Reported by a user · Fixed in #1849 · Released in 2.3.1

Somebody filled in a form with a table question - a purchase list, say, with an item, a quantity and a date on each row - submitted it, and downloaded the submission as a PDF.

The PDF showed the question's heading, ITEMS, and then nothing. The fields after it printed normally. It read exactly as though the person had skipped the table.

They had not. The rows were there the whole time.

The same submission was missing the same rows on screen and in the CSV export, for the same reason. It was reported as a PDF fault only because that is where somebody happened to look first, and a heading with a gap under it is more obviously wrong than a table you were not expecting to see.

A table question keeps its two halves in two different places:

• the columns - what the table asks - live in form_fields.config
• the answers - what somebody typed - live in form_submission_data.field_value

The endpoint that feeds the submissions screen selected this:

SELECT id, field_type, label, is_deleted
  FROM form_fields

No config. So the answers came back and the columns did not.

FormLogic.gridColumns() then had nothing to read and returned an empty list, and every reader laid three rows of perfectly good data out against no columns at all:

All three failed silently. Nothing errored, nothing was logged, and the JSON response genuinely contained the answers - so anyone checking the data would have found it present and concluded the export was fine.

Because the sibling endpoint was already right. FormsService::collectionSubmissions(), which feeds the collection view, selects options and config - so the very same table rendered correctly there.

🔑 Two endpoints feeding the same JavaScript different field shapes. One was complete, one was not, and the incomplete one was the one most people use. Nothing compared them.

The endpoint selects options, config as well, which is a one-line change.

The part worth more than the fix: both endpoints now return the same field shape, rather than this one merely being made to work. The test asserts that equality directly - it checks that the submissions endpoint returns every field key the collection view gets - so the two cannot drift apart again without something saying so.

• The new test builds a form with a table, submits against it, and drives the real endpoint with a forged analyst session - asserting the columns arrive, their labels are intact, and a choice column keeps its options.
• Run against the old query as a control it fails 6 of 12.
• The row-count assertion passes even in that control run, which is the detail that proves the diagnosis: the answers were never the missing half.

Nothing was lost and nothing needs repairing. The rows were always in the database, and they reappear as soon as you upgrade.

It is worth re-opening any submission you looked at and assumed had been left blank - on screen and in any CSV you exported, not only in a PDF.

• Forms - what a table question is
• The approvals inbox said "Error" and nothing else - the same field type, a different reader, found the same week
• Bugs resolved