dplyr translation for approx_percentile()

[Sean Wilson]

The default translation of the R function quantile() is incorrect:

> translate_sql(quantile(x, 0.9))
<SQL> PERCENTILE_CONT(0.9) WITHIN GROUP (ORDER BY `x`) OVER ()

I realize that Presto does not implement exact quantiles, but it does provide approx_percentile(). This would enable a translation like

> translate_sql(quantile(x, 0.9))
<SQL> APPROX_PERCENTILE(`x`, 0.9)

This would also allow a translation of median() as a special case. There is an approach to this in dbplyr for Teradata, and it seems like it could be implemented with minimal effort for RPresto, possibly adding a once-per-session warning that the result is approximate.

Is this something we could consider doing? If so, I'd be happy to create a PR.

[Ismail Onur Filiz]

Thanks for identifying the issue. Actually I'd much more prefer disabling the quantile translation instead. We'd want the users to explicitly specify their intention by referencing the approx_percentile function directly if they are OK with an approximation.

[Sean Wilson]

Okay, then I think there are maybe two pieces of work.

1. Implement an R function approx_percentile() that is translated to the matching Presto function. The benefit is that users can call it as they would quantile() without having to know about build_sql() and so on. So, something like mutate(presto_tbl, q = approx_percentile(x, 0.9)) where x is the unquoted name of a column in presto_tbl.

2. Disable the quantile() default translation and throw an error that directs the user to the function described in (1).

How does that sound?

[Ismail Onur Filiz]

You don't need (1), unrecognized functions get passed directly to the backend:

(
  src_presto(...)
  %>% tbl(sql('select 1 as x'))
  %>% summarise(p=approx_percentile(x, 0.1))
  %>% show_query()
)

[Sean Wilson]

I agree. Submitted #120.