Add support for quantile()

[halldc]

Would it be possible to add support for the quantile() function?

Quite a few databases support PERCENTILE_CONT and PERCENTILE_DISC, which I think could make this possible (e.g. Google BigQuery, PostgreSQL, and Redshift).

I'd be willing to help of course, but would need some pointers on where to start.

[hadley]

@edavidaja that is perfect, thank you!

[hadley]

A few more tweaks to make it a bit easier for me to parse:

Aggregation function:

• bigquery: PERCENTILE_CONT(x, p)
• hive: PERCENTILE(x, array(p1, p2, ...))
• impala:
• monetdb: QUANTILE(x, p)
• mysql: not supported
• sqlite: not supported
• teradata: APPROX_PERCENTILE(x, percentage)

Window function:

• sql-server: PERCENTILE_CONT(p) WITHIN GROUP (ORDER BY x)
• postgres: PERCENTILE_CONT(p) WITHIN GROUP (ORDER BY x)
• redshift: PERCENTILE_CONT(p) WITHIN GROUP (ORDER BY x)
• salesforce: PERCENTILE_CONT(p) WITHIN GROUP (ORDER BY x)
• mariadb: PERCENTILE_CONT(p) WITHIN GROUP (ORDER BY x)

[hadley]

For the databases that support the PERCENTILE_CONT() window function, I don't see how we can use this in aggregation (i.e. summarise()) context.

The SQL server docs seem to be the only place that addresses this problem and suggests using DISTINCT:

SELECT DISTINCT DepartmentName  
,PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY BaseRate)  
    OVER (PARTITION BY DepartmentName) AS MedianCont  
,PERCENTILE_DISC(0.5) WITHIN GROUP (ORDER BY BaseRate)  
    OVER (PARTITION BY DepartmentName) AS MedianDisc  
FROM dbo.DimEmployee;  

So maybe the best we can do is supply translations for those as window functions, and then suggest the user use distinct()?

[hadley]

Oops, those aren't window functions, but are "ordered-set aggregate" functions, and they work just fine with GROUP BY:

library(DBI)

con <- dbConnect(RPostgres::Postgres())
dbGetQuery(con, "
  SELECT cyl, PERCENTILE_CONT(0.5) WITHIN GROUP (ORDER BY mpg) 
  FROM mtcars 
  GROUP BY cyl  
")
#>   cyl percentile_cont
#> 1   4            26.0
#> 2   6            19.7
#> 3   8            15.2

^{Created on 2019-02-06 by the reprex package (v0.2.1.9000)}

[halldc]

Thanks @hadley! 🎉

[hadley]

Looks like the link to teradata was actually to teradata's distribution of presto. It seems like teradata actually is ansi compliant: https://docs.teradata.com/reader/756LNiPSFdY~4JcCCcR5Cw/RgAqeSpr93jpuGAvDTud3w