hubvault v0.0.1

First public release

hubvault v0.0.1 is the first public release of an API-first, embedded, portable repository for local ML artifacts.

This release ships a working local repository surface with:

• Git-like commits, trees, refs, tags, reflog, reset, and merge semantics
• Hugging Face style file APIs for upload, download, listing, and metadata access
• Detached download and snapshot views so read paths do not expose writable repository truth
• A Phase 15 SQLite truth-store design centered on repo-root metadata.sqlite3
• Chunked large-file storage with filesystem payload bytes still kept in objects/blobs/*.data and chunks/packs/*.pack
• Public maintenance APIs and CLI commands for verification, storage analysis, garbage collection, and history squashing
• A git-like CLI exposed as both hubvault and hv

Release hardening for v0.0.1

For this first release, the packaging and documentation surface was tightened up before tagging:

• fixed the local make package path by adding a Python 3.7 compatible build dependency to the maintained requirement sets
• refreshed setup.py metadata, project URLs, and package classifiers
• switched the public documentation links to Read the Docs
• aligned the docs index structure and Sphinx metadata for the new public docs entry point
• refreshed the GitHub repository description, homepage, and topic tags

Current performance snapshot

Benchmark evidence is currently based on the preserved Phase 15 records under build/benchmark/phase15/, measured on Linux x86_64 with CPython 3.10.10.

Standard/full snapshot:

• large upload: 230.69 MiB/s
• large range read: 1113.59 MiB/s
• cold file download: 846.98 MiB/s
• warm file download: 13761.47 MiB/s
• cache-heavy warm download: 19704.43 MiB/s
• deep history listing: 15221.94 ops/s
• recursive tree listing: 31185.03 ops/s
• heavy non-fast-forward merge: 126.65 MiB/s
• squash history with cleanup: 146.83 MiB/s
• chunk threshold sweep: 74.20 MiB/s
• small-file read-all path: 5.76 MiB/s, 1473.64 ops/s

Pressure snapshot:

• large upload: 332.13 MiB/s
• large range read: 910.23 MiB/s
• cold file download: 422.80 MiB/s
• warm file download: 637608.97 MiB/s on cache/view-hit paths
• cache-heavy warm download: 39457.46 MiB/s

The current performance profile is clear:

• metadata-heavy and maintenance-heavy paths improved meaningfully after the SQLite truth-store migration
• large-range reads under pressure are stronger than the pre-migration baseline
• the main remaining optimization targets are warm-download paths, small-file hot reads, and detached-view read-side short-circuiting

Notes

This is a 0.0.1 first release, not a stability promise. The project is already usable as a serious local artifact repository foundation, but it should still be treated as early-stage software while the hot-read path continues to be optimized.