Traditionally we've preferred to release new language features in major
releases only, because we can then use the beta cycle to gather feedback
on the feature and learn about any usability challenges or other
situations we didn't consider during our design in time to make those
changes before inclusion in a stable release.
This "experiments" feature is intended to decouple the feedback cycle for
new features from the major release rhythm, and thus allow us to release
new features in minor releases by first releasing them as experimental for
a minor release or two, adjust for any feedback gathered during that
period, and then finally remove the experiment gate and enable the feature
for everyone.
The intended model here is that anything behind an experiment gate is
subject to breaking changes even in patch releases, and so any module
using these experimental features will be broken by a future Terraform
upgrade.
The behavior implemented here is:
- Recognize a new "experiments" setting in the "terraform" block which
allows module authors to explicitly opt in to experimental features.
terraform {
experiments = [resource_for_each]
}
- Generate a warning whenever loading a module that has experiments
enabled, to avoid accidentally depending on experimental features and
thus risking unexpected breakage on next Terraform upgrade.
- We check the enabled experiments against the configuration at module
load time, which means that experiments are scoped to a particular
module. Enabling an experiment in one module does not automatically
enable it in any other module.
This experiments mechanism is itself an experiment, and so I'd like to
use the resource for_each feature to trial it. Because any configuration
using experiments is subject to breaking changes, we are free to adjust
this experiments feature in future releases as we see fit, but once
for_each is shipped without an experiment gate we'll be blocked from
making significant changes to it until the next major release at least.
* huge change to weave new addrs.Provider into addrs.ProviderConfig
* terraform: do not include an empty string in the returned Providers /
Provisioners
- Fixed a minor bug where results included an extra empty string
* terraform/context: use new addrs.Provider as map key in provider factories
* added NewLegacyProviderType and LegacyString funcs to make it explicit that these are temporary placeholders
This PR introduces a new concept, provider fully-qualified name (FQN), encapsulated by the `addrs.Provider` struct.
Add deprecation warning for references from destroy provisioners or
their connections to external resources or values. In order to ensure
resource destruction can be completed correctly, destroy nodes must be
able to evaluate with only their instance state.
We have sufficient information to validate destroy-time provisioners
early on during the config loading process. Later on these can be
converted to hard errors, and only allow self, count.index, and each.key
in destroy provisioners. Limited the provisioner and block evaluation
scope later on is tricky, but if the references can never be loaded,
then they will never be encountered during evaluation.
`terraform 0.12upgrade` assumes that the configuration has passed 0.11
init, but did not explicitly check that the configuration was valid.
Certain issues would not get caught because the configuration was
syntactically valid. In this case, int or float values out of range
resulted in a panic from `Value()`.
Since running a 0.11 validate command is a breaking change, this PR
merely moves the `Value()` logic for ints and floats into `configupgrade` so
the error can be returned to the user, instead of causing a panic.
Following on from de652e22a26b, this introduces deprecation warnings for
when an attribute value expression is a template with only a single
interpolation sequence, and for variable type constraints given in quotes.
As with the previous commit, we allowed these deprecated forms with no
warning for a few releases after v0.12.0 to ensure that folks who need to
write cross-compatible modules for a while during upgrading would be able
to do so, but we're now marking these as explicitly deprecated to guide
users towards the new idiomatic forms.
The "terraform 0.12upgrade" tool would've already updated configurations
to not hit these warnings for those who had pre-existing configurations
written for Terraform 0.11.
The main target audience for these warnings are newcomers to Terraform who
are learning from existing examples already published in various spots on
the wider internet that may be showing older Terraform syntax, since those
folks will not be running their configurations through the upgrade tool.
These warnings will hopefully guide them towards modern Terraform usage
during their initial experimentation, and thus reduce the chances of
inadvertently adopting the less-readable legacy usage patterns in
greenfield projects.
Terraform 0.12.0 removed the need for putting references and keywords
in quotes, but we disabled the deprecation warnings for the initial
release in order to avoid creating noise for folks who were intentionally
attempting to maintain modules that were cross-compatible with both
Terraform 0.11 and Terraform 0.12.
However, with Terraform 0.12 now more widely used, the lack of these
warnings seems to be causing newcomers to copy the quoted versions from
existing examples on the internet, which is perpetuating the old and
confusing quoted form in newer configurations.
In preparation for phasing out these deprecated forms altogether in a
future major release, and for the shorter-term benefit of giving better
feedback to newcomers when they are learning from outdated examples, we'll
now re-enable those deprecation warnings, and be explicit that the old
forms are intended for removal in a future release.
In order to properly test this, we establish a new set of test
configurations that explicitly mark which warnings they are expecting and
verify that they do indeed produce those expected warnings. We also
verify that the "success" tests do _not_ produce warnings, while removing
the ones that were previously written to succeed but have their warnings
ignored.
During the 0.12 work we intended to move all of the variable value
collection logic into the UI layer (command package and backend packages)
and present them all together as a unified data structure to Terraform
Core. However, we didn't quite succeed because the interactive prompts
for unset required variables were still being handled _after_ calling
into Terraform Core.
Here we complete that earlier work by moving the interactive prompts for
variables out into the UI layer too, thus allowing us to handle final
validation of the variables all together in one place and do so in the UI
layer where we have the most context still available about where all of
these values are coming from.
This allows us to fix a problem where previously disabling input with
-input=false on the command line could cause Terraform Core to receive an
incomplete set of variable values, and fail with a bad error message.
As a consequence of this refactoring, the scope of terraform.Context.Input
is now reduced to only gathering provider configuration arguments. Ideally
that too would move into the UI layer somehow in a future commit, but
that's a problem for another day.
Previously we were using the experimental HCL 2 repository, but now we'll
shift over to the v2 import path within the main HCL repository as part of
actually releasing HCL 2.0 as stable.
This is a mechanical search/replace to the new import paths. It also
switches to the v2.0.0 release of HCL, which includes some new code that
Terraform didn't previously have but should not change any behavior that
matters for Terraform's purposes.
For the moment the experimental HCL2 repository is still an indirect
dependency via terraform-config-inspect, so it remains in our go.sum and
vendor directories for the moment. Because terraform-config-inspect uses
a much smaller subset of the HCL2 functionality, this does still manage
to prune the vendor directory a little. A subsequent release of
terraform-config-inspect should allow us to completely remove that old
repository in a future commit.
copyDir is used in configload/getter.go to copy previously downloaded modules instead of using the go-getter client every time. The go-getter client downloads dotfiles, but copyDir did not copy dotfiles, leading to inconsistent behaviour when reusing the same module source.
In order to allow lazy evaluation of resource indexes, we can't index
resources immediately via GetResourceInstance. Change the evaluation to
always return whole Resources via GetResource, and index individual
instances during expression evaluation.
This will allow us to always check for invalid index errors rather than
returning an unknown value and ignoring it during apply.
We can only validate MinItems >= 1 (equiv to "Required") during
decoding, as dynamic blocks each only decode as a single block. MaxItems
cannot be validated at all, also because of dynamic blocks, which may
have any number of blocks in the config.
Due to both the nature of dynamic blocks, and the need for resources to
sometimes communicate incomplete values, we cannot validate MinItems and
MaxItems in CoerceValue.
A provider may not have the data to fill in required block values in all
cases during the resource Read operation. This is more common in import,
because there is no initial configuration or state, and it's possible
some values are only provided in the configuration.
The original intent of MinItems and MaxItems in the schema was to
enforce configuration constraints, not to enforce what the resource
could save in the state. Since the configuration is already statically
validated, and the Schema is validated against the configuration in a
separate step, we can drop these extra validation constraints in
CoerceValue and relax it to only ensure the types conform to what is
expected.
If a block was defined via "dynamic", there will be only one block value
until the expansion is known. Since we can't detect dynamic blocks at
this point, don't verify MinItems while there are unknown values in the
config.
The decoder spec can also only check for existence of a block, so limit
the check to 0 or 1.
This also fixes a few things with resource for_each:
It makes validation more like validation for count.
It makes sure the index is stored in the state properly.
In some cases (see #22020 for a specific example), the parsed hilNode
can be nil. This causes a series of panics. Instead, return an error and
move on.
When loading nested modules, the child module diagnostics were dropped
in the recursive function. This mean that the config from the submodules
wasn't fully loaded, even though no errors were reported to the user.
This caused further problems if the plan was stored in a plan file, when
means only the partial configuration was stored for the subsequent apply
operation, which would result in unexplained "Resource node has no
configuration attached" errors later on.
Also due to the child module diagnostics being lost, any newly added
nested modules would be silently ignored until `init` was run again
manually.
Previously, adding a version constraint to a module that was previously
recorded without a version in the module manifest would cause a panic.
Instead, we now use a slight variant of the "dependencies have changed"
error that doesn't try to print out a specific version number.
Previously we were trying to access a field of the analysis object before
checking if analysis produced errors. The analysis function usually
returns a nil analysis on error, so this would result in a panic whenever
that happened.
Now we'll dereference the analysis object pointer only after checking for
errors, so we'll get a chance to report the analysis error to the user.
The expression upgrade functionality mostly ignores comments because in
the old language the syntax prevented comments from appearing in the
middle of expressions, but there was one exception: object expressions.
Because HCL 1 used ObjectType both for blocks and for object expressions,
that is the one situation where something we consider to be an expression
could have inline attached comments in the old language.
We migrate these here so we don't lose these comments that don't appear
anywhere else. Other comments get gathered up into a general comments
set maintained inside the analysis object and so will be printed out as
required _between_ expressions, just as they did before.
Previously we were treating "dynamic" blocks in configuration the same as
any other block type when merging config bodies, so that dynamic blocks
in the override would override any dynamic blocks present in the base,
without considering the dynamic block type.
It's more useful and intuitive for us to treat dynamic blocks as if they
are instances of their given block type for the purpose of overriding.
That means a foo block can be overridden by a dynamic "foo" block and
vice-versa, and dynamic blocks of different types do not interact at all
during overriding.
This requires us to recognize dynamic blocks and treat them specially
during decoding of a merged body. We leave them unexpanded here because
this package is not responsible for dynamic block expansion (that happens
in the sibling "lang" package) but we do decode them enough to recognize
their labels so we can treat them as if they were blocks of the labelled
type.
