Calculate median on different grouping sets in Vertica

When you use cube function, it is a Multi Level aggregation. whereas using just group by is single level aggregation. Using cube with APPROXIMATE_MEDIAN is not a supported feature. Hence it is expected behaviour.

You must remember that GROUP BY CUBE or GROUP BY ROLLUP often are multi-pass query access plans, even if you don't code them as such. And it does not hurt, as Vertica's query engine has pipeline parallelism - many parts of the plan are executed by different threads in parallel. So while it might look awkward to see it in code, it's not so un-elegant after all if you UNION SELECT all levels of aggregation together:

WITH
indata(fld_a,fld_b,fld_c) AS (
            SELECT 1,1,  1
  UNION ALL SELECT 1,2,  2
  UNION ALL SELECT 1,3,  3
  UNION ALL SELECT 1,4,  4
  UNION ALL SELECT 1,5,  5
  UNION ALL SELECT 2,1, 10
  UNION ALL SELECT 2,2, 20
  UNION ALL SELECT 2,3, 30
  UNION ALL SELECT 3,4, 40
  UNION ALL SELECT 3,5, 50
  UNION ALL SELECT 3,1,100
  UNION ALL SELECT 3,2,200
  UNION ALL SELECT 3,3,300
  UNION ALL SELECT 3,4,400
  UNION ALL SELECT 3,5,500
) 
SELECT
  'fld_a' AS grain
, fld_a
, NULL::int   AS fld_b
, APPROXIMATE_MEDIAN(fld_c) AS app_med_c
FROM indata
GROUP BY fld_a
UNION ALL
SELECT
  'fld_b'     AS grain
, NULL::int   AS fld_a
, fld_b
, APPROXIMATE_MEDIAN(fld_c) AS app_med_c
FROM indata
GROUP BY fld_b
UNION ALL
SELECT
  'both' AS grain
, fld_a
, fld_b
, APPROXIMATE_MEDIAN(fld_c) AS app_med_c
FROM indata
GROUP BY fld_a,fld_b
ORDER BY 1,2,3
;