Perform multiple rowwise validations for joint validity

The conjointly() validation function, the expect_conjointly() expectation function, and the test_conjointly() test function all check whether test units at each index (typically each row) all pass multiple validations. We can use validation functions that validate row units (the col_vals_*() series), check for column existence (col_exists()), or validate column type (the col_is_*() series). Because of the imposed constraint on the allowed validation functions, the ensemble of test units are either comprised rows of the table (after any common preconditions have been applied) or are single test units (for those functions that validate columns).

Each of the functions used in a conjointly() validation step (composed using multiple validation function calls) ultimately perform a rowwise test of whether all sub-validations reported a pass for the same test units. In practice, an example of a joint validation is testing whether values for column a are greater than a specific value while adjacent values in column b lie within a specified range. The validation functions to be part of the conjoint validation are to be supplied as one-sided R formulas (using a leading ~, and having a . stand in as the data object). The validation function can be used directly on a data table or with an agent object (technically, a ptblank_agent object) whereas the expectation and test functions can only be used with a data table.

Usage

conjointly(
  x,
  ...,
  .list = list2(...),
  preconditions = NULL,
  segments = NULL,
  actions = NULL,
  step_id = NULL,
  label = NULL,
  brief = NULL,
  active = TRUE
)

expect_conjointly(
  object,
  ...,
  .list = list2(...),
  preconditions = NULL,
  threshold = 1
)

test_conjointly(
  object,
  ...,
  .list = list2(...),
  preconditions = NULL,
  threshold = 1
)

Arguments

x: A data frame, tibble (tbl_df or tbl_dbi), Spark DataFrame (tbl_spark), or, an agent object of class ptblank_agent that is created with create_agent().
...: A collection one-sided formulas that consist of validation functions that validate row units (the col_vals_*() series), column existence (col_exists()), or column type (the col_is_*() series). An example of this is ~ col_vals_gte(., vars(a), 5.5), ~ col_vals_not_null(., vars(b)).
.list: Allows for the use of a list as an input alternative to ....
preconditions: An optional expression for mutating the input table before proceeding with the validation. This can either be provided as a one-sided R formula using a leading ~ (e.g., ~ . %>% dplyr::mutate(col = col + 10) or as a function (e.g., function(x) dplyr::mutate(x, col = col + 10). See the Preconditions section for more information.
segments: An optional expression or set of expressions (held in a list) that serve to segment the target table by column values. Each expression can be given in one of two ways: (1) as column names, or (2) as a two-sided formula where the LHS holds a column name and the RHS contains the column values to segment on. See the Segments section for more details on this.
actions: A list containing threshold levels so that the validation step can react accordingly when exceeding the set levels. This is to be created with the action_levels() helper function.
step_id: One or more optional identifiers for the single or multiple validation steps generated from calling a validation function. The use of step IDs serves to distinguish validation steps from each other and provide an opportunity for supplying a more meaningful label compared to the step index. By default this is NULL, and pointblank will automatically generate the step ID value (based on the step index) in this case. One or more values can be provided, and the exact number of ID values should (1) match the number of validation steps that the validation function call will produce (influenced by the number of columns provided), (2) be an ID string not used in any previous validation step, and (3) be a vector with unique values.
label: An optional label for the validation step. This label appears in the agent report and for the best appearance it should be kept short.
brief: An optional, text-based description for the validation step. If nothing is provided here then an autobrief is generated by the agent, using the language provided in create_agent()'s lang argument (which defaults to "en" or English). The autobrief incorporates details of the validation step so it's often the preferred option in most cases (where a label might be better suited to succinctly describe the validation).
active: A logical value indicating whether the validation step should be active. If the validation function is working with an agent, FALSE will make the validation step inactive (still reporting its presence and keeping indexes for the steps unchanged). If the validation function will be operating directly on data (no agent involvement), then any step with active = FALSE will simply pass the data through with no validation whatsoever. Aside from a logical vector, a one-sided R formula using a leading ~ can be used with . (serving as the input data table) to evaluate to a single logical value. With this approach, the pointblank function has_columns() can be used to determine whether to make a validation step active on the basis of one or more columns existing in the table (e.g., ~ . %>% has_columns(vars(d, e))). The default for active is TRUE.
object: A data frame, tibble (tbl_df or tbl_dbi), or Spark DataFrame (tbl_spark) that serves as the target table for the expectation function or the test function.
threshold: A simple failure threshold value for use with the expectation (expect_) and the test (test_) function variants. By default, this is set to 1 meaning that any single unit of failure in data validation results in an overall test failure. Whole numbers beyond 1 indicate that any failing units up to that absolute threshold value will result in a succeeding testthat test or evaluate to TRUE. Likewise, fractional values (between 0 and 1) act as a proportional failure threshold, where 0.15 means that 15 percent of failing test units results in an overall test failure.

Value

For the validation function, the return value is either a ptblank_agent object or a table object (depending on whether an agent object or a table was passed to x). The expectation function invisibly returns its input but, in the context of testing data, the function is called primarily for its potential side-effects (e.g., signaling failure). The test function returns a logical value.

Column Names

If providing multiple column names in any of the supplied validation steps, the result will be an expansion of sub-validation steps to that number of column names. Aside from column names in quotes and in vars(), tidyselect helper functions are available for specifying columns. They are: starts_with(), ends_with(), contains(), matches(), and everything().

Preconditions

Providing expressions as preconditions means pointblank will preprocess the target table during interrogation as a preparatory step. It might happen that a particular validation requires a calculated column, some filtering of rows, or the addition of columns via a join, etc. Especially for an agent-based report this can be advantageous since we can develop a large validation plan with a single target table and make minor adjustments to it, as needed, along the way.

The table mutation is totally isolated in scope to the validation step(s) where preconditions is used. Using dplyr code is suggested here since the statements can be translated to SQL if necessary (i.e., if the target table resides in a database). The code is most easily supplied as a one-sided R formula (using a leading ~). In the formula representation, the . serves as the input data table to be transformed (e.g., ~ . %>% dplyr::mutate(col_b = col_a + 10)). Alternatively, a function could instead be supplied (e.g., function(x) dplyr::mutate(x, col_b = col_a + 10)).

Segments

By using the segments argument, it's possible to define a particular validation with segments (or row slices) of the target table. An optional expression or set of expressions that serve to segment the target table by column values. Each expression can be given in one of two ways: (1) as column names, or (2) as a two-sided formula where the LHS holds a column name and the RHS contains the column values to segment on.

As an example of the first type of expression that can be used, vars(a_column) will segment the target table in however many unique values are present in the column called a_column. This is great if every unique value in a particular column (like different locations, or different dates) requires it's own repeating validation.

With a formula, we can be more selective with which column values should be used for segmentation. Using a_column ~ c("group_1", "group_2") will attempt to obtain two segments where one is a slice of data where the value "group_1" exists in the column named "a_column", and, the other is a slice where "group_2" exists in the same column. Each group of rows resolved from the formula will result in a separate validation step.

If there are multiple columns specified then the potential number of validation steps will be m columns multiplied by n segments resolved.

Segmentation will always occur after preconditions (i.e., statements that mutate the target table), if any, are applied. With this type of one-two combo, it's possible to generate labels for segmentation using an expression for preconditions and refer to those labels in segments without having to generate a separate version of the target table.

Actions

Often, we will want to specify actions for the validation. This argument, present in every validation function, takes a specially-crafted list object that is best produced by the action_levels() function. Read that function's documentation for the lowdown on how to create reactions to above-threshold failure levels in validation. The basic gist is that you'll want at least a single threshold level (specified as either the fraction of test units failed, or, an absolute value), often using the warn_at argument. This is especially true when x is a table object because, otherwise, nothing happens. For the col_vals_*()-type functions, using action_levels(warn_at = 0.25) or action_levels(stop_at = 0.25) are good choices depending on the situation (the first produces a warning when a quarter of the total test units fails, the other stop()s at the same threshold level).

Briefs

Want to describe this validation step in some detail? Keep in mind that this is only useful if x is an agent. If that's the case, brief the agent with some text that fits. Don't worry if you don't want to do it. The autobrief protocol is kicked in when brief = NULL and a simple brief will then be automatically generated.

YAML

A pointblank agent can be written to YAML with yaml_write() and the resulting YAML can be used to regenerate an agent (with yaml_read_agent()) or interrogate the target table (via yaml_agent_interrogate()). When conjointly() is represented in YAML (under the top-level steps key as a list member), the syntax closely follows the signature of the validation function. Here is an example of how a complex call of conjointly() as a validation step is expressed in R code and in the corresponding YAML representation.

R statement:

agent %>% 
  conjointly(
    ~ col_vals_lt(., columns = vars(a), value = 8),
    ~ col_vals_gt(., columns = vars(c), value = vars(a)),
    ~ col_vals_not_null(., columns = vars(b)),
    preconditions = ~ . %>% dplyr::filter(a < 10),
    segments = b ~ c("group_1", "group_2"),
    actions = action_levels(warn_at = 0.1, stop_at = 0.2), 
    label = "The `conjointly()` step.",
    active = FALSE
  )

YAML representation:

steps:
- conjointly:
    fns:
    - ~col_vals_lt(., columns = vars(a), value = 8)
    - ~col_vals_gt(., columns = vars(c), value = vars(a))
    - ~col_vals_not_null(., columns = vars(b))
    preconditions: ~. %>% dplyr::filter(a < 10)
    segments: b ~ c("group_1", "group_2")
    actions:
      warn_fraction: 0.1
      stop_fraction: 0.2
    label: The `conjointly()` step.
    active: false

In practice, both of these will often be shorter as only the expressions for validation steps are necessary. Arguments with default values won't be written to YAML when using yaml_write() (though it is acceptable to include them with their default when generating the YAML by other means). It is also possible to preview the transformation of an agent to YAML without any writing to disk by using the yaml_agent_string() function.

Examples

For all examples here, we'll use a simple table with three numeric columns (a, b, and c). This is a very basic table but it'll be more useful when explaining things later.

tbl <-
  dplyr::tibble(
    a = c(5, 2, 6),
    b = c(3, 4, 6),
    c = c(9, 8, 7)
  )
  
tbl

## # A tibble: 3 × 3
##       a     b     c
##   <dbl> <dbl> <dbl>
## 1     5     3     9
## 2     2     4     8
## 3     6     6     7

A: Using an `agent` with validation functions and then `interrogate()`

Validate a number of things on a row-by-row basis using validation functions of the col_vals* type (all have the same number of test units): (1) values in a are less than 8, (2) values in c are greater than the adjacent values in a, and (3) there aren't any NA values in b. We'll determine if this validation has any failing test units (there are 3 test units, one for each row).

agent <-
  create_agent(tbl = tbl) %>%
  conjointly(
    ~ col_vals_lt(., columns = vars(a), value = 8),
    ~ col_vals_gt(., columns = vars(c), value = vars(a)),
    ~ col_vals_not_null(., columns = vars(b))
    ) %>%
  interrogate()

Printing the agent in the console shows the validation report in the Viewer. Here is an excerpt of validation report, showing the single entry that corresponds to the validation step demonstrated here.

What's going on? Think of there being three parallel validations, each producing a column of TRUE or FALSE values (pass or fail) and line them up side-by-side, any rows with any FALSE values results in a conjoint fail test unit.

B: Using the validation function directly on the data (no `agent`)

This way of using validation functions acts as a data filter. Data is passed through but should stop() if there is a single test unit failing. The behavior of side effects can be customized with the actions option.

tbl %>%
  conjointly(
    ~ col_vals_lt(., columns = vars(a), value = 8),
    ~ col_vals_gt(., columns = vars(c), value = vars(a)),
    ~ col_vals_not_null(., columns = vars(b))
  )

## # A tibble: 3 × 3
##       a     b     c
##   <dbl> <dbl> <dbl>
## 1     5     3     9
## 2     2     4     8
## 3     6     6     7

C: Using the expectation function

With the expect_*() form, we would typically perform one validation at a time. This is primarily used in testthat tests.

expect_conjointly(
  tbl,
  ~ col_vals_lt(., columns = vars(a), value = 8),
  ~ col_vals_gt(., columns = vars(c), value = vars(a)),
  ~ col_vals_not_null(., columns = vars(b))
)

D: Using the test function

With the test_*() form, we should get a single logical value returned to us.

tbl %>%
  test_conjointly(
    ~ col_vals_lt(., columns = vars(a), value = 8),
    ~ col_vals_gt(., columns = vars(c), value = vars(a)),
    ~ col_vals_not_null(., columns = vars(b))
  )

## [1] TRUE

Function ID

2-34