terraform

Commit Graph

Author	SHA1	Message	Date
Martin Atkins	0f845adfdb	moduletest: EvalContext lives outside of the "terraform" package The initial implementation of the module testing language included some elements deeply embedded inside "package terraform", which is primarily concerned with the traditional Terraform language, which we might now call the "module language" to distinguish it from the test language and the stacks language. That was a reasonable design decision due to the modules runtime needing to evaluate some expressions in a similar way to how the modules language does it, but unfortunately it caused the test language runtime to depend a little too closely on some current implementation details of how the modules runtime deals with expression evaluation, and those details are about to change as we teach Terraform to be able to deal with unknown values in new places, as part of resolving this issue: https://github.com/hashicorp/terraform/issues/30937 This commit therefore refactors the test runtime a little so that it doesn't intrude so deeply into the module runtime's business. Instead of directly extending the module runtime's expression evaluator, we'll instead call into it in a similar way to how the "terraform console" command does, by wrapping the evaluation data produced by the module runtime to extend it with the only two special things that the test runtime needs: references to the output values of previous runs, and references to test-only variables that aren't declared as part of the module under test. This effectively means that the test runtime can now evaluate a superset of the expressions that would be valid in the global scope of the module under test using only the public API of the modules runtime, with no testing-specific evaluation logic embedded inside the modules runtime. This new strategy is also somewhat similar to how the stacks runtime calls into the modules runtime to evaluate components, but the stacks runtime only actually cares about output values and so doesn't need a full evaluation scope. This means the result here is a little more symmetrical with the approach we took in stacks, which will hopefully allow experience working on one to translate better to the other.	2 years ago

Author

SHA1

Message

Date

Martin Atkins

0f845adfdb

moduletest: EvalContext lives outside of the "terraform" package

The initial implementation of the module testing language included some
elements deeply embedded inside "package terraform", which is primarily
concerned with the traditional Terraform language, which we might now call
the "module language" to distinguish it from the test language and the
stacks language.

That was a reasonable design decision due to the modules runtime needing
to evaluate some expressions in a similar way to how the modules language
does it, but unfortunately it caused the test language runtime to depend
a little too closely on some current implementation details of how the
modules runtime deals with expression evaluation, and those details are
about to change as we teach Terraform to be able to deal with unknown
values in new places, as part of resolving this issue:
    https://github.com/hashicorp/terraform/issues/30937

This commit therefore refactors the test runtime a little so that it
doesn't intrude so deeply into the module runtime's business. Instead of
directly extending the module runtime's expression evaluator, we'll
instead call into it in a similar way to how the "terraform console"
command does, by wrapping the evaluation data produced by the module
runtime to extend it with the only two special things that the test
runtime needs: references to the output values of previous runs, and
references to test-only variables that aren't declared as part of the
module under test.

This effectively means that the test runtime can now evaluate a superset
of the expressions that would be valid in the global scope of the module
under test using only the public API of the modules runtime, with no
testing-specific evaluation logic embedded inside the modules runtime.

This new strategy is also somewhat similar to how the stacks runtime calls
into the modules runtime to evaluate components, but the stacks runtime
only actually cares about output values and so doesn't need a full
evaluation scope. This means the result here is a little more symmetrical
with the approach we took in stacks, which will hopefully allow experience
working on one to translate better to the other.

1 Commits (5848d8a2b4e0c500c384bcbdcc84b5feec3f6549)