Bug: `run --all` fails to match dependencies in queue when symlinks are used

[jean-mercier-hivebrite]

Bug: run --all fails to match dependencies in queue when symlinks are used

Describe the bug

When using terragrunt run --all apply with a symlinked directory, dependencies are not correctly matched to queue entries because paths are stored in incompatible formats:

• Queue entries use relative paths: ./namespace
• Dependencies use broken hybrid paths after symlink resolution: ../../../../../real/path/to/stacks/.../namespace

This causes all dependencies to be marked as "not in queue, considering it ready", resulting in all units running in parallel regardless of their actual dependencies.

Environment

• Terragrunt version: v0.97.2
• Terraform version: 1.14.3
• OS: Linux (container)

Steps to reproduce

1. Create a directory structure:

/app/
├── terraform/
│   └── stacks/
│       └── myproject/
│           ├── namespace/
│           │   └── terragrunt.hcl
│           ├── service-a/
│           │   └── terragrunt.hcl  # depends on namespace
│           └── service-b/
│               └── terragrunt.hcl  # depends on namespace

2. Create a symlink to the terraform directory:

ln -s /app/terraform /terraform

3. In the terragrunt.hcl files, use the symlink path for dependencies:

# service-a/terragrunt.hcl
dependency "namespace" {
  config_path = "/terraform/stacks/myproject/namespace"
}

inputs = {
  namespace = dependency.namespace.outputs.name
}

4. Run terragrunt from the project directory:

cd /app/terraform/stacks/myproject
terragrunt run --all apply

Expected behavior

Dependencies should be correctly matched to queue entries regardless of symlink resolution. Units should run in dependency order:

1. namespace runs first
2. service-a and service-b wait for namespace to complete

Actual behavior

All units start simultaneously because dependencies are not matched in the queue.

Debug output shows the path mismatch:

Unit queue will be processed for apply in this order:
- Unit namespace
- Unit service-a
- Unit service-b

Runner Pool Controller: starting with 3 tasks, concurrency 2147483647

Dependency /app/terraform/stacks/myproject/namespace is not in queue, considering it ready

And when listing units with dependencies:

=> Unit service-a (excluded: false, assume applied: false, dependencies: [../../../../../app/terraform/stacks/myproject/namespace])
=> Unit service-b (excluded: false, assume applied: false, dependencies: [../../../../../app/terraform/stacks/myproject/namespace])
=> Unit namespace (excluded: false, assume applied: false, dependencies: [])

Notice:

• Queue entries use: ./namespace
• Dependencies use: ../../../../../app/terraform/stacks/myproject/namespace

The path comparison in entryByPathUnsafe fails:

// internal/queue/queue.go line ~193
func (q *Queue) entryByPathUnsafe(path string) *Entry {
    for _, entry := range q.Entries {
        if entry.Component.Path() == path {  // This comparison fails!
            return entry
        }
    }
    return nil
}

Root cause analysis

The issue appears to be in path normalization when symlinks are involved:

1. config_path = "/terraform/stacks/.../namespace" is an absolute path through a symlink
2. During discovery, filepath.EvalSymlinks() resolves this to /app/terraform/stacks/.../namespace
3. This is then converted to a relative path using filepath.Rel() from the working directory
4. The resulting path ../../../../../app/terraform/stacks/.../namespace is a broken hybrid (goes up then uses absolute path components)
5. Queue entries are stored with simple relative paths like ./namespace
6. The string comparison == fails because the formats don't match

Workaround

Use bind mounts instead of symlinks, or use the real absolute path in config_path instead of the symlinked path.

Additional context

This worked correctly in Terragrunt v0.93.x. The regression appears to be related to path handling changes in the v0.95+ releases.

Related issues that may share root cause:

• Symlinks in Terragrunt #1611 - Symlinks in Terragrunt
• dependency block with a relative path does not resolve anymore when included with expose=true #3475 - dependency block with relative path does not resolve with expose=true
• path_relative_to_include and symlinks #543 - path_relative_to_include and symlinks

[amveremes]

This problem persists in Terragrunt version v0.99.0.
The "quickestart" project is not working :
https://github.com/gruntwork-io/terragrunt/blob/pull-request-4400/docs-starlight/src/content/docs/01-getting-started/01-quick-start.mdx

[yhakbar]

Hey folks. Have you enabled the symlink experiment?
https://terragrunt.gruntwork.io/docs/reference/experiments/#symlinks

As pointed out in the symlinks experiment, support for symlinks is still experimental and requires quite a lot of comprehensive testing to be made generally available by default.

[jean-mercier-hivebrite]

Hello thanks for pointing it, i didn't test the experiment, what is "strange" is that it's was working before
it's the new version that introduce the lost of the support ?

[lidor-ra]

Similiar\Maybe related to 5314.
On my example experiment is enabled.