Tests asserting correctness of nbrOfWorkers() could produce an
"invalid format '%d'; use format %f, %e, %g or %a for numeric
objects" error when trying to produce an assertion error on
nbrOfWorkers() having an incorrect value. This could happen if
nbrOfWorkers() returned a non-integer value, e.g. +Inf.
Test asserting that the workers argument can be a function would
not always work if testing with a hardcoded number of workers
according to plan().
Test asserting that the workers argument can be a function would
not work if the backend's default value was non-numeric, e.g. the
cluster backend defaults to the character vector
parallelly::availableWorkers().
Test asserting that lazy futures would be automatically launched
and resolved relied on a legacy version of the Future API, where
calling resolved() on a lazy future could leave it in a lazy
state, which is no longer correct. A lazy future will always be
launched if one calls resolved() on it.
Test asserting that the ff package worked across multiple futures assumed that the package is loaded automatically by the future, which it is not. The could cause the test to fail for some future backends.
Now tests can be formally skipped by calling
future.tests::skip_test() from within the test. Skipped tests
are counted and reported in the summary.
Now check_plan() outputs the reason for a test is being skipped.
Now check_plan() outputs also the error message, error class, the
call, and any standard output, whenever there's is a test error.
Now check_plan() outputs also the test iteration index.
Add Rscript -e future.tests::check --version.
workers
argument, which is not true for all future backends. For example,
'sequential' does not take argument workers. Previously, we
avoided this problem by only testing if the evaluator inherited
multiprocess, but that is not sufficient, e.g. upcoming
future.redis::redis inherits multiprocess, but still does not
have a workers argument. Now we check for the workers argument
instead.future(), run(), result() and value()
update the RNG state.Assert that rm(a) in a future expression only removes a local
variable a, but never a global variable a.
Assert that packages data.table and ff are not affected when a future resets the R options on the worker when resolved.
Assert that a global that is a copy of a non-exported package
object (e.g. utils:::str2logical()) is not dropped because it
belongs to a package namespace.
Assert that ... can be exported as a global to a future, and
used as-is inside a function that does not have ... arguments.
check() and check_plan() gained argument local, which is
passed down to run_test().
check() gained argument envir, which is passed down to
run_test().
The TestResult class did not record whether the test was evaluated in a local environment or not.
A too strict internal assertion would give Error in evaluate_expr(test$expr, envir = envir, local = FALSE, output = output, : identical(Sys.getenv(), old$envvars) is not TRUE for R
4.2.x and R-devel on MS Windows. This was because it is not
possible to remove environment variables on MS Windows; they can
only be set to an empty value.
Assert that future() doesn't change the RNG kind.
Assert that future(..., conditions = character(0L)) muffles all
conditions.
For robustness, using explicit stringsAsFactors = FALSE
internally.
evaluate_expr(), which is used for running all tests, now reset
options, environment variables, the RNG kind, and the random seed
afterward to what it was before being called.
Added a package vignettes.
resolve() would use deprecated argument value.Update the package description to use single quotes.
Add example to check().
resolved() will launch lazy futures.In non-interactive mode, check() will quit R with an exit code
that reflects whether all tests passed (0) or not (1).
check() gained arguments so that it can be easily called from R
too.