Database write-back

This article reviews two possible scenarios for “writing back” predictions to the database, and without importing the data into memory first. Both scenarios share a common model preparation method.

Example setup

To keep the example reproducible, a SQLite database will be used to simulate a larger scale deployment.

First the data is prepared in memory. The article will use the nycflights13::flights data, with a couple of modifications

library(dplyr)
library(tidypredict)
library(randomForest)
library(dbplyr)

flights_table <- nycflights13::flights %>%
  mutate(
    current_score = 0, 
    flight_id = row_number()
    ) 

A new database is created using RSQLite.

library(DBI)

con <- dbConnect(RSQLite::SQLite(), path = ":memory:")
db_fligths <- copy_to(con,flights_table )

Model preparation

A sample is downloaded from the database for modeling. This example already selects the needed variables.

df <- db_fligths %>%
  select(dep_delay, hour, distance) %>%
  head(1000) %>%
  collect() 

A linear model is fitted using lm()

model <- lm(dep_delay ~ ., data = df)

It is highly recommendable to always run tidypredict_test() to make sure that the predictions are in line with what the predict() command returns.

The lm and glm models contain the data that they were fitted with, so it’s easy to run a test. tidypredict_test() uses the model’s internal data set by default

tidypredict_test(model)
#> tidypredict test results
#> Difference threshold: 1e-12
#> 
#>  All results are within the difference threshold

In cases where the model re-fitting is automated, the tidypredict_test() function returns an alert in case the threshold is exceeded, which can be used to fail the automated script.

if(tidypredict_test(model)$alert) stop("Threshold exceeded!")

Scenario 1 - Update scores

In this scenario, The table that supplies the term values is also the recipient of the new score. This is done by updating a specific field in the table. For this example, the field is called current_score. The following SQL UPDATE statement should work in most databases:

library(dbplyr)

update_statement <- build_sql("UPDATE flights_table SET current_score  = ", tidypredict_sql(model, con = con), con = con)

update_statement
#> <SQL> UPDATE flights_table SET current_score  = -3.5984422918702 + (`hour` * 1.38710560882252) + (`distance` * -0.00307606912118567)

This statement can be then passed on to the database team, via documentation or an automated process. In cases where the analyst has the responsibility to run the new SQL statement, or if R is being used to automate the scoring, the next line can be used:

dbSendQuery(con, update_statement)
#> <SQLiteResult>
#>   SQL  UPDATE flights_table SET current_score  = -3.5984422918702 + (`hour` * 1.38710560882252) + (`distance` * -0.00307606912118567)
#>   ROWS Fetched: 0 [complete]
#>        Changed: 336776

Here is a sample of the newly populated field:

db_fligths %>%
  select(current_score) %>%
  head(10) 
#> Warning: Closing open result set, pending rows
#> # Source:   lazy query [?? x 1]
#> # Database: sqlite 3.38.5 []
#>    current_score
#>            <dbl>
#>  1       -0.969 
#>  2       -1.02  
#>  3       -0.0128
#>  4       -1.51  
#>  5        2.38  
#>  6        1.13  
#>  7        1.45  
#>  8        4.02  
#>  9        1.82  
#> 10        2.47

Scenario 2- Append new scores

There may be the need to retain not only the new score, but when it was determined and its history. This is usually possible because the source of record possesses a unique key identifier per entity, such as transaction ID or customer ID. In the example, flights_id is the unique identifier.

In this example, the new scores will be stored in a new table called daily_scores. The following code is part of the example preparation, it creates the table and seeds it with a single row. This is not the best way to create an empty table, but it’ll do for the purposes of this example.

dbWriteTable(con, "daily_scores", 
             tibble(
               flight_id = 0,
               score = 0,
               date = ""
             ))

The plan is to use a SQL statement that most vendors support, is called INSERT INTO SELECT. The idea is to use dplyr laziness to prepare the data transformation and predictions, but it’s not going to be executed until is parsed into SQL and sent as part of a another statement. The INSERT INTO SELECT statement allows for the results of a query to be saved in a table, and without leaving the database.

In this example, predictions are going to be executed for just the records in the month of December. The data is filtered, and then tidypredict_to_column() is used to create the new fit field. The results are then transformed to match to the structure of the new daily_scores table.

new_predictions <- db_fligths %>%
  filter(month == 12) %>% 
  tidypredict_to_column(model, vars = "score") %>%
  select(
    flight_id,
    score) %>%
  mutate(date = "01/01/2018")
insert_scores <- build_sql("INSERT INTO daily_scores ", sql_render(new_predictions, con = con), con = con)
insert_scores
#> <SQL> INSERT INTO daily_scores SELECT `flight_id`, `score`, '01/01/2018' AS `date`
#> FROM (SELECT `flight_id`, -3.5984422918702 + (`hour` * 1.38710560882252) + (`distance` * -0.00307606912118567) AS `score`
#> FROM `flights_table`
#> WHERE (`month` = 12.0))

As in the first scenario, this statement can be then passed on to the database team, via documentation or an automated process. In cases where the analyst has the responsibility to run the new SQL statement, or if R is being used to automate the scoring, the next line can be used:

dbSendQuery(con, insert_scores)
#> <SQLiteResult>
#>   SQL  INSERT INTO daily_scores SELECT `flight_id`, `score`, '01/01/2018' AS `date`
#> FROM (SELECT `flight_id`, -3.5984422918702 + (`hour` * 1.38710560882252) + (`distance` * -0.00307606912118567) AS `score`
#> FROM `flights_table`
#> WHERE (`month` = 12.0))
#>   ROWS Fetched: 0 [complete]
#>        Changed: 28135

A simple table join can be used to confirm that the new update worked. For real life scenarios, a more sophisticated query should be performed in order to only get the latest score. For this example, we simple filter on the same date we inserted

tbl(con, "daily_scores") %>%
  inner_join(tbl(con, "flights_table"), by = "flight_id") %>%
  filter(date == "01/01/2018") %>%
  select(dep_delay, hour, distance, score, date)
#> Warning: Closing open result set, pending rows
#> # Source:   lazy query [?? x 5]
#> # Database: sqlite 3.38.5 []
#>    dep_delay  hour distance   score date      
#>        <dbl> <dbl>    <dbl>   <dbl> <chr>     
#>  1        14    23     1617 23.3    01/01/2018
#>  2        18    23     1576 23.5    01/01/2018
#>  3        -7     5      529  1.71   01/01/2018
#>  4         5     5     1400 -0.969  01/01/2018
#>  5        -4     5     1089 -0.0128 01/01/2018
#>  6       -10     5     1576 -1.51   01/01/2018
#>  7        -4     5      569  1.59   01/01/2018
#>  8         1     5     1416 -1.02   01/01/2018
#>  9       -11     6      214  4.07   01/01/2018
#> 10       -10     6     1065  1.45   01/01/2018
#> # … with more rows