» Upgrading to Terraform v0.14

Terraform v0.14 is a major release and so it includes some small changes in behavior that you may need to consider when upgrading. This guide is intended to help with that process.

The goal of this guide is to cover the most common upgrade concerns and issues that would benefit from more explanation and background. The exhaustive list of changes will always be the Terraform Changelog. After reviewing this guide, we recommend reviewing the Changelog to check for specific notes about less-commonly-used features. This guide is also not intended as an overview of the new features in Terraform v0.14. This release includes other enhancements that don't need any special attention during upgrade, but those are described in the changelog and elsewhere in the Terraform documentation.

This guide focuses on changes from v0.13 to v0.14. Terraform supports upgrade tools and features only for one major release upgrade at a time, so if you are currently using a version of Terraform prior to v0.13 please upgrade through the latest minor releases of all of the intermediate versions first, reviewing the previous upgrade guides for any considerations that may be relevant to you.

In particular, Terraform v0.14 no longer includes the terraform 0.13upgrade command for automatically inserting provider requirements into existing modules, and the automatic mechanisms to upgrade legacy provider references in the Terraform state. You will need to successfully complete a terraform apply at least once under Terraform v0.13 before upgrading an existing configuration to Terraform v0.14.

Upgrade guide sections:

» Before You Upgrade

Terraform v0.14 does not support legacy Terraform state snapshot formats from prior to Terraform v0.13, so before upgrading to Terraform v0.14 you must have successfully run terraform apply at least once with Terraform v0.13 so that it can complete its state format upgrades.

When upgrading between major releases, we always recommend ensuring that you can run terraform plan and see no proposed changes on the previous version first, because otherwise pending changes can add additional unknowns into the upgrade process. Terraform v0.14 has the additional requirement of running terraform apply, as described above, because that allows Terraform v0.13 to commit the result of its automatic state format upgrades.

» Provider Dependency Lock File

In Terraform v0.13 and earlier, the terraform init command would always install the newest version of any provider in the configuration that would meet the configured version constraints.

That meant that unless the configuration author manually entered exact version constraints (for a particular version alone), a later provider release could potentially cause a change in behavior for an existing configuration even though the configuration itself had not changed.

We believe that, as far as possible, the behavior of a configuration that has already been written and tested should remain consistent unless it is intentionally changed by its author, and that intentional changes should be represented in files that can be included in a version control system and code review process.

To better meet that goal, Terraform v0.14 introduces a new dependency lock file, which Terraform will generate automatically after running terraform init in the same directory as your configuration's root module. This file includes the specific version numbers selected for each provider, and also includes the package checksums for the selected version to help ensure that the provider packages you depend on are not changed in-place upstream, whether accidentally or maliciously.

This new behavior is designed so that for most users it will not require a significant change in workflow. After running terraform init for the first time after upgrading you will find a new file .terraform.lock.hcl in your root module directory, and terraform init will automatically read and respect the entries in that file on future runs with no additional action on your part. We strongly recommend that you commit this file to your version control system, but if you do not then Terraform's behavior will be very similar to the old v0.13 behavior.

There are some situations that require some further consideration though, and those are discussed in the following sections.

» Opting out of dependency locking

We understand that not all teams share our belief that upgrades should always be represented as changes to the code in a version control repository. Those teams may have previously intentionally used a non-exact version constraint for one or more providers in order to automatically adopt any future provider releases and then make any necessary configuration changes in retrospect.

You can continue with a model similar to the v0.13 behavior after upgrading to v0.14 by placing .terraform.lock.hcl in your version control system's "ignore" file, such as .gitignore for Git. In that case, Terraform will see the lock file in the same way as the internal index that Terraform v0.13 would generate under the .terraform directory, preserving the selections only with in a particular working directory until you delete the file.

With that said, we do recommend that teams consider carefully the benefits of a persistent lock file, and make a considered decision about which path to take. We believe that a lock file under version control will be the best choice for most teams, because we've seen this pattern used successfully in many other programming language ecosystems.

» In-house providers and internal mirrors

Terraform v0.13 introduced a new hierarchical namespace for providers, which was an important pre-requisite for introducing a dependency lock file in v0.14 which can support a mixture of official, partner, community and in-house providers in a single configuration.

If you followed the advice from the Terraform v0.13 upgrade guide about assigning your in-house providers their own unique source addresses, and you're distributing your in-house providers to Terraform through one of the documented mechanisms, Terraform should handle selecting a version and recording its checksums in the same way for all of the providers you use.

However, the full functionality of the lock file does depend on some information only available from a real provider registry, so there are some special considerations for providers installed from either filesystem or network mirrors:

  • Mirror sources don't provide package checksums that are signed by the original provider distributor, so by default terraform init will record only the checksum for the provider package on the platform you are currently using.

For example, if you run terraform init on a macOS system and then commit the lock file, a collegue running terraform init on a Linux system may encounter a checksum error, because the mirror-based installation was only able to safely record the checksum for the package it actually installed.

This problem doesn't arise for installation from a provider registry because the registry can provide signed checksums for all packages for a particular provider version, across all supported platforms.

If you use mirrors exclusively in your environment and you use Terraform across a mixture of platforms then, in addition to making sure that your mirrors include packages for all of the necessary platforms, you may choose to use the new terraform providers lock command to pre-enter the required lock file entries for all of the platforms you intend to use.

  • Terraform expects a particular version of a provider to have identical package checksums across all installation methods you work with in your team.

For example, if you use direct installation from Terraform registries in your development environment but you use a mirror in your production automation, you must ensure that the packages available for a particular provider version in your mirror are identical to the official packages in the origin registry.

If your internal mirrors intentionally have different packages than are available upstream, you must either use your internal mirrors consistently (so Terraform never uses or verifies an official package) or you must publish your own builds so that Terraform can understand your intent for them to be distinct.

If you are only making minor or temporary changes to a provider, such as building for a platform that Terraform doesn't official support or including a bugfix patch that isn't yet in an upstream release, the simplest answer is to number your local build with semantic versioning build metadata, such as v2.1.0+companyname.1 where v2.1.0 is the upstream release you derived yours from, companyname is a short mnemonic for your organization, and .1 is an internal build id that you can potentially increment if you need to make ongoing new builds from the same upstream version.

If you are making more substantial changes to a provider, such as adding entirely new features that your modules depend on, it may be better to instead publish the provider under a separate namespace you control, such as publishing a fork of hashicorp/aws as companyname/aws in the public registry or tf.example.com/companyname/aws for in-house distribution only. This is a more drastic approach in that Terraform will understand your release as an entirely separate provider, but it also allows your modules to clearly indicate that they depend on the features of your fork rather than the features of the upstream release.

In both cases the dependency lock file will see your releases as distinct from the upstream ones and thus expect the two to have a different set of checksums each.

» External module dependencies are not locked

Although we do hope to eventually include a means to lock version selections for external modules in addition to providers, this new capability is limited only to providers in Terraform v0.14.

Terraform modules have a different approach to distribution and versioning than Terraform providers, with many different supported installation methods that each require careful consideration in designing a dependency locking mechanism.

If you wish to lock your module dependencies then for now you must continue to use the same strategy as for v0.13 and earlier: specify exact version constraints for modules distributed via a module registry, or use the source-type-specific mechanisms to lock to a particular version of module packages retrieved directly using other protocols.

Note that Terraform also does not currently track checksums for external module dependencies. If you are concerned about the possibility of external modules being altered in-place without your knowledge, we recommend using modules only from sources directly under your control, such as a private Terraform module registry.

» The local provider cache directory

As an implementation detail of automatic provider installation, Terraform has historically unpacked auto-installed plugins under the local cache directory in .terraform/plugins. That directory was only intended for Terraform's internal use, but unfortunately due to a miscommunication within our team it was inadvertently documented as if it were a "filesystem mirror" directory that you could place local providers in to upload them to Terraform Cloud.

Unfortunately the implementation details have changed in Terraform v0.14 in order to move the authority for provider version selection to the new dependency lock file, and so manually placing extra plugins into that local cache directory is no longer effective in Terraform v0.14.

We've included a heuristic in terraform init for Terraform v0.14 which should detect situations where you're relying on an unofficial provider manually installed into the cache directory and generate a warning like the following:

Warning: Missing provider is in legacy cache directory

Terraform supports a number of local directories that can serve as automatic
local filesystem mirrors, but .terraform/plugins is not one of them because
Terraform v0.13 and earlier used this directory to cache copies of provider
plugins retrieved from elsewhere.

If you intended to use this directory as a filesystem mirror for
tf.example.com/awesomecorp/happycloud, place it instead in the following

The error message suggests using the terraform.d directory, which is a local search directory originally introduced in Terraform v0.10 in order to allow sending bundled providers along with your configuration up to Terraform Cloud. The error message assumes that use-case because it was for Terraform Cloud in particular that this approach was previously mis-documented.

If you aren't intending to upload the provider plugin to Terraform Cloud as part of your configuration, we recommend instead installing to one of the other implied mirror directories, or you can explicitly configure some custom provider installation methods if your needs are more complicated.

» Concise Terraform Plan Output

In Terraform v0.11 and earlier, the output from terraform plan was designed to show only the subset of resource instance attributes that had actually changed compared to the prior state.

Although that made the output very concise, we heard from several users that the lack of context in the output had led to some misunderstandings that in turn caused production outages. We know that reviewing a Terraform plan can be a point of anxiety for those working on production infrastructure, so we responded to that feedback in Terraform v0.12 by having the plan output instead show the full context of each resource instance that has a planned action, and then use extra annotations (+, -, ~) to mark the specific attributes that will change.

Based on further feedback since the v0.12 release, we understand that the new detailed output has been very overwhelming for resource types that have a large number of attributes or deeply nested block structures. Terraform v0.14 introduces a new compromise that aims to still address the concern about context while allowing better focus on the parts of each object that are changing.

For this initial release, Terraform will omit from the plan output any attribute that has not changed, with the exception of a number of attribute names whose values often contain human-recognizable identifying information. When attributes or blocks are omitted, Terraform will always include a summary of what isn't included, to avoid ambiguity with an argument merely being unset.

This is intended as an incremental step to improve the worst cases of verbose output in Terraform v0.12 and v0.13, but the compromises we made here may not be suitable for all situations. If you'd like to retain the fully-verbose output from Terraform v0.13, you can temporarily re-enable it by setting the environment variable TF_X_CONCISE_DIFF=0 when you run Terraform.

If you choose to opt out of the new concise mode, please open a feature request issue to let us know what you found lacking in the new output. We intend to continue iterating on the design tradeoffs here to find the best compromise to suit the needs of most users. We expect to remove the opt-out environment variable in Terraform v0.15.

» Sensitive Values in Plan Output

In Terraform v0.13 and earlier, Terraform allowed provider authors to mark certain resource type attributes as being "sensitive", and similarly allowed module authors to mark certain output values as "sensitive". Terraform would then show the placeholder string (sensitive value) in the plan output, instead of the actual value.

Terraform v0.14 introduces a more extensive version of that behavior where Terraform will track when you write an expression whose result is derived from a sensitive input variable or sensitive output value, and so after upgrading to Terraform v0.14 you may find that more values are obscured in the Terraform plan output than would have been in Terraform v0.13.

If a sensitive value (either derived from a sensitive input variable or a sensitive output variable) is used in another module output, that output must be marked sensitive as well to be explicit about this data being passed through Terraform:

variable "foo" {
  sensitive = true

output "bar" {
  value     = var.foo
  # sensitive must be true when referencing a sensitive input variable
  sensitive = true

There is also experimental behavior that will extend this sensitivity-awareness to attributes providers define as sensitive. You can enable this feature by activating the experiment in the terraform block:

terraform {
  experiments = [provider_sensitive_attrs]

If you enable this experiment, attributes that are defined by a given provider as sensitive will have the same sensitivity-tracking behavior as sensitive input values and outputs. For example, the vault_generic_secret data source has an attribute data that is sensitive according to this provider's schema.

# mod/main.tf

terraform {
  experiments = [provider_sensitive_attrs]

data "vault_generic_secret" "foobar" {
  path = "secret/foobar"

output "token" {
  value = vault_generic_secret.foobar.data["token"]
  # a error will display if sensitive = true is not here

If you do not add sensitive = true to the output referencing that sensitive attribute, you will get an error:

Error: Output refers to sensitive values

  on mod/main.tf line 6:
   6: output "token" {

Expressions used in outputs can only refer to sensitive values if the
sensitive attribute is true.

For this feature we've taken the approach that it's better to be conservative and obscure potentially-sensitive values at the expense of potentially also obscuring some values that aren't sensitive. Unfortunately this means that if you've written a module in a generic or dynamic way then Terraform may over-generalize which values are sensitive, leading to less helpful plan output.

Due to the security implications of this feature, Terraform offers no direct way to opt out of this change. However, the obscuring of these values is done at the UI layer only and so you can still access the raw values, if needed, by saving your plan to an plan file and then asking Terraform to present it in machine-readable JSON format:

terraform plan -out=tfplan
terraform show -json tfplan

Please note that the binary file tfplan and the JSON output produced from it can both include cleartext representations of sensitive values, so writing these to disk on a multi-user system or viewing the JSON output on-screen may cause those values to become visible to others.

Sensitive values are also still saved in state snapshots stored in your configured backend. Use the access control and audit mechanisms offered by the remote system to control who can access that data.

» Other Important Workflow Changes

» Terraform Output Formatting

We've modified the formatting of terraform output to match the formatting of terraform show.

We consider the console output of Terraform human readable; specifically designed and optimized for operators and practitioners to review themselves. As a result we occasionally (maybe even regularly) intend to tweak that output to help improve consistency, clarity, actionability and more.

If you rely on terraform output in automation, please use terraform output -json.