CalCOFI release v2026.07.17

2026-07-17 · 18 tables · 236,577,241 rows · 1.90 GB consolidated

table	rows	size
_spatial_attr.parquet	40,298	121 KB
_spatial.parquet	3,373	25.0 MB
cruise.parquet	691	14 KB
dataset_taxon.parquet	1,781	36 KB
dataset.parquet	13	6 KB
grid.parquet	218	98 KB
lookup.parquet	26	2 KB
measurement_type.parquet	120	7 KB
obs/ 12 parts	17,705,061	229.0 MB
obs_attribute.parquet	452,682	1.8 MB
obs_ctd_full/ supplemental 97 parts	216,427,608	1.41 GB
obs.parquet	17,705,061	225.3 MB
region.parquet	4	1 KB
sample_measurement.parquet	555,623	1.7 MB
sample.parquet	1,385,959	11.8 MB
ship.parquet	49	2 KB
taxon_group.parquet	155	2 KB
taxon.parquet	3,580	69 KB

Sidecars

Machine-readable descriptions of this release.

file	size
RELEASE_NOTES.md	2 KB
catalog.json	2 KB
erd.mmd	7 KB
index.html	12 KB
metadata.json	77 KB
test_results.json	4 KB

Why obs appears twice. Each is published both as a Hive-partitioned directory and as a single .parquet file, on purpose — they are the same rows.

The partitioned form lets DuckDB, R and Python skip whole partitions when you filter on the partition column, so it is the faster choice for most analysis. The single-file form exists because plain HTTPS has no directory listing: browser DuckDB-WASM (which powers calcofi.io/db-query) and other HTTPS-only consumers cannot expand a /**/*.parquet glob against cloud storage, so they need one addressable object. Removing either breaks a real consumer.

Read directly with DuckDB — no download needed. A single-file table:

SELECT * FROM read_parquet('https://storage.googleapis.com/calcofi-db/ducklake/releases/v2026.07.17/parquet/sample.parquet') LIMIT 10;

A partitioned table (note the /**/*.parquet glob and hive_partitioning, which turns the directory names into a real column):

SELECT * FROM read_parquet('https://storage.googleapis.com/calcofi-db/ducklake/releases/v2026.07.17/parquet/obs_ctd_full/**/*.parquet',
hive_partitioning = true) LIMIT 10;