Support percentile_cont via approx_percentile

[kwannoel]

Now that we support approx percentile, perhaps we can support streaming percentile_cont via approx_percentile too, by simply setting relative_error to 0.

Most of the work just needs to take place in the frontend.

[chagelo]

SELECT percentile_cont(x) WITHIN GROUP (ORDER BY column1) FROM table1; gets nonzero, while SELECT approx_percentile(x, 0) WITHIN GROUP (ORDER BY column1) FROM table1; always gets zero, but I'm not familar with "streaming" and hdr, and I'm not sure whether it's correct that approx_percentile return zero.

[kwannoel]

SELECT percentile_cont(x) WITHIN GROUP (ORDER BY column1) FROM table1; gets nonzero, while SELECT approx_percentile(x, 0) WITHIN GROUP (ORDER BY column1) FROM table1; always gets zero, but I'm not familar with "streaming" and hdr, and I'm not sure whether it's correct that approx_percentile return zero.

It's a bug. Thanks for catching it. Will submit a PR to fix it shortly.

[kwannoel]

Actually it is tricky to support it via approx_percentile. That's because we will have to override the semantics of bucket_id to handle relative_error = 0.0. In order to keep the precision of original value when relative_error = 0.0, we have to store bucket_id (this is also the original value when relative_error = 0.0) as a f32/f64 instead of i32. That means the range of values we can do approx_percentile for will be less.

To avoid this extra complexity and trade-off, we should implement percentile_cont first, and map approx_percentile(x, 0.0) to it instead. So it's the other way around.

Tracking issue to support percentile_cont: #12857.

Closing this issue.