Wrong PERCENT_RANK computation in Firebird 5

[lukaseder]

Run this query:

select
  a,
  percent_rank() over (order by a desc) "p1"
from (
  select 1 a from rdb$database union all
  select 2 a from rdb$database union all
  select 3 a from rdb$database union all
  select 4 a from rdb$database
) t
order by a asc

The correct result as produced by Firebird 4 was:

1	1
2	0.66666666666667
3	0.33333333333333
4	0

Firebird 5.0.3 now produces:

1	0.1111111111111111
2	0.16666666666666666
3	0.3333333333333333
4	0.0

I'm using Firebird from here: https://hub.docker.com/r/firebirdsql/firebird.

A workaround, I can calculate percent_rank() over (order by a desc) myself using rank() over (order by a desc) - 1.0) / (count(*) over () - 1.0)

[sim1984]

Fixed in
#8970
#8971
#8972

[dyemanov]

@sim1984 I see you've fixed it in the v4 branch too, but the author says it works correctly in v4. Please clarify.

[sim1984]

@sim1984 I see you've fixed it in the v4 branch too, but the author says it works correctly in v4. Please clarify.

Perhaps the author was testing an older version of Firebird 4.0. This bug was introduced by another fix #8589, which is present everywhere PERCENT_RANK appears.

[lukaseder]

Indeed, I was testing 4.0.1 as provided here:
https://dbfiddle.uk/_dyqncBf

[mrotteveel]

This also seems to indicate a lack of test coverage for PERCENT_RANK.

[lukaseder]

This also seems to indicate a lack of test coverage for PERCENT_RANK.

Probably, a smaller data set is being used? Here, the result is correct:

select
  a,
  percent_rank() over (order by a desc) "p1"
from (
  select 1 a from rdb$database union all
  select 2 a from rdb$database
) t
order by a asc

[sim1984]

To test the fix for this bug, it's best to use this set:

select
  a,
  percent_rank() over (order by a desc) "p1"
from (
  select 1 a from rdb$database union all
  select 2 a from rdb$database union all
  select 3 a from rdb$database union all
  select 4 a from rdb$database union all
  select 5 a from rdb$database
) t
order by a asc;

There will be no infinite fractions or rounding errors here.

           A                      p1
============ =======================
           1       1.000000000000000
           2      0.7500000000000000
           3      0.5000000000000000
           4      0.2500000000000000
           5       0.000000000000000