troubleshoot differences between DB and local behavior about dm

[TSchiefer]

The above was created with dbplyr v2.3.0, with v2.2.0 we get the expected behavior, i.e. a 2-row tibble without the need for compute():

suppressPackageStartupMessages({
  library(dplyr)
  library(tibble)
  library(duckdb)
})
packageVersion("dbplyr")
#> [1] '2.2.0'
lcl_tbl <- tibble(
  a = as.integer(c(1, 2, 1)),
  e = c("c", "b", "c"),
  f = c(TRUE, FALSE, TRUE)
)

src_duckdb <- dbplyr::src_dbi(DBI::dbConnect(duckdb::duckdb()), auto_disconnect = TRUE)

rmt_tbl <- copy_to(src_duckdb, lcl_tbl)
rmt_tbl %>% 
  distinct() %>%
  mutate(aef = row_number(a))
#> # Source:   SQL [2 x 4]
#> # Database: DuckDB 0.5.0 [root@Darwin 22.2.0:R 4.2.2/:memory:]
#>       a e     f       aef
#>   <int> <chr> <lgl> <dbl>
#> 1     1 c     TRUE      1
#> 2     2 b     FALSE     2

^{Created on 2023-02-14 with reprex v2.0.2}

This explains why we see the unexpected behavior on all remote sources. The duckdb version has apparently nothing to do with our problem.

[TSchiefer]

Problem seems to lie in the query created by dbplyr v2.3.0:

suppressPackageStartupMessages({
  library(dplyr)
  library(tibble)
  library(duckdb)
})
packageVersion("dbplyr")
#> [1] '2.3.0'
lcl_tbl <- tibble(
  a = as.integer(c(1, 2, 1)),
  e = c("c", "b", "c"),
  f = c(TRUE, FALSE, TRUE)
)

src_duckdb <- dbplyr::src_dbi(DBI::dbConnect(duckdb::duckdb()), auto_disconnect = TRUE)

rmt_tbl <- copy_to(src_duckdb, name = "rmt_tbl", lcl_tbl)

rmt_tbl %>% 
  distinct() %>% 
  show_query()
#> <SQL>
#> SELECT DISTINCT *
#> FROM rmt_tbl

rmt_tbl %>% 
  distinct() %>%
  mutate(aef = row_number(a)) %>% 
  show_query()
#> <SQL>
#> SELECT DISTINCT
#>   *,
#>   CASE
#> WHEN (NOT((a IS NULL))) THEN ROW_NUMBER() OVER (PARTITION BY (CASE WHEN ((a IS NULL)) THEN 1 ELSE 0 END) ORDER BY a)
#> END AS aef
#> FROM rmt_tbl

^{Created on 2023-02-15 with reprex v2.0.2}

whereas for dbplyr c2.2.0 it was:

# [...] (same as above)
rmt_tbl %>% 
  distinct() %>%
  mutate(aef = row_number(a)) %>% 
  show_query()
#> <SQL>
#> SELECT *, ROW_NUMBER() OVER (ORDER BY a) AS aef
#> FROM (
#>   SELECT DISTINCT *
#>   FROM rmt_tbl
#> ) q01

^{Created on 2023-02-15 with reprex v2.0.2}

It looks like the order of distinct() and mutate() is switched by the translation to SQL code for dbplyr v2.3.0.

[TSchiefer]

As to the problem with flattening:

suppressPackageStartupMessages({
  library(dplyr)
  library(tibble)
  library(duckdb)
})
packageVersion("dbplyr")
#> [1] '2.3.0'
lcl_tbl_1 <- tibble(
  a = 1:2,
  e = letters[1:2],
)

lcl_tbl_2 <- tibble(
  a = 1:2,
  e = letters[3:4],
)
lcl_tbl_3 <- tibble(
  a = 1:2,
  e = letters[5:6],
)
lcl_tbl_4 <- tibble(
  a = 1:2,
  e = letters[7:8],
)
lcl_tbl_1 %>% 
  left_join(rename(lcl_tbl_2, e_2 = e), by = "a") %>% 
  left_join(rename(lcl_tbl_3, e_3 = e), by = "a") %>% 
  left_join(rename(lcl_tbl_4, e_4 = e), by = "a")
#> # A tibble: 2 × 5
#>       a e     e_2   e_3   e_4  
#>   <int> <chr> <chr> <chr> <chr>
#> 1     1 a     c     e     g    
#> 2     2 b     d     f     h


src_duckdb <- dbplyr::src_dbi(DBI::dbConnect(duckdb::duckdb()), auto_disconnect = TRUE)

rmt_tbl_1 <- copy_to(src_duckdb, name = "rmt_tbl_1", lcl_tbl_1)
rmt_tbl_2 <- copy_to(src_duckdb, name = "rmt_tbl_2", lcl_tbl_2)
rmt_tbl_3 <- copy_to(src_duckdb, name = "rmt_tbl_3", lcl_tbl_3)
rmt_tbl_4 <- copy_to(src_duckdb, name = "rmt_tbl_4", lcl_tbl_4)
rmt_tbl_1 %>% 
  left_join(rename(rmt_tbl_2, e_2 = e), by = "a") %>% 
  left_join(rename(rmt_tbl_3, e_3 = e), by = "a") %>% 
  left_join(rename(rmt_tbl_4, e_4 = e), by = "a")
#> # Source:   SQL [2 x 5]
#> # Database: DuckDB 0.6.2-dev1166 [root@Darwin 22.2.0:R 4.2.2/:memory:]
#>       a e     e_2   e_3   e_4  
#>   <int> <chr> <chr> <chr> <chr>
#> 1     1 a     c     e     e    
#> 2     2 b     d     f     f
rmt_tbl_1 %>% 
  left_join(rename(rmt_tbl_2, e_2 = e), by = "a") %>% 
  left_join(rename(rmt_tbl_3, e_3 = e), by = "a") %>% 
  left_join(rename(rmt_tbl_4, e_4 = e), by = "a") %>% 
  dbplyr::sql_render()
#> <SQL> SELECT rmt_tbl_1.*, rmt_tbl_2.e AS e_2, rmt_tbl_3.e AS e_3, rmt_tbl_3.e AS e_4
#> FROM rmt_tbl_1
#> LEFT JOIN rmt_tbl_2
#>   ON (rmt_tbl_1.a = rmt_tbl_2.a)
#> LEFT JOIN rmt_tbl_3
#>   ON (rmt_tbl_1.a = rmt_tbl_3.a)
#> LEFT JOIN rmt_tbl_4
#>   ON (rmt_tbl_1.a = rmt_tbl_4.a)

^{Created on 2023-02-15 with reprex v2.0.2}
In the SELECT part of the query rmt_tbl_3.e is wrongly selected twice, once as e_3 (correct) and once as e_4 (wrong).
This behavior leads to the differences in the flatten-snapshot.

[TSchiefer]

see tidyverse/dbplyr#1158 and tidyverse/dbplyr#1159

[TSchiefer]

Both of these problems will be fixed on GHA after the next patch-release of dbplyr.