Getting Started with Pointblank Operatives

Introduction

The pointblankops package provides specialized data validation operations using lightweight operatives for focused intelligence gathering. Operatives are streamlined alternatives to pointblank agents, designed for efficient row-level failure detection without the overhead of full reporting capabilities.

The use case that this solves is the following:

• data is large so can be out of memory
• we run tests on data to understand which rows fail which test, because downstream we exclude different rows in different situations depending on the purpose of the analysis
• so we need per-row validation results to use in post-processing

Extracting this from an interrogated agent it tedious and memory-intensive.

To preserve memory and allow working on large datasets, operatives focus on extracting validation failures directly, without the full reporting overhead of pointblank agents.

• Per-row validation results are returned in a tidy format, making it easy to integrate with other data processing workflows.
• They can be stored directly in a database or saved to a file format like Parquet for further analysis, all done efficiently with minimal memory footprint.
• Validation failure information can also be returned as a tibble for immediate use in R.

Creating Operatives

Operatives are created using the create_operative() function, which is a lightweight wrapper around pointblank’s create_agent():

# Create test data
test_data <- data.frame(
  batch = c("A", "A", "B", "B", "C"),
  id = c(1, 2, 3, 4, 5),
  value = c(10, NA, 15, 8, 12),
  category = c("X", "Y", "X", "Z", "Y")
)

# Create an operative
operative <- create_operative(test_data, tbl_name = "test_data", label = "Test Operative")

operative

Adding Validation Steps

Just like pointblank agents, operatives can have validation steps added to them:

operative <- operative |>
  col_vals_not_null(columns = vars(value)) |>
  col_vals_between(columns = vars(value), left = 5, right = 20) |>
  col_vals_in_set(columns = vars(category), set = c("X", "Y", "Z"))

Debriefing Operatives

The core functionality is the debrief() function, which extracts only the validation failures:

# Get failures as a tibble
failures <- debrief(operative, row_id_col = c("batch", "id"))

failures
#> # A tibble: 1 × 6
#>   batch id    test_name test_type         column_name failure_details           
#>   <chr> <chr> <chr>     <chr>             <chr>       <chr>                     
#> 1 A     2     step_1    col_vals_not_null value       Failed col_vals_not_null …

Output Options

The debrief() function supports multiple output formats:

1. Return as Tibble (default)

failures <- debrief(operative, row_id_col = c("batch", "id"))

failures
#> # A tibble: 1 × 6
#>   batch id    test_name test_type         column_name failure_details           
#>   <chr> <chr> <chr>     <chr>             <chr>       <chr>                     
#> 1 A     2     step_1    col_vals_not_null value       Failed col_vals_not_null …

2. Save to Parquet File

debrief(operative, 
        row_id_col = c("batch", "id"), 
        parquet_path = "validation_failures.parquet")

read_parquet("validation_failures.parquet")
#> # A tibble: 1 × 6
#>   batch id    test_name test_type         column_name failure_details           
#>   <chr> <chr> <chr>     <chr>             <chr>       <chr>                     
#> 1 A     2     step_1    col_vals_not_null value       Failed col_vals_not_null …

3. Save to Database

con <- DBI::dbConnect(duckdb::duckdb(), ":memory:")

# Copy test data to database
DBI::dbWriteTable(con, "test_data", test_data)

# Create operative from database table
db_operative <- create_operative(test_data) |>
  col_vals_not_null(columns = vars(value)) |>
  col_vals_between(columns = vars(value), left = 5, right = 20)

# Save failures to database table
debrief(db_operative, 
        row_id_col = c("batch", "id"), 
        con = con, 
        output_tbl = "validation_failures")

tbl(con, "validation_failures") |>
  collect()
#> # A tibble: 1 × 6
#>   batch id    test_name test_type         column_name failure_details           
#>   <chr> <chr> <chr>     <chr>             <chr>       <chr>                     
#> 1 A     2     step_1    col_vals_not_null value       Failed col_vals_not_null …

Memory Efficiency

For large datasets, debrief() processes data in chunks to maintain memory efficiency:

# Process in smaller chunks for memory efficiency
failures <- debrief(operative, 
                   row_id_col = c("batch", "id"),
                   chunk_size = 500)  # Process 500 rows at a time

failures
#> # A tibble: 1 × 6
#>   batch id    test_name test_type         column_name failure_details           
#>   <chr> <chr> <chr>     <chr>             <chr>       <chr>                     
#> 1 A     2     step_1    col_vals_not_null value       Failed col_vals_not_null …

Database Compatibility

Operatives work seamlessly with database tables via dbplyr:

con <- DBI::dbConnect(duckdb::duckdb(), ":memory:")
DBI::dbWriteTable(con, "large_table", test_data)

# Create operative from database table
db_operative <- create_operative(dplyr::tbl(con, "large_table")) |>
  col_vals_not_null(columns = vars(value)) |> 
  col_vals_gt(value, 8)

# Debrief processes the query efficiently in the database
failures <- debrief(db_operative, row_id_col = c("batch", "id"))

failures
#> # A tibble: 2 × 6
#>   batch id    test_name test_type         column_name failure_details           
#>   <chr> <chr> <chr>     <chr>             <chr>       <chr>                     
#> 1 A     2     step_1    col_vals_not_null value       Failed col_vals_not_null …
#> 2 B     4     step_2    col_vals_gt       value       Failed col_vals_gt on col…

Supported Validation Types

The following pointblank validation functions are supported:

• col_vals_not_null() / col_vals_null()
• col_vals_between() / col_vals_not_between()
• col_vals_in_set() / col_vals_not_in_set()
• col_vals_gt() / col_vals_gte() / col_vals_lt() / col_vals_lte()
• col_vals_equal() / col_vals_not_equal()

Unsupported validation types are automatically skipped with a message.