hscida 0.4.4

HSCI data access provider

hscida

hscida is a small Python/R data access package for local or remote data files. It provides lazy dataframe and SQL access to datasets such as Parquet or CSV, including files stored remotely in object storage.

By default, the package reads configuration details such as dataset locations and access credentials from .env/.env.secret files. Then, users can access the configured data using the same conventions from Python or R:

• DuckDB discovers dataset files from a configurable GLOB_PATTERN.
• Remote HTTPS/S3 access is configured with DuckDB INIT_SQL.
• Datasets are registered lazily as DuckDB views only when first requested.
• Python callers can work with lazy DuckDB relations, SQLFrame (PySpark) dataframes, and Narwhals lazy frames, pulling these into Polars dataframes, Pandas dataframes, or rendering the lazy queries as SQL strings.
• R callers get a DBI/DuckDB connection and lazy dplyr tables.

Installation

From PyPI: https://pypi.org/project/hscida/ From R-Universe: https://hsci-r.r-universe.dev/hscida

Configuration

Data access is configured through environment variables. In normal downstream projects these come from .env and .env.secret files in the project root.

.env is usually checked in and contains non-sensitive defaults, such as public S3/HTTPS paths, DuckDB config, and DuckDB initialization SQL.

.env.secret is not checked in. It typically contains credentials for remote data access, such as object storage keys or service tokens. Place it in the project root next to .env:

my-analysis-project/
|-- .env
|-- .env.secret
`-- ...

Ask the downstream project maintainer for the expected .env.secret contents.

Environment Variables

GLOB_PATTERN is a DuckDB table expression used to discover files for a dataset. It may use {projroot} and {dataset} placeholders. For example:

GLOB_PATTERN=glob('{projroot}/data/{dataset}/*.parquet')

INIT_SQL, plus any other variables whose names start with INIT_SQL, are concatenated in sorted key order and run when the DuckDB connection starts. This is useful for loading extensions, creating secrets, setting S3 endpoints, or installing project-specific macros. The ordering is lexical, so use zero-padded suffixes such as INIT_SQL_010 and INIT_SQL_020 if fragments may reach two digits.

INIT_SQL=INSTALL httpfs; LOAD httpfs;
INIT_SQL_010=CREATE SECRET s3_secret (...);
INIT_SQL_020=SET s3_region='auto';

DUCKDB_CONFIG is a comma-separated list of DuckDB configuration values:

DUCKDB_CONFIG=parquet_metadata_cache=true,preserve_insertion_order=false,enable_fsst_vectors=true

PROJROOT overrides the project root used for {projroot} substitution. If it is unset, hscida uses the current project root detected by the language-specific helper (here in R, hereutil in Python).

Remote and Local Data

By default, downstream projects commonly read data directly from remote object storage. Public read-only data can often be accessed through HTTPS without credentials. S3-compatible access, private buckets, or write access usually requires credentials in .env.secret and matching DuckDB startup SQL.

Project may also support a local mirror to avoid repeated remote fetches. For example, a project may use a GLOB_PATTERN that prefers data/filter/ if files exist locally, while falling back to remote paths otherwise. A typical mirror command might look like:

rclone sync s3://project-data data/local

For very large datasets, it can be useful to limit the local mirror:

rclone sync s3://project-data data/local --max-size 500M

The local data directory should usually be gitignored. Re-running the same rclone sync command keeps the mirror up to date.

Python Usage

from hscida import DataAccess

with DataAccess() as da:
    rel = da.duckdb_dataframe("my_dataset")
    preview = rel.limit(10).pl()

DataAccess can expose the same dataset through several lazy dataframe APIs:

with DataAccess() as da:
    duckdb_rel = da.duckdb_dataframe("my_dataset")
    sqlframe_df = da.spark_dataframe("my_dataset")
    narwhals_rel = da.narwhals_duckdb_dataframe("my_dataset")
    narwhals_sqlframe = da.narwhals_spark_dataframe("my_dataset")

You can also start from SQL and convert between supported representations:

with DataAccess() as da:
    query = da.duckdb_dataframe_from_sql("SELECT * FROM my_dataset LIMIT 10")
    polars_df = da.to_polars(query)
    sql = da.to_sql(query, pretty=True)

Short aliases are available for interactive work:

• da.ddf() / da.duckdb_dataframe()
• da.sdf() / da.spark_dataframe()
• da.nddf() / da.narwhals_duckdb_dataframe()
• da.nsdf() / da.narwhals_spark_dataframe()
• da.d() / da.to_duckdb()
• da.s() / da.to_spark()
• da.n() / da.to_narwhals()
• da.p() / da.to_polars()
• da.q() / da.to_sql()

R Usage

library(hscida)

da <- data_access()
on.exit(DBI::dbDisconnect(da$con, shutdown = TRUE), add = TRUE)

tbl <- da$f("my_dataset")

tbl |>
  dplyr::filter(year >= 2020) |>
  dplyr::collect()

da$f("my_dataset") lazily registers the dataset as a DuckDB view, then returns a dplyr table backed by DuckDB. You can pass explicit paths to bypass GLOB_PATTERN:

tbl <- da$f("my_dataset", "data/my_dataset/part-000.parquet", replace = TRUE)

Development

Run the Python tests with:

uv run pytest

Run the R tests against the local source tree with:

Rscript -e 'pkgload::load_all(export_all = FALSE); testthat::test_dir("tests/testthat")'