CSV and Excel import: the whole wizard, column by column

The import wizard turns a file into a table in three moves: read the file, show you what it thinks each column is, then — only after you've agreed — validate everything and write. The guessing is honest about being a guess: every column shows its verdict next to real sample values, and every verdict is yours to overrule.

The file

The ⋮ button at the right of the table tabs opens Import data… → Upload a CSV or Excel file. (Starting fresh? The New database picker's Import a spreadsheet card lands in the same wizard.) It takes .csv, .tsv, .xlsx and .xls.

A CSV or TSV has no size limit. Your browser reads the file where it already sits and sends the rows in batches, so nothing has to be uploaded whole — a 500,000-row export is a normal import, with a progress bar and a Stop import button. The only ceiling is the one your database already has: 1,000,000 records.

An Excel workbook is capped at 20 MB and 50,000 rows per run, because a .xlsx is a zip and has to be opened whole to be read. Saving it as CSV removes the cap entirely.

The first row is read as headers. An Excel workbook imports its first sheet only; the header counts the rest — "2 extra sheet(s) ignored" — so nothing vanishes silently. A blank header becomes "Column N", and duplicate headers get numbered, so the default plan is always valid.

What the sniffer looks for

Types are guessed from each column's first 50 non-empty values (you see three of them as samples). In order of precedence:

• Checkbox — every value is true/false/1/0/yes/no.
• Currency — numbers with at most two decimals under a header that smells like money (price, cost, amount, total, fee).
• Number, percent — plain numerics; values ending in %.
• Date — ISO dates (2026-08-05, with or without a time), and slash dates, below. If any sample carries a time, the field includes one.
• Email, web address — every value matches the shape.
• Multi select — every value is a delimited list with 20 or fewer distinct entries overall.
• Single select — 20 or fewer distinct short values.
• Long text — line breaks, or anything past 200 characters.
• Anything else: single line text.

One guard runs before all of that: leading zeros. A column holding 00123 is never guessed as a number, because a number field would eat the zeros — id columns stay text.

Slash dates, the honest way

03/04/2026 is March 4th or April 3rd depending on who wrote the file, and nothing in the value says which. The sniffer looks for a day above 12 — the one piece of real evidence — and sets Read dates as accordingly. When no sample settles it, the control defaults to month-first and says so out loud: "No day above 12 here, so both readings fit — worth a look." That sentence is the difference between an import and a quietly wrong one.

The mapping step

Every column starts as Create new field with the guessed name and type — rename it, change the type (fourteen to pick from, text through multi select), or Skip it. Importing into a table you already have adds the third option, Map to existing field, which has its own guide.

Two field kinds never appear as create targets: links and attachments. A CSV cell has no record ids or file bytes to offer them. Import the plain columns, then wire up linked records inside Tabla.

For select columns, the wizard's full-file scan means every distinct value in the column — not only the sampled ones — becomes a choice, colored from the standard palette. A "select" that turns out to hold more than 100 distinct values is refused by name: "…has 214 distinct values — too many for a select field. Import it as text instead."

Validate everything, then write

Import N rows runs the whole file through the plan before creating anything. A value that won't fit its column's type stops the run with the exact spot:

Row 3: "tall" isn't a number, needed for field "Qty".

Nothing is written on failure — in new-table mode, not even the table. Fix the cell, run again. The only failure this ordering can't prevent is a dropped connection mid-write; the wizard then says plainly that some rows may already have landed, and append mode reports exactly how many.

One import runs at a time per person — a second attempt while one is writing gets "You already have an import running."

Round trips and the API twin

The importer reads Tabla's own export files back cleanly: the UTF-8 byte-order mark is stripped, and the apostrophe the exporter prefixes to formula-like cells ('+20 100…) is removed on the way in, so a phone number survives the loop unchanged.

The same wizard exists as one REST call — POST https://tabladb.com/api/databases/:id/import, multipart, with the file alone returning the sniffed preview and file + plan running the import. The API reference has the exact shapes.

The questions people ask

Excel dates arrive as dates, right? Yes. Real date cells import as dates; date-only cells stay date-only rather than growing a spurious midnight.

Where do imported rows land? In the table, unfiltered. If the view you're looking at has a filter, some imported rows may be filtered out of sight — the footer count tells the truth.

Can I import into a linked table and have links resolve? Not from a file. The Airtable importer rebuilds links because it can read both sides; a lone CSV can't name records.