Fix the str_flatten function in Redshift

[hdplsa]

Fixes #804

This pull request adds a Redshift specific translation to str_flatten. Currently the default Postgres translation (string_agg) is used, but it is not supported in Redshift. The equivalent function in Redshift is LISTAGG but it has a slightly different syntax (requires WITHIN GROUP (ORDER BY ...) syntax for ordering).

Please check below the reprex with the expected output:

library(dbplyr)
library(DBI)
library(reprex)

con <- dbConnect(RPostgres::Redshift(), 
                 host = ,
                 dbname = , 
                 port = ,
                 user = , 
                 password = ) 

example_table <- dplyr::tribble(
  ~customer, ~day, ~item,
  "A", 1, "WATER",
  "A", 3, "BREAD",
  "A", 2, "JUICE",
  "B", 1, "APPLE",
  "B", 4, "BANANA",
  "C", 1, "MILK"
)

table_db <- dplyr::copy_to(con, example_table, temporary = T)

table_db %>%
  dplyr::group_by(customer) %>%
  dplyr::summarize(flat_string = str_flatten(item, "-"))
#> # Source:   [?? x 2]
#> # Database: postgres
#> #   []
#>   customer flat_string      
#>   <chr>    <chr>            
#> 1 B        APPLE-BANANA     
#> 2 C        MILK             
#> 3 A        WATER-BREAD-JUICE

table_db %>%
  dplyr::group_by(customer) %>%
  dplyr::summarize(flat_string = str_flatten(item, "-")) %>% 
  dplyr::show_query()
#> <SQL>
#> SELECT "customer", LISTAGG("item", '-') AS "flat_string"
#> FROM "example_table"
#> GROUP BY "customer"

table_db %>%
  dplyr::group_by(customer) %>%
  dbplyr::window_order(day) %>%
  dplyr::mutate(flat_string = str_flatten(item, "-")) 
#> # Source:     [?? x 4]
#> # Database:   postgres
#> #   []
#> # Groups:     customer
#> # Ordered by: day
#>   customer   day item   flat_string      
#>   <chr>    <dbl> <chr>  <chr>            
#> 1 A            1 WATER  WATER-JUICE-BREAD
#> 2 A            2 JUICE  WATER-JUICE-BREAD
#> 3 A            3 BREAD  WATER-JUICE-BREAD
#> 4 C            1 MILK   MILK             
#> 5 B            1 APPLE  APPLE-BANANA     
#> 6 B            4 BANANA APPLE-BANANA

table_db %>%
  dplyr::group_by(customer) %>%
  dbplyr::window_order(day) %>%
  dplyr::mutate(flat_string = str_flatten(item, "-")) %>%
  dplyr::show_query()
#> <SQL
#> SELECT
#>   "customer",
#>   "day",
#>   "item",
#>   LISTAGG("item", '-') WITHIN GROUP (ORDER BY "day") OVER (PARTITION BY "customer") AS "flat_string"
#> FROM "example_table"

^{Created on 2022-03-25 by the reprex package (v2.0.0)}

[hadley]

Are there other functions in redshift that use this syntax? I'm surprised that (say) windowed mean() doesn't need the same syntax? Or is LISTAGG() the same sort of function as PERCENTILE_DISC()?

[hdplsa]

From searching the documentation only the LISTAGG(), PERCENTILE_DISC(), PERCENTILE_CONT(), and ST_COLLECT() functions in redshift require the WITHIN GROUP clause for the ordering. The remaining windowed functions use the more general OVER (PARTITION BY ... ORDER BY ...) clause.

Unfortunately, I cannot explain why LISTAGG() in particular uses the WITHIN GROUP syntax. But this syntax seems to be shared by other database types (e.g. Oracle).

[hadley]

Thanks for the investigation! Unfortunately there seem to be few satisfying explanation for why SQL things are the way they are.

[hadley]

Can you please add a bullet to the top of NEWS.md? It should briefly describe the change and end with (@yourname, #issuenumber).

[hdplsa]

Just added the bullet to the NEWS.md. Thanks @hadley.

[hadley]

Thanks!