MySQL/MariaDB's algorithm for aggregate_func(DISTINCT x) uses Unique class to remove duplicates. This approach cannot be easily extended to handle the case where rows move in and out of the window frame.

Consider a query

SELECT agg_func(DISTINCT c) OVER (PARTITION BY a ORDER BY b ) FROM ...

Normally, window function code will sort the query output by (a, b). What if we sorted by (a,b,c) instead? We would be able to compute
agg_func(DISTINCT c) easily:

Consider some value of c=$c1. As soon as the first row with c=$c1 goes into window frame, we can compute the current agg_func value $val. All subsequent rows with c=$c1 have the same value agg_func(...)=$val.

The idea in the above comment is incorrect.

All subsequent rows with c=$c1 have the same value agg_func(...)=$val.

what to do about rows that have the same value of c but different value of b? The sorting is done by a,b,c, so these rows won't be adjacent to one another.

http://www.vldb.org/pvldb/vol8/p1058-leis.pdf is a paper that is not relevant to this specific task but it says this on p. 1060:

Finally, it is
important to note that the framing clause only affects some
window functions, namely intra-window navigation functions
(first value, last value, nth value), and non-distinct
aggregate functions (min, max, count, sum, avg). The remaining
window functions (row number, rank, lead, . . . ) and distinct
aggregates are always evaluated on the entire partition.

Note: distinct aggregates are always evaluated on the entire partition.

The standard says this:

13) If the window ordering clause or the window framing clause of the window structure descriptor that

describes the <window name or specification> is present, then no <aggregate function> simply contained

in <window function> shall specify DISTINCT or <ordered set function>.

So, if DISTINCT is used, then ORDER BY is not allowed. Then, frame definition is not allowed either, and we have

• all rows are peers
• the window frame is always the same, and always includes all rows in the partition.

Are you still planning to do it in 10.2? If not, let's change the fix version.