Pull actual observation rows from an AQUAVIEW data source — not catalog metadata.

Use this whenever the question needs the measurements themselves (temperature profiles, float positions, buoy readings), rather than which datasets exist. Call list_data_sources first for a source's canonical variable names, its qc_modes, and its engine (gridded "xarray" sources have no rows to pull).

READ 'returned' FIRST. It is one of three values:

• 'complete': 'rows' holds EVERY matching row (or your whole 'limit'). Safe to compute anything from.

• 'head': 'rows' holds the FIRST rows in the source's natural order (see 'head.order'), out of 'head.of' total. These are not a random sample — later rows differ systematically — so never average or take a maximum from them. To get more, set 'limit' up to 'head.max_inline'; beyond that, narrow the request or fetch 'download_url'.

• 'none': no rows. 'refusal.reason' says why. 'unscoped_query' means add a bbox/datetime/depth/filter. 'no_natural_order' means the source has no column to order a partial read by, so no partial set can be given honestly — narrow until it fits, or use 'stats'. The rest are transient; retry.

OTHER FIELDS:

• 'stats': exact minimum, maximum, mean and spread computed over EVERY matching row. Use these for any numeric claim — exact at any data size.

• 'stats_basis': what shaped those numbers. If 'qc' is null the values are UNSCREENED and may include implausible outliers from bad sensor readings; say so, or re-run with a quality filter. When a mode was applied, 'qc_screened' lists the variables it actually checked — a variable not in it was NOT flag-checked, and an empty list means nothing was screened despite 'qc' naming a mode. 'qc_note', when present, explains that the screen drops whole rows, so it shapes every variable's figures.

• 'structure': how many distinct floats/stations/profiles matched — often the real answer ("16 profiles from 5 floats"), which no row count gives.

• 'sample': a genuinely random sample, included only when rows were withheld. For seeing what the data looks like — do NOT compute from it.

• 'download_url': the full result as a file, in 'download_format'. Fetch it if you can run code; otherwise give it to the user. Absent for restricted sources and for matches too large to stage synchronously.

• 'schema': each column's real type. Needed when reading the CSV download, which stores no types — without it a reader infers them and can get them wrong (an id column read as a number loses leading zeros).

• 'omitted': names any part that could not be computed. An absent field is unknown, NOT zero.

Args: source: Source id from list_data_sources (e.g. "gadr", "wod", "ndbc"). variables: Comma-separated canonical variable names (e.g. "temperature,salinity,depth"). bbox: Bounding box "west,south,east,north" (e.g. "-161,18,-154,23"). datetime: ISO-8601 instant or range ("2022-06-01/2022-06-30"; ".." for an open end). depth: Depth range in metres below surface, "min,max" (e.g. "0,200"). filters: JSON object of extra column filters on ordinary columns, e.g. {"data_mode": "D"}. For QUALITY screening prefer qc below — a source that defines modes applies them to each measurement correctly, where filtering a raw flag column yourself is easy to get wrong. qc: Named quality mode, and the right way to screen. Unfiltered ocean data contains sentinel values that make maxima and means physically implausible, so screen before any scientific claim. list_data_sources reports each source's qc_modes; pass one of those here. Passing a mode to a source with none is an error. A source with no modes may still expose per-measurement quality columns you can filter directly via filters — list_data_sources shows its variables. limit: Maximum rows to return, honored up to what fits inline. Omit for a default-sized head. When a response comes back as a head, head.max_inline tells you the largest limit worth asking for on that request. format: Encoding of download_url. Default "parquet" — it keeps every column's type, is ~5x smaller and ~2x faster, and pd.read_parquet(url) reads it. Choose "csv" when the link is going to a person who will open it by hand rather than into code, or when you can fetch a URL but cannot run code; then use schema to set the column types yourself. include_sample: Set False to skip the random sample and return slightly faster. output_format: 'csv' (compact, default), 'json', or 'toon'.