Terraform v1.x Compatibility Promises
The release of Terraform v1.0 represents an important milestone in the development of the Terraform language and workflow. Terraform v1.0 is a stable platform for describing and managing infrastructure.
In this release we're defining a number of Terraform behaviors that we intend to remain compatible with throughout the 1.x releases:
- A large subset of Terraform language features.
- A more conservative subset of the Terraform CLI workflow commands.
- The wire protocol for communication between Terraform Core and Terraform providers.
- The wire protocols for installation of Terraform providers and external Terraform modules.
Our intention is that Terraform modules written for Terraform v1.0 will continue to plan and apply successfully, without required changes, throughout the v1.x releases.
We also intend that automation built around the workflow subset described in this document will work without changes in all future v1.x releases.
Finally, we intend that providers built against the currently-documented provider wire protocol will be compatible with all future Terraform v1.x releases targeting the same operating system and architecture, without the need for source code modification or binary recompilation.
In short, we aim to make upgrades between v1.x releases straightforward, requiring no changes to your configuration, no extra commands to run upgrade steps, and no changes to any automation you've set up around Terraform.
The Terraform v1.x series will be actively maintained for at least 18 months after v1.0.
The following sections include some specific guidance on what we will promise throughout the v1.x series, for those who would like full details. At a higher level though, we don't intend to make any changes that would cause existing modules or automation to require changes when upgrading to a new v1.x release. We will generally treat compatibility problems in new Terraform CLI releases as bugs to be fixed unless there was a very significant justification for the change, such as in addressing a critical security problem or matching with a breaking change to a remote dependency that isn't directly under our control.
The Terraform Language
The main Terraform Language includes the language syntax, the top-level
structures such as
provider blocks, the
"meta-arguments" in those blocks, and the documented semantics and behaviors
for the operators and built-in functions available for use in expressions.
There is not a single formal specification for the Terraform language, but the Configuration section of the documentation on the Terraform website serves as a description of the language features and their intended behaviors.
The following top-level blocks and their defined "meta-arguments" (that is, arguments defined by Terraform Core rather than by external plugins such as providers) will retain their current functionality:
datablocks to declare resources, including their nested block types
provisioner, and their meta-argument
moduleblocks to call other modules, and its meta-argument
providerblocks to configure providers, and the
localsblocks for declaring the various kinds of named values in a module.
terraformblocks, including the nested
required_providersarguments, and nested
backendblocks for backend configuration.
We also intend to keep compatibility with all
expression operators and
built-in functions, with the exception of
whose behavior may change as part of future changes to the workspace model.
We intend to retain broad compatibility with Terraform language features, with a few specific caveats:
We consider a configuration to be valid if Terraform can create and apply a plan for it without reporting any errors.
A configuration that currently produces errors might generate different errors or exhibit other non-error behaviors in a future version of Terraform. A configuration that generates errors during the apply phase might generate similar errors at an earlier phase in future, because we generally consider it better to detect errors in as early a phase as possible.
Generally-speaking, the compatibility promises described in this document apply only to valid configurations. Handling of invalid configurations is always subject to change in future Terraform releases.
If the actual behavior of a feature differs from what we explicitly documented as the feature's behavior, we will usually treat that as a bug and change the feature to match the documentation, although we will avoid making such changes if they seem likely to cause broad compatibility problems. We cannot promise to always remain "bug-compatible" with previous releases, but we will consider such fixes carefully to minimize their impact.
Any experimental features may change or may be removed entirely from future releases. Terraform always produces a warning when an experimental language feature is active, to make you aware of that risk. We don't recommend using experimental features in production modules.
We will introduce new language features, and if you start using them then your configuration won't work with prior Terraform versions that didn't support those features yet.
Terraform Providers are separate plugins which can change independently of Terraform Core and are therefore not subject to these compatibility promises. If you upgrade any of the providers you are using then you might need to change provider or resource configurations related to those providers.
A small number of features remain deprecated with explicit warnings in Terraform v1.0. Those deprecation cycles will end in a future v1.x release, at which point we will remove the corresponding features.
There is a set of often used Terraform workflows, which we are calling protected workflows. We will not remove these commands, subcommands, and flags or make backward-incompatible changes to protected workflow functionality. If we accidentally change these, we will consider backwards-incompatible changes to core workflows as bugs to be fixed. For a list of the command and option combinations that are part of protected workflows, see Protected Workflow Commands. There is another set of commands that we are explicitly not making compatibility promises about, because we expect their functionality to change in v1.x releases: see Commands That Might Change.
The supported ways for external software to interact with Terraform are via the JSON output modes offered by some commands and via exit status codes. We may extend certain JSON formats with new object properties but we will not remove or make breaking changes to the definitions of existing properties.
Natural language command output or log output is not a stable interface and may change in any new version. If you write software that parses this output then it may need to be updated when you upgrade Terraform. If you need access to data that is not currently available via one of the machine-readable JSON interfaces, we suggest opening a feature request to discuss your use-case.
Upgrading and Downgrading
Throughout the v1.x series of releases, we intend that you should be able to switch to a newer Terraform version and use it just as before, without any special upgrade steps.
You should be able to upgrade from any v1.x release to any later v1.x release. You might also be able to downgrade to an earlier v1.x release, but that isn't guaranteed: later releases may introduce new features that earlier versions cannot understand, including new storage formats for Terraform state snapshots.
If you make use of features introduced in a later v1.x release, your configuration won't be compatible with releases that predate that feature. For example, if a language feature is added in v1.3 and you start using it, your Terraform configuration will no longer be compatible with Terraform v1.2.
Terraform providers are separate plugins which communicate with Terraform using a documented protocol. Therefore these compatibility promises can only cover the "client" side of this protocol as implemented by Terraform Core; the behaviors of individual providers, including which resource types they support and which arguments they expect, are decided by the provider development teams and can change independently of Terraform Core releases.
If you upgrade to a new version of a provider then you might need to change the parts of your configuration which are interpreted by that provider, even if you are still using a Terraform v1.x release.
Provider Installation Methods
Terraform normally installs providers from a provider registry implementing the Provider Registry Protocol, version 1. All Terraform v1.x releases will remain compatible with that protocol, and so correctly-implemented provider registries will stay compatible.
Terraform also supports installation of providers from local filesystem directories (filesystem mirrors) and from network mirrors (implementing the Provider Mirror Protocol. All Terraform v1.x releases will remain compatible with those installation methods, including the Implied Local Mirror Directories.
Specific provider registries or network mirrors are run independently from Terraform itself and so their own behaviors are not subject to these compatibility promises.
Provider Protocol Versions
The current major version of the provider plugin protocol as of Terraform v1.0 is version 5, which is defined by a combination of a Protocol Buffers schema describing the physical wire formats and by additional prose documentation describing the expected provider behaviors.
We will support protocol version 5 throughout the Terraform v1.x releases. If we make new minor revisions to protocol version 5 in later releases then we will design them such that existing plugins will continue to work, as long as they correctly implemented the protocol.
We may introduce new major versions of the protocol during the v1.x series. If so, we will continue to support protocol version 5 alongside those new versions. Individual provider teams might decide to remove support for protocol version 5 in later releases, in which case those new provider releases will not be compatible with all of the Terraform v1.x releases.
Terraform modules are reusable infrastructure components written in the Terraform language. Some modules are "external" in the sense that Terraform automatically installs them from a location other than the current configuration directory, in which case their contents could change independently of changes to your local modules, of the providers you use, and of Terraform itself.
Module Installation Methods
Terraform supports installing child modules from a number of different module source types. We will continue to support all of the existing source types throughout the v1.x releases.
One of the supported source types is a module registry implementing the Module Registry Protocol version 1. All Terraform v1.x releases will remain compatible with correct implementations of that protocol.
Some module source types work directly with services or protocols defined and run by third parties. Although we will not remove Terraform's own client-side support for those, we cannot guarantee that their owners will keep those services running or that they will remain compatible with Terraform's client implementations.
External Module Compatibility
If your configuration depends on external modules, newer versions of those modules may include breaking changes. External modules are not part of Terraform and are therefore not subject to these compatibility promises.
We will maintain compatibility for the
provisioner types through all v1.x releases.
Some additional vendor-specific provisioners were available in earlier Terraform versions but were deprecated in Terraform v0.13 and removed in Terraform v0.15.
Terraform supports loading additional provisioners as plugins from certain local filesystem directories. We'll continue to support that throughout the Terraform v1.x releases, but since such plugins are separate from Terraform Core itself their own behaviors cannot be subject to these compatibility promises. However, we will continue to support the plugin wire protocol as defined in Terraform v1.0 throughout the v1.x releases, and so correctly-implemented provisioner plugins should remain compatible with future Terraform releases.
State Storage Backends
When you use remote state, Terraform interacts with remote services over the network in order to store and manage locks for Terraform state.
For historical reasons, all supported state storage backends are included as part of Terraform CLI but not all are supported directly by the Terraform Team. Only the following backends maintained by the Terraform team are subject to compatibility promises:
The other state storage backends are maintained by external teams via contributions to the Terraform CLI codebase, and so their expected configuration arguments or behaviors might change even in v1.x releases, although we will aim to still ensure a good migration path in such cases, where possible.
We are considering allowing external state storage backend implementations via plugins, similar to provider plugins. If we introduce such a mechanism during the v1.x releases then you may need to make configuration changes in order to use those plugins, and state storage backends other than those listed above may be removed from later versions of Terraform CLI once equivalent plugins are available.
remote Backend and Terraform Cloud
remote backend is maintained by the Terraform Cloud team and so its
behavior may change along with ongoing changes to Terraform Cloud.
There will be a supported mechanism to use Terraform CLI with Terraform Cloud throughout the v1.x releases, but the exact details may change. Terraform Cloud evolves independently of Terraform CLI and is therefore not subject to these compatibility promises.
Community-maintained State Storage Backends
kubernetes backends are maintained by
other teams at HashiCorp. Those teams intend to continue basic maintenence at
the level of bug fixes through the v1.x releases, unless we implement a plugin
protocol for backends at which point development of these backends is likely
to continue in the external plugins only, which may require configuration
changes to switch to the plugin equivalents.
etcdv3 backends are maintained by outside
contributors and are not subject to these compatibility promises.
Unmaintained State Storage Backends
swift state storage backends do not
currently have any maintainers and thus remain in Terraform CLI releases on
a best-effort basis. They may be removed in later v1.x releases, and will not
be updated in case of any breaking changes to the services they integrate with.
Throughout the v1.x series we will continue to produce official releases for the following platforms, and make changes as necessary to support new releases of these operating systems:
- macOS on x64 CPUs (
- Windows on x64 CPUs (
- Linux on x64, 32-bit ARMv6, and 64-bit ARMv8 (
Over time we may require newer versions of these operating systems. For example, subsequent Terraform releases in the v1.x series might end support for earlier versions of macOS or Windows, or earlier Linux kernel releases.
We have historically produced official releases for a number of other platforms as a convenience to users of those platforms, and we have no current plans to stop publishing them but we cannot promise ongoing releases or bug fixes for the other platforms throughout the v1.x series. We do not routinely test Terraform on any platforms other than those listed above.
We might add support for new platforms in later v1.x releases. If so, earlier Terraform releases prior to that support will not be available on those platforms.
All Terraform plugins, including provider plugins, are separate programs that have their own policies for which platforms they support. We cannot guarantee that all providers currently support or will continue to support the platforms listed above, even though Terraform CLI itself will support them.
Later Revisions to These Promises
We may extend or refine these promises throughout the v1.x series in order to describe promises related to new features or to clarify existing promises if we find via feedback that our earlier statements had been unclear.
Promises for new features will be additive in the sense that they will add further promises without retracting any existing ones. For promises that only apply to later v1.x releases we will mention the earliest version(s) those promises apply to.
Even if we don't add an explicit statement to this document, we intend that any non-experimental features added in later v1.x releases will remain compatible at least through the remainder of the v1.x series, unless otherwise stated.
Protected Workflow Commands
The following is the list of Terraform CLI subcommands and options that are subject to these compatibility promises. If you build automation around these commands then it should be compatible with all later v1.x releases.
As noted above, compatibility with external software is limited to
explicitly-machine-readable output (
-raw modes) and exit codes.
Any natural-language output from these commands might change in later releases.
providers(with no subcommand)
For commands or options not in the above list, we will still avoid breaking changes where possible, but can't promise full compatibility throughout the v1.x series. If you are building automation around Terraform, use only the commands above to avoid the need for changes when upgrading.
Please note that although Terraform's internal logs (via the
environment variable) are available in a JSON format, the particular syntax
or structure of those log lines is not a supported integration interface.
The logs are available as JSON only to help with ad-hoc filtering and
processing of logs by Terraform developers.
Commands That Might Change
All of the following commands and their subcommands/options are not subject to compatibility promises, either because we have existing plans to improve them during the v1.x series or because we are aware of shortcomings in their design that might require breaking changes for ongoing maintenence.
While you can freely use these commands when running Terraform interactively as long as they remain supported, we don't recommend using them as part of any automation unless you are willing to potentially update that automation when upgrading to a later v1.x release.
terraform apply -destroyinstead)
terraform apply -refresh-onlyinstead)
- all subcommands of
workspace(and its deprecated alias
While we do intend to retain support for the main use-cases associated with these commands in future releases, we cannot promise to retain the exact command names or options used to meet those use-cases.
How We Will Keep These Promises
Automated Regression Testing
The Terraform codebase includes various unit and integration tests intended to help us to notice accidental behavior regressions before they ship in a stable version.
However, Terraform is a relatively complex system with many different features that can interact in interesting ways. In the past we've seen reports of behavior differences that appeared only when combining two or more features in a way we hadn't previously anticipated or written automated tests for.
In each case we have both implemented a change to resolve the compatibility problem and added one or more integration tests representing the behavior of that combination of features. We intend to continue this approach, so we can improve Terraform's test coverage over time.
We intend that most accidental changes in behavior covered by these promises will be caught by existing tests. However, we also accept that our test suite can never have perfect coverage of all possible feature interactions or other edge cases, and so we aim for each significant change to be included in both alpha and beta releases before eventual inclusion in a final release.
For minor releases we will typically also issue at least one release candidate prior to the final release. A release candidate represents that planned development is concluded and that we've fixed any regressions reported based on the alpha and beta releases, and thus the final release that follows should typically match exactly or almost exactly its most recent release candidate.
Regressions in Final Releases
For more obscure combinations of features it is possible that a regression could be undetected during prerelease the prerelease periods and thus included in a final release.
If someone finds and reports such a regression soon after its release then we will treat it as a bug and fix it to restore the previous behavior in future releases, unless there is a very significant justification such as a security advisory. In these cases, we'll typically recommend anyone affected by the regression remain on the previous version until the problem is fixed and then skip forward directly to the new release containing that fix.
You can minimize the risk of being affected by missed regressions in final releases by proactively testing modules against alpha, beta, and release candidate packages. We recommend doing so only in isolated development or staging environments rather than against your production infrastructure. If you find a change in behavior in a prerelease build that seems contrary to the promises in this document, please open an issue in Terraform's GitHub repository to discuss it.
In the most extreme case, there may be a regression with a combination of features that is so rare that it remains undetected for a long time.
After a change has been included in more releases it becomes increasingly likely that other users will have depended on the newer behavior and thus we will need to make a tradeoff to decide whether restoring the behavior would have a greater negative impact than retaining the new behavior. We will always make this decision with due consideration to the implications of each unique situation.
You can minimize the risk of your modules being affected by late-reported regressions by upgrading promptly to new minor and patch releases of Terraform and reporting any compatibility problems you encounter in Terraform's GitHub repository.
We are making the promises above in good faith, with the intent that your investment in writing Terraform modules or automation will not be invalidated by future changes to Terraform. However, broad promises like the above can't possibly cover all nuances of practical problems that might arise as we continue to develop Terraform.
For that reason, there are some situations where we may still need to make changes that may impact existing modules or automation:
- Security problems: We may become aware of a design problem that has an important security impact. Depending on our determination of the associated risk, we may choose to break compatibility to achieve a more secure system.
- External Dependencies: Terraform's behavior depends on interfaces provided by external codebases, including your chosen operating system and including some remote network services for situations such as module and provider installation. These external systems can change outside of our control, including potentially removing or changing features that Terraform's own features depend on. In that case, if there is no suitable replacement mechanism then we may need to change Terraform's design to work within the new constraints.
- Opt-in Compatibility Breaks: The design of a language new feature may require changing the behavior or configuration representation of an existing feature. If so, we will typically make the new feature opt-in only in order to avoid breaking existing modules, but if you change your module to opt in to the new feature then you may also then be required to change other parts of your configuration to work with the new language design.
- Bugs in New Features: If we introduce a new feature to Terraform and the initial implementation has problems that cause it to not match the documented design intent at release, we may make a follow-up release that corrects the implementation to match the documented design, even if that represents a minor compatibility regression compared to the initial implementation. However, once a feature is well-established and in common use we will usually defer to the implemented behavior and instead change the documentation to reflect it.
- Regressions in Existing Features: If we learn that a new Terraform release includes a regression for an existing feature that wasn't detected during the development and prerelease periods, and that learning comes promptly after the new release, we will typically restore the previous behavior at the expense of technically therefore breaking compatibility with the behavior of the new release, under the assumption that more users will have systems affected by the regression than will have systems depending on the newly-introduced behavior.
- Late-reported regressions: As described in the previous section, if we learn that there was an unintentional regression of a rarely-used feature or combination of features in a much earlier release then restoring the previous behavior may appear as a regression to later adopters. If we believe that fixing the regression would affect more users than the regression itself affects then we may choose to accept the regression as the new promised behavior.
- Situations we cannot anticipate: Although we've made an effort to consider various specific exceptional situations here, Terraform and its development process are not isolated from broader context, and so we must consider that there may be situations that we cannot possibly anticipate that would affect the future of Terraform. In those situations, we will always do our best to find a course of action that will minimize as much as possible the impact to existing modules and automation.
Our intent with these pragmatic exceptions is only to acknowledge that there will always be situations that general compatibility promises cannot address. We will use these exceptions only with due consideration and as a last resort.