1 datasets.bundles.repack

datasets.bundles.repack

Atomic physical root/rollout composition with exact actor-source remapping.

Storage owners retain their schemas and writers. This owner coordinates their new identities, validates the QH join and a real CPU batch, and publishes the whole derived bundle once. Acquisition receipts remain archived unchanged.

1.1 Classes

Name	Description
BundleRepackResult	Published composition, exact consumer paths and transformation receipt.

1.1.1 BundleRepackResult

datasets.bundles.repack.BundleRepackResult(
    directory,
    roots,
    rollouts,
    receipt_path,
)

Published composition, exact consumer paths and transformation receipt.

1.1.1.1 Attributes

Name	Description
directory	Atomically published directory containing every new dependency.
roots	Exact final actor configurations for source-qualified QH admission.
rollouts	Final whole-chain rollout stores grouped by actor, physical split and campaign assignment.
receipt_path	Receipt authenticating original inventories and child transformations.

1.2 Functions

Name	Description
repack_dataset_bundle	Merge/split a complete explicit root/rollout dependency population.

1.2.1 repack_dataset_bundle

datasets.bundles.repack.repack_dataset_bundle(
    roots,
    rollouts,
    destination,
    *,
    samples_per_root=None,
    rows_per_shard=64,
    read_chunk_rows=1,
    chains_per_store=8,
)

Merge/split a complete explicit root/rollout dependency population.

Old root manifest/version, sample index/key, scene/snippet, physical split, shard/row remain in authenticated mappings and verbatim acquisition archives. New root-local sample indices and all rollout source/target/chain/state/ candidate/successor IDs are remapped by their storage owners. Campaign split is retained separately. Each new source cohort receives its recomputed split-manifest hash and transformation-owned configuration summaries.

All roots must have compatible representation/scientific contracts, all rollout sources must resolve exactly, and unsupported retained schemas reject. Distinct replicas remain distinct unless full inventories prove aliases. No generation, inference or relabeling runs. Existing destinations reject. Stage failure/interruption leaves no partial published bundle; original inventories are rechecked before the final directory rename.

Real QH source admission and CPU batch collation run before and after the transformation. This is consumer integrity proof, not trained-fit admission or empirical qualification. Root tensors are bounded by shard rows; rollout materialization is bounded by complete chains per output store.

# datasets.bundles.repack { #aria_nbv.datasets.bundles.repack } `datasets.bundles.repack` Atomic physical root/rollout composition with exact actor-source remapping. Storage owners retain their schemas and writers. This owner coordinates their new identities, validates the QH join and a real CPU batch, and publishes the whole derived bundle once. Acquisition receipts remain archived unchanged. ## Classes | Name | Description | | --- | --- | | [BundleRepackResult](#aria_nbv.datasets.bundles.repack.BundleRepackResult) | Published composition, exact consumer paths and transformation receipt. | ### BundleRepackResult { #aria_nbv.datasets.bundles.repack.BundleRepackResult } ```python datasets.bundles.repack.BundleRepackResult( directory, roots, rollouts, receipt_path, ) ``` Published composition, exact consumer paths and transformation receipt. #### Attributes | Name | Description | | --- | --- | | [directory](#aria_nbv.datasets.bundles.repack.BundleRepackResult.directory) | Atomically published directory containing every new dependency. | | [roots](#aria_nbv.datasets.bundles.repack.BundleRepackResult.roots) | Exact final actor configurations for source-qualified QH admission. | | [rollouts](#aria_nbv.datasets.bundles.repack.BundleRepackResult.rollouts) | Final whole-chain rollout stores grouped by actor, physical split and campaign assignment. | | [receipt_path](#aria_nbv.datasets.bundles.repack.BundleRepackResult.receipt_path) | Receipt authenticating original inventories and child transformations. | ## Functions | Name | Description | | --- | --- | | [repack_dataset_bundle](#aria_nbv.datasets.bundles.repack.repack_dataset_bundle) | Merge/split a complete explicit root/rollout dependency population. | ### repack_dataset_bundle { #aria_nbv.datasets.bundles.repack.repack_dataset_bundle } ```python datasets.bundles.repack.repack_dataset_bundle( roots, rollouts, destination, *, samples_per_root=None, rows_per_shard=64, read_chunk_rows=1, chains_per_store=8, ) ``` Merge/split a complete explicit root/rollout dependency population. Old root manifest/version, sample index/key, scene/snippet, physical split, shard/row remain in authenticated mappings and verbatim acquisition archives. New root-local sample indices and all rollout source/target/chain/state/ candidate/successor IDs are remapped by their storage owners. Campaign split is retained separately. Each new source cohort receives its recomputed split-manifest hash and transformation-owned configuration summaries. All roots must have compatible representation/scientific contracts, all rollout sources must resolve exactly, and unsupported retained schemas reject. Distinct replicas remain distinct unless full inventories prove aliases. No generation, inference or relabeling runs. Existing destinations reject. Stage failure/interruption leaves no partial published bundle; original inventories are rechecked before the final directory rename. Real QH source admission and CPU batch collation run before and after the transformation. This is consumer integrity proof, not trained-fit admission or empirical qualification. Root tensors are bounded by shard rows; rollout materialization is bounded by complete chains per output store.