Network reads are bounded: omicidx_connect() sets DuckDB’s http_timeout (30 s) and retries each request up to 4 times with backoff, including on HTTP 429, and a failed connect names the release URL and suggests what to do (#16). omicidx_download() retries for up to 6 minutes so a copy of a monthly release waits out HuggingFace’s rate limit.
Monthly releases use HF_TOKEN, if set, for HuggingFace’s higher signed-in rate limit. ?omicidx_connect, the README and the vignette recommend querying "latest" or dated builds in place and downloading monthly releases with omicidx_download().
The vignette and tests use "latest" rather than HuggingFace, so builds do not fail on HuggingFace’s rate limit.
omicidx_connect(version) opens a DuckDB connection to an OmicIDX release: "latest", a dated nightly build "YYYY-MM-DD" or a monthly release "YYYY-MM". Nothing is downloaded up front, and DuckDB extensions are cached in tools::R_user_dir("omicidxr", "cache") rather than ~/.duckdb.
omicidx_release() reports a connection’s release and its provenance from manifest.json.
omicidx_download(version, path, tables) copies a release, or only the chosen tables or schemas, into a local DuckDB file for offline use, checked against manifest.json row counts. It reports the size of the Parquet it will read before it starts.