Ungrouped (count (distinct col)) in a select projection returns the row count (STR/SYM) or a type error (I64)

[alexey-milovidov]

Summary

(count (distinct col)) used as an ungrouped select projection (no by:)
does not compute a distinct count. On a STR / SYM column it silently returns
the row count; on an integer column it fails with a type error. The same
expression is correct under by:, and the whole-vector form
(count (distinct (at t 'col))) is correct too — so only the ungrouped
select-projection path is wrong.

Reproduction

(set t (table [i s g] (list [1 1 2 2 3] ["aa" "aa" "bb" "bb" "cc"] [9 9 9 9 9])))

(count (distinct (at t 'i)))                    ; 3  correct
(count (distinct (at t 's)))                    ; 3  correct
(select {from: t by: g u: (count (distinct i))}) ; u = 3  correct

(select {from: t u: (count (distinct s))})      ; u = 5  <-- wrong, that is (count s)
(select {from: t u: (count (distinct i))})      ; error: type: distinct: argument must be a list, got i64

Expected

(select {from: t u: (count (distinct s))}) → u = 3, matching both the
grouped form and the whole-vector reducer. An ungrouped aggregate select is just
the one-group case of by:.

Notes

• The STR/SYM case is the dangerous one: no error, and the answer is a
  plausible-looking number. On a 2M-row table the same query returned 1999939
  (the row count) where the correct distinct count was 100665, so it does not
  look obviously wrong at a glance either.
• The integer case at least fails loudly, and the message
  (distinct: argument must be a list, got i64) suggests distinct is being
  applied per row rather than to the column, i.e. the projection is evaluated
  element-wise instead of being lowered to the count-distinct aggregate that
  the by: path uses (ops/idiom.c rw_count_distinct →
  exec_count_distinct).
• Whichever way it is resolved, an error would be better than the current
  silent row count.

Versions

Reproduced on v2.5.14 (release tag, make release) and on dev at
95defb7aef2ff0863e70ab3b79785036cd529b1d, Linux aarch64.

Found while porting Rayforce to
ClickBench, where several queries
are SELECT COUNT(DISTINCT c) FROM hits.

[singaraiona]

Fixed in 55c6e22 (on dev). Root cause: the ungrouped scalar-aggregate fallback evaluated the aggregate's argument per row, so distinct saw one scalar cell at a time — STR/SYM collapsed to the row count, integers errored, as you observed. Whole-column verbs (distinct/asc/desc/reverse) at the argument head are now evaluated once against the full column, matching the by: path and the whole-vector reducer. Regression test in test/rfl/query/ungrouped_count_distinct.rfl. On the 2M-row shape from the report the ungrouped form now agrees with the grouped count. Thanks for the report.