How Terraform Uses Unicode

The Terraform language uses the Unicode standards as the basis of various different features. The Unicode Consortium publishes new versions of those standards periodically, and we aim to adopt those new versions in new minor releases of Terraform in order to support additional characters added in those new versions.

Unfortunately due to those features being implemented by relying on a number of external libraries, adopting a new version of Unicode is not as simple as just updating a version number somewhere. This document aims to describe the various steps required to adopt a new version of Unicode in Terraform.

We typically aim to be consistent across all of these dependencies as to which major version of Unicode we currently conform to. The usual initial driver for a Unicode upgrade is switching to new version of the Go runtime library which itself uses a new version of Unicode, because Go itself does not provide any way to select Unicode versions independently from Go versions. Therefore we typically upgrade to a new Unicode version only in conjunction with upgrading to a new Go version.

Unicode tables in the Go standard library

Several Terraform language features are implemented in terms of functions in the Go strings package, the Go unicode package, and other supporting packages in the Go standard library.

The Go team maintains the Go standard library features to support a particular Unicode version for each Go version. The specific Unicode version for a particular Go version is available in unicode.Version.

We adopt a new version of Go by editing the .go-version file in the root of this repository. Although it‘s typically possible to build Terraform with other versions of Go, that file documents the version we intend to use for official releases and thus the primary version we use for development and testing. Adopting a new Go version typically also implies other behavior changes inherited from the Go standard library, so it’s important to review the relevant version changelog(s) to note any behavior changes we'll need to pass on to our own users via the Terraform changelog.

The other subsystems described below should always be set up to match unicode.Version. In some cases those libraries automatically try to align themselves with unicode.Version and generate an error if they cannot, but that isn't true of all of them.

Unicode Identifier Rules in HCL

Identifier and Pattern Syntax (TF31) is a Unicode standards annex which describe a set of rules for tokenizing “identifiers”, such as variable names in a programming language.

HCL uses a superset of that specification for its own identifier tokenization rules, and so it includes some code derived from the TF31 data tables that describe which characters belong to the “ID_Start” and “ID_Continue” classes.

Since Terraform is the primary user of HCL, it‘s typically Terraform’s adoption of a new Unicode version which drives HCL to adopt one. To update the Unicode tables to a new version:

  • Edit hclsyntax/generate.go's line which runs unicode2ragel.rb to specify the URL of the DerivedCoreProperties.txt data file for the intended Unicode version.
  • Run go generate ./hclsyntax to run the generation code to update both unicode_derived.rl and, indirectly, scan_tokens.go. (You will need both a Ruby interpreter and the Ragel state machine compiler on your system in order to complete this step.)
  • Run all the tests to check for regressions: go test ./...
  • If all looks good, commit all of the changes and open a PR to HCL.
  • Once that PR is merged and released, update Terraform to use the new version of HCL.

Unicode Text Segmentation

Text Segmentation (TR29) is a Unicode standards annex which describes algorithms for breaking strings into smaller units such as sentences, words, and grapheme clusters.

Several Terraform language features make use of the grapheme cluster algorithm in particular, because it provides a practical definition of individual visible characters, taking into account combining sequences such as Latin letters with separate diacritics or Emoji characters with gender presentation and skin tone modifiers.

The text segmentation algorithms rely on supplementary data tables that are not part of the core set encoded in the Go standard library's unicode packages, and so instead we rely on the third-party module github.com/apparentlymart/go-textseg to provide those tables and a Go implementation of the grapheme cluster segmentation algorithm in terms of the tables.

The go-textseg library is designed to allow calling programs to potentially support multiple Unicode versions at once, by offering a separate module major version for each Unicode major version. For example, the full module path for the Unicode 13 implementation is github.com/apparentlymart/go-textseg/v13.

If that external library doesn‘t yet have support for the Unicode version we intend to adopt then we’ll first need to open a pull request to contribute new language support. The details of how to do this will unfortunately vary depending on how significantly the Text Segmentation annex has changed since the most recently-supported Unicode version, but in many cases it can be just a matter of editing that library's make_tables.go, make_test_tables.go, and generate.go files to point to the URLs where the Unicode consortium published new tables and then run go generate to rebuild the files derived from those data sources. As long as the new Unicode version has only changed the data tables and not also changed the algorithm, often no further changes are needed.

Once a new Unicode version is included, the maintainer of that library will typically publish a new major version that we can depend on. Two different codebases included in Terraform all depend directly on the go-textseg module for parts of their functionality:

  • hashicorp/hcl uses text segmentation as part of producing visual column offsets in source ranges returned by the tokenizer and parser. Terraform in turn uses that library for the underlying syntax of the Terraform language, and so it passes on those source ranges to the end-user as part of diagnostic messages.
  • The third-party module github.com/zclconf/go-cty provides several of the Terraform language built in functions, including functions like substr and length which need to count grapheme clusters as part of their implementation.

As part of upgrading Terraform's Unicode support we therefore typically also open pull requests against these other codebases, and then adopt the new versions that produces. Terraform work often drives the adoption of new Unicode versions in those codebases, with other dependencies following along when they next upgrade.

At the time of writing Terraform itself doesn't directly depend on go-textseg, and so there are no specific changes required in this Terraform codebase aside from the go.sum file update that always follows from changes to transitive dependencies.

The go-textseg library does have a different “auto-version” mechanism which selects an appropriate module version based on the current Go language version, but neither HCL nor cty use that because the auto-version package will not compile for any Go version that doesn‘t have a corresponding Unicode version explicitly recorded in that repository, and so that would be too harsh a constraint for libraries like HCL which have many callers, many of which don’t care strongly about Unicode support, that may wish to upgrade Go before the text segmentation library has been updated.