pytail 0.1.3

Incremental PyPI Simple API caching mirror

PyTail

pytail is a minimal incremental PyPI caching mirror.

It is no longer a devpi clone. The new server only does four things:

• expose a valid Python Simple Repository API root at /simple/
• lazily fetch and cache /simple/<project>/ from an upstream index
• rewrite file links to a devpi-style root/pypi/+f/... path
• serve cached project pages and cached files on later requests

Plan

1. Implement the smallest Simple API surface that pip and uv actually need: root index, project index, normalized names, file links, and content negotiation for HTML vs JSON responses.
2. Keep the cache incremental: do not mirror the full upstream project list, only remember projects that have already been requested.
3. Use conditional project refresh: cache upstream project pages, preserve ETag, and refresh a project after a TTL with If-None-Match.
4. Keep package files immutable: once a file URL is requested, cache it under +files/ and serve it from a stable devpi-style file route on later requests.
5. Store the index in SQLite: project pages, file metadata, and known-project state live in SQLite, while package bytes live on disk.
6. Handle concurrent requests safely inside one process: per-project and per-file locks avoid duplicate upstream fetches and duplicate downloads.
7. Explicitly drop old devpi goals: no users, no ACLs, no upload API, no index inheritance, no replication, no snapshot format, no mirror whitelist, no web UI.

Spec Coverage

The implementation is intentionally narrow:

• PEP 503 HTML Simple API for /simple/ and /simple/<project>/
• PEP 691 JSON Simple API responses when Accept asks for application/vnd.pypi.simple.v1+json
• preservation of requires-python, yanked, gpg-sig, dist-info-metadata, and core-metadata markers when they are present on the upstream project page
• lazy local root index containing only already-cached projects

The implementation does not currently fetch or synthesize a full global upstream project list. That is deliberate: the root index is only a local catalogue of known projects, while project pages are fetched on demand.

Why This Is Enough

Package resolution for pip and uv depends primarily on per-project Simple API pages. A full pre-fetched mirror root is not required for normal dependency resolution as long as:

• /simple/<normalized-project>/ is available and correct
• file links are reachable
• project metadata such as hashes and requires-python are preserved

Run

cargo run -- \
  --bind 127.0.0.1:3141 \
  --upstream-base-url https://pypi.org \
  --torch-url https://download.pytorch.org/whl/ \
  --cache-dir .cache/pytail

Then point tools at it:

uv pip install --index-url http://127.0.0.1:3141/simple/ requests
pip install --index-url http://127.0.0.1:3141/simple/ requests

PyTorch wheel indexes are also exposed under /pytorch-wheels/, so a command that uses:

pip install torch --index-url https://download.pytorch.org/whl/cu126

can use the local caching endpoint instead:

pip install torch --index-url http://127.0.0.1:3141/pytorch-wheels/cu126

The local path is mapped directly under the configured PyTorch wheels upstream:

/pytorch-wheels/torch/        -> <torch-url>/torch/
/pytorch-wheels/cu126/torch/  -> <torch-url>/cu126/torch/

For packages that need both PyPI and PyTorch wheels, keep the indexes separate:

pip install torch \
  --index-url http://127.0.0.1:3141/pytorch-wheels/cu126 \
  --extra-index-url http://127.0.0.1:3141/simple/

Configuration

• --bind: listen address, default 127.0.0.1:3141
• --upstream-base-url: upstream index origin, default https://pypi.org
• --torch-url: PyTorch wheels upstream root, default https://download.pytorch.org/whl/
• --cache-dir: local cache directory, default .cache/pytail
• --project-cache-ttl-secs: refresh age for cached project pages, default 900
• --request-timeout-secs: upstream HTTP timeout, default 15
• --stats-interval-secs: cache hit-rate log interval, default 60; set to 0 to disable periodic stats
• --verbose: enable debug logging for pytail

Package

Build a Python wheel with maturin:

maturin build --release

Publish to PyPI:

maturin publish --release

The wheel installs the pytail command as a native Rust binary.

Cache Layout

<cache-dir>/
  index.sqlite3
  +files/
    root/
      pypi/
        +f/
          f0a/
            f3fc39e7459b0/
              gradio_client-1.0.2-py3-none-any.whl
          _url/
            <url-digest>/
              <filename>

• SQLite stores project cache entries, parsed file rows, upstream ETag, and known project names
• +files/root/pypi/+f/<first-3-sha256>/<next-13-sha256>/<filename> stores hashed files in the same shape as devpi's filesystem layout
• +files/root/pypi/+f/_url/<url-digest>/<filename> is used only when an upstream file link does not provide a usable sha256 hash
• PyTorch wheel endpoints reuse the same blob store and download pipeline; only their project cache keys and client-facing file URLs are namespaced

Non-Goals

• full PyPI mirroring
• authentication or private indexes
• package upload
• merge of local and upstream package sources
• replica mode or background synchronization