revision.

Any validation-only runtime instrumentation must remain outside the production downloader.