S4 methods not registered for packages loaded in test file

[kevinushey]

(I apologize in advance...)

Normally, loading the Matrix package forces a generic for t to be registered; e.g.

library(Matrix)
inherits(t, "standardGeneric")  # TRUE

However, this doesn't seem to happen if the package is loaded in a test file. For example:

library(Matrix)
test_that("t is a standardGeneric", {
  expect_s4_class(t, "standardGeneric")
})

If I run testthat::test_local(), I see:

> testthat::test_local()
✔ | F W S  OK | Context
✖ | 1       0 | hello
─────────────────────────────────────────────────────────
Failure (test-hello.R:4:3): t is a standardGeneric
`t` is not an S4 object
─────────────────────────────────────────────────────────

══ Results ══════════════════════════════════════════════
[ FAIL 1 | WARN 0 | SKIP 0 | PASS 0 ]
Error: Test failures

And yet, this works fine with test_file():

> testthat::test_file("tests/testthat/test-hello.R")

══ Testing test-hello.R ═════════════════════════════════
[ FAIL 0 | WARN 0 | SKIP 0 | PASS 1 ] Done!

Any idea what might be going on?

[hadley]

No 😢

[lionel-]

I think this is because test_local() uses the load_all() strategy and then evaluates in the namespace environment (technically, a clone of the namespace), whereas test_file() doesn't load the package and evaluates in the global environment.

If you don't explicitly import the t() generic from the Matrix package in your namespace, then the non-generic definition of base::t() from the base imports of your package gets picked up.

A related and even more subtle problem is r-lib/pkgload#88. In that issue the package calls setMethod("t"), which causes the creation of a "t" generic but only if one doesn't already exists. So when devtools::test(), which calls devtools::load_all(), is run for the first time, the generic is created in the package's namespace. Then the tests call library(Matrix) which creates another "t" generic on the search path. When devtools::test() is called a second time, setMethod() finds Matrix's "t" generic and doesn't create a local generic this time around. And then the base non-generic "t" is found in the imports.

I think this suggests two recommendations:

• Do not call library(Matrix) in tests.
• Explicitly import all the generics for which you define methods in your namespace. The S4 documentation recommends an explicit call to setGeneric() but I think it's better to import the generic from another authoritative package if one exists, this way the methods can be shared.

[hadley]

Closing this because I think @lionel- proposed a solution, and few people seem to have stumbled into the same problem.