refs_here is the refs file(s) for the slice on screen (cheap);
refs is every refs file for this variable (a scan downloads them all).
Both have i (flat chunk index), one *_c chunk coordinate per dimension, path, offset, size.
This page reads the reference stores in
mdsumner/virtualized:
BRAN2023 daily fields from 180-192 monthly netCDF files on NCI THREDDS, indexed as kerchunk Parquet.
remote/*.parq stores were made by VirtualiZarr (Python) on NCI;
zarr/*.zarr stores were made by blocklist (R, rhdf5 and GDAL multidim).
Nothing is preprocessed for this page. It fetches .zmetadata as JSON and reads
refs.N.parq files on raw.githubusercontent.com with HTTP range requests: the map and profiles use
hyparquet (footer plus a row window), and the SQL tab loads
DuckDB-Wasm on first use.
Chunk index i is the C-order position in the chunk grid; it lives in file
floor(i / record_size) at row i mod record_size.
bytes mode needs only the index: compressed chunk size is a surprisingly good picture of the field (land, sea ice, fronts and noise all compress differently). values mode fetches the referenced byte ranges from THREDDS, then undoes zlib and the HDF5 shuffle filter, applies the fill value, scale_factor and add_offset, and draws the result. It only works if the data server allows cross-origin range requests from the browser.
refs via switches how the reference rows are read. hyparquet reads the Parquet footer and a row window directly. DuckDB SQL runs the slice as a SQL query over the same refs files (shown under the status line): row windows per file, chunk coordinates computed from the flat index, and predicates on the dimensions held fixed by the sliders. The map, bytes and values are identical either way.
view reprojects the display into any PROJ projection (orthographic globe, polar stereographic, Equal Earth, UTM, Lambert conic, EPSG polar and regional grids, or a custom PROJ string with optional {lon0}, {lat0}, {zone} and {south}). Centred presets take a centre (type it, use the tapped point, or double-tap); reach caps the initial extent to a great-circle distance from the centre or pole, so polar and azimuthal views do not stretch out to the whole globe. The data are not resampled or copied: each screen pixel is inverse-projected back to the grid as it is stored, using a coarse mesh of exact inverses interpolated within cells (exact per pixel along the limb and wraps). Chunk boundaries, coastline, GEBCO contours and the selected chunk are forward-projected over it, so you can see what a chunk looks like on the globe. Works in both bytes and values modes.
rows / cols choose any two dimensions to display (for example depth x longitude for a section); the others get sliders. Zooming to a chunk also stretches the colours to that chunk; all resets both. GEBCO reads the GEBCO 2026 overview on source.coop through the virtual-COG byte refs in virtual-cog-pilot (about 15 MB once, zlib and TIFF predictor undone here), then draws the contour at the current depth level, or the seafloor along a depth section.