pkgnet is an R package designed for the analysis of R packages! The goal of the package is to build graph representations of a package’s various types of dependencies. This can inform a variety of activities, including:

prioritizing functions to unit test based on their centrality or influence
examining the recursive dependencies you are taking on by using a given package
exploring the structure of a new package provided by a coworker or downloaded from the internet

Below is a brief tour of pkgnet and its features.

Packages as a Graph

pkgnet represents aspects of R packages as graphs. The two default reporters, which we will discuss in this vignette, model their respective aspects as directed graphs: a package’s dependencies on other packages, and the interdependencies of functions within a package. Before we look at the output of pkgnet, here are few core concepts to keep in mind.

Example Dependency Graph

Dependency

Units of the analysis are represented as nodes, and their dependency relationships are represented as edges (a.k.a. arcs or arrows). In pkgnet, the nodes could be functions in the package you are examining, or other packages that the package depends on. The direction of edges point in the direction of dependency—the tail node depends on the head node.¹

In the example dependency graph above:

C depends on both A and B.
D depends both C and B.
D indirectly depends upon A through C via the transitive property.
B does not depend on A as there is no path from A to B by following the edges.

Following the direction of the edges allows you to figure out the dependencies of a node—the other nodes that it depends on. On the flip side, tracing the edges backwards allows you to figure out the reverse dependencies (i.e., dependents) of a node—the other nodes that depend on it.

Running pkgnet

pkgnet can analyze any R package locally installed. (Run installed.packages() to see the full list of packages installed on your system.) For this example, let’s say we are analyzing a custom built package, baseballstats.

To analyze baseballstats, run the following two lines of code:

library(pkgnet)
report1 <- CreatePackageReport(pkg_name = "baseballstats")

That’s it! You have generated a lot of valuable information with that one call for an installed package.

However, if the full source repository for the package is available on your system, you can supplement this report with other information such as code coverage from covr. To do so, specify the path to the repository in CreatePackageReport.

library(pkgnet)
report2 <- CreatePackageReport(
  pkg_name = "baseballstats"
  , pkg_path = <path to the repo>
)

Examining the Results

CreatePackageReport has written an HTML report with the pertinent information, and it also returned a list object with the same information and more.

The Report

The location of the HTML report is specified in the messages in the terminal.

This report has three sections:

Package Summary – general information about the package and package level statistics
Dependency Network – information regarding the packages upon which the current package under analysis depends upon
Function Network – information regarding the functions within the current package under analysis and their interdependencies

Each section has helpful tables and visuals.

As a sample, here’s how the Function Network Visualization looks for baseballstats:

Default

All functions and their dependencies are visible. For example, we can see that both batting_avg and slugging_avg functions depend upon the at_bats function.

We also see that nothing depends on the on_base_pct function. This might be valuable information to an R package developer.

With Coverage Information

Same as the default visualization except we can see coverage information as well (Pink = 0%, Green = 100%).

It appears the function with the most dependencies, at_bats, is well covered. However, no other functions are covered by unit tests.

Check out the full HTML report for more results

The List Object

The CreatePackageReport() function returns a list with three items:

SummaryReporter
DependencyReporter
FunctionReporter

Each items contains information visible in the report and more. We can use this information for a more detailed analysis of the results and/or more easily incorporate pkgnet results into other R processes.

Here are a few notable items available within the list object:

Node Information

Both the DependencyReporter and the FunctionReporter contain metrics about their package dependencies or functions (a.k.a network nodes) in a nodes table.

dim(report2$FunctionReporter$nodes)

#> [1]  5 14

names(report2$FunctionReporter$nodes)

#>  [1] "node"                "type"                "isExported"         
#>  [4] "coveredLines"        "totalLines"          "coverageRatio"      
#>  [7] "meanCoveragePerLine" "filename"            "betweenness"        
#> [10] "outDegree"           "inDegree"            "numRecursiveDeps"   
#> [13] "numRecursiveRevDeps" "pageRank"

Note, a few of these metrics provided by default are from the field of Network Theory. You can leverage the Network Object described below to derive many more.

Network Measures

Both the DependencyReporter and the FunctionReporter contain graph-level measures based on their network structure in a network_measures list.

report2$FunctionReporter$network_measures

#> $packageTestCoverage.mean
#> [1] 0.1
#> 
#> $packageTestCoverage.betweenessWeightedMean
#> [1] 0
#> 
#> $graphOutDegree
#> [1] 0.3
#> 
#> $graphInDegree
#> [1] 0.3
#> 
#> $graphBetweenness
#> [1] 0.03125

Network Graph Model Object

Both the DependencyReporter and the FunctionReporter have an object called pkg_graph that contains the graph model of their respective networks. This object has methods to calculate additional node-level and graph-level measures. It is powered by igraph, and the igraph object itself is directly accessible with pkg_graph$igraph.

report2$FunctionReporter$pkg_graph$node_measures(c('hubScore', 'authorityScore'))

#>            node hubScore authorityScore
#> 1:          OPS        1              0
#> 2:      at_bats        0              1
#> 3:  batting_avg        1              1
#> 4:  on_base_pct        0              0
#> 5: slugging_avg        1              1

report2$FunctionReporter$pkg_graph$igraph

#> IGRAPH f3f0cd3 DN-- 5 4 -- 
#> + attr: name (v/c)
#> + edges from f3f0cd3 (vertex names):
#> [1] slugging_avg->at_bats      batting_avg ->at_bats     
#> [3] OPS         ->slugging_avg OPS         ->batting_avg

A Deeper Look

With the reports and objects produced by pkgnet by default, there is plenty to inform us on the inner workings of an R package. However, we may want to know MORE! Since the igraph objects are available, we can leverage those graphs for further analysis.

In this section, let’s examine a larger R package, such as lubridate.

If you would like to follow along with the examples in this section, run these commands in your terminal to download and install lubridate².

# Create a temporary workspace
mkdir -p ~/pkgnet_example && cd ~/pkgnet_example

# Grab the lubridate source code
git clone https://github.com/tidyverse/lubridate
cd lubridate

# If you want the examples to match exactly
git reset --hard 9797d69abe1574dd89310c834e52d358137669b8

# Install it
R CMD install .

Coverage of Most Depended-on Functions

Let’s examine lubridate’s functions through the lens of each function’s total number of dependents (i.e., the other functions that depend on it) and its code’s unit test coverage. In our graph model for the FunctionReporter, the subgraph of paths leading into a given node is the set of functions that directly or indirectly depend on the function that node represents.

# Run pkgnet
library(pkgnet)
report2 <- CreatePackageReport(
    pkg_name = "lubridate"
    , pkg_path = "~/pkgnet_example/lubridate"
)

# Extract Nodes Table
funcNodes <- report2$FunctionReporter$nodes

# List Coverage For Most Depended-on Functions
mostRef <- funcNodes[order(numRecursiveRevDeps, decreasing = TRUE),
                     .(node, numRecursiveRevDeps, coverageRatio, totalLines)
                     ][1:10]

#>             node numRecursiveRevDeps coverageRatio totalLines
#>  1:        month                  81             1          1
#>  2:           tz                  79             1          1
#>  3: reclass_date                  68             1          1
#>  4:         date                  67             1          1
#>  5:      is.Date                  60             1          1
#>  6:    is.POSIXt                  57             1          1
#>  7:         wday                  56             1          1
#>  8:   is.POSIXct                  55             1          1
#>  9:  .deprecated                  55             0         10
#> 10:      as_date                  52             1          1

Inspecting results such as these can help an R package developer decide which function to cover with unit tests next.

In this case, check_duration, one of the most depended-on functions (either directly or indirectly), is not covered by unit tests. However, it appears to be a simple one line function that may not be necessary to cover in unit testing. check_interval, on the other hand, might benefit from some unit test coverage as it is a larger, uncovered function with a similar number of dependencies.

Discovering Similar Functions

Looking at that same large package, let’s say we want to explore options for consolidating functions. One approach might be to explore consolidating functions that share the same dependencies. In that case, we could use the igraph object to highlight functions with the same out-neighborhood via Jaccard similarity.

# Get igraph object
funcGraph <- report2$FunctionReporter$pkg_graph$igraph
funcNames <- igraph::vertex_attr(funcGraph, name = "name")

# Jaccard Similarity
sim <- igraph::similarity(graph = funcGraph
                          , mode = "out"
                          , method = "jaccard")
diag(sim) <- 0
sim[sim < 1] <- 0

simGraph <- igraph::graph_from_adjacency_matrix(adjmatrix = sim, mode = "undirected")

# Find groups with same out-neighbors (similarity == 1)
sameDeps <- igraph::max_cliques(graph = simGraph
                                , min = 2
                                )

# Write results
for (i in seq_along(sameDeps)) {
    cat(paste0("Group ", i, ": "))
    cat(paste(funcNames[as.numeric(sameDeps[[i]])], collapse = ", "))
    cat("\n")
}

#> Group 1: divisible_period, make_date
#> Group 2: parse_date_time2, fast_strptime
#> Group 3: .deprecated_fun, .deprecated_arg
#> Group 4: stamp_date, stamp_time
#> Group 5: epiweek, isoweek
#> Group 6: ms, hm
#> Group 7: quarter, semester
#> Group 8: am, .roll_hms
#> Group 9: modulo_interval_by_duration, modulo_interval_by_period
#> Group 10: .difftime_from_pieces, .duration_from_units
#> Group 11: divide_period_by_period, xtfrm.Period
#> Group 12: int_diff, %--%
#> Group 13: isoyear, epiyear
#> Group 14: nanoseconds, microseconds, picoseconds, milliseconds
#> Group 15: period_to_seconds, check_period, multiply_period_by_number, format.Period, divide_period_by_number, add_period_to_period
#> Group 16: myd, dmy, yq, ymd, dym, mdy, ydm
#> Group 17: hours, weeks, minutes, years, days, months.numeric, seconds, seconds_to_period
#> Group 18: C_force_tz, hour.default, mday.default, c.POSIXct, .mklt, yday.default, year.default, minute.default, second.default
#> Group 19: ehours, emilliseconds, eyears, eseconds, epicoseconds, enanoseconds, eminutes, olson_time_zones, edays, emicroseconds, eweeks
#> Group 20: dmy_h, ydm_hms, ymd_hms, dmy_hm, ymd_h, ydm_hm, ydm_h, dmy_hms, ymd_hm, mdy_hms, mdy_hm, mdy_h

Now, we have identified twenty different groups of functions within lubridate that share the exact same dependencies. We could explore each group of functions for potential consolidation.

Edge direction was previously Independent -> Dependent. It was changed to Dependent -> Independent in version v0.3.0. The new convention follows the Unified Modeling Language (UML) framework, a widely used standard for software system modeling.↩︎
Examples from version 1.7.3 of Lubridate↩︎

Exploring the Structure and Dependencies of an R Package