PoC: read inputs from GitHub Releases (issue #136)

[lbm364dl]

Draft / proof of concept — not for merge. Demonstrates the GitHub-Releases file-store idea from #136 end to end.

What it does

Adds an opt-in input source that fetches <alias>.parquet from a GitHub release into a local cache, wired into the single .download_pin_paths() choke point (R/github_inputs.R + a 4-line hook in R/read_raw_inputs.R). Default behaviour is unchanged — still the Nextcloud pins board. Enable with:

options(whep.input_source = "github")   # repo/tag also configurable via options or env vars

Data

The 10 inputs build_primary_production() needs were mirrored from the current Nextcloud board to a release in eduaguilera/whep_inputs:

• Release: https://github.com/eduaguilera/whep_inputs/releases/tag/poc
• 10 assets, 215 MB total, largest 44 MB — all well under the 2 GiB per-file limit, so no splitting needed for these tabular inputs (answers one of Reconsider input-data hosting: move whep_inputs off the Nextcloud pins board #136's open questions for this subset).

Proof it runs

build_primary_production(start_year = 2015, end_year = 2016) with whep.input_source = "github", cold cache:

• All 10 inputs fetched from the release assets (216 MB pulled into the whep cache).
• Output: 91,363 rows, 12 columns, years 2015–2016, sources FAOSTAT_prod (90,680) + LUH2_grassland (683). Ran in ~82 s over the network.

Deliberately out of scope (PoC)

• Only the parquet download path is routed to GitHub (enough for build_primary_production; whep_read_file's other branches untouched).
• Fixed release tag instead of the per-alias versioning whep_inputs currently tracks — the versioning-scheme decision from Reconsider input-data hosting: move whep_inputs off the Nextcloud pins board #136 still needs settling.
• No tests; reads via base download.file rather than piggyback (kept dependency-free for the demo).

Relates to #136.

🤖 Generated with Claude Code