From e34b93975233feafe80da522777bb2d6f58c3c0d Mon Sep 17 00:00:00 2001
From: Masayuki Morita <minamijoyo@gmail.com>
Date: Thu, 2 Mar 2023 15:47:36 +0900
Subject: [PATCH] Export TF_PLUGIN_CACHE_MAY_BREAK_DEPENDENCY_LOCK_FILE=true in
 CI

Starting from Terraform v1.4, launching terraform providers in the
acceptance test has been failing more frequently with a text file busy
error.

```
--- FAIL: TestAccMultiStateMigratorApplySimple (1.07s)
    multi_state_migrator_test.go:123: failed to run terraform init: failed to run command (exited 1): terraform init -input=false -no-color
        stdout:

        Initializing the backend...

        Successfully configured the backend "s3"! Terraform will automatically
        use this backend unless the backend configuration changes.

        Initializing provider plugins...
        - Finding latest version of hashicorp/null...
        - Installing hashicorp/null v3.2.1...

        stderr:

        Error: Failed to install provider

        Error while installing hashicorp/null v3.2.1: open
        /tmp/plugin-cache/registry.terraform.io/hashicorp/null/3.2.1/linux_amd64/terraform-provider-null_v3.2.1_x5:
        text file busy
```

After some investigation, I found Go's `os/exec.Cmd.Run()` does not wait
for the grandchild process to complete; from the point of view of
tfmigrate, the terraform command is the child process, and the provider
is the grandchild process.

https://github.com/golang/go/issues/23019

If I understand correctly, this is not a Terraform issue and
theoretically should occur in versions older than v1.4; the changes in
v1.4 may have broken the balance of execution timing and made the test
very flaky. I experimented with inserting some sleep but could not get
the test to stabilize correctly. After trying various things, I found
that the test became stable by enabling the
`TF_PLUGIN_CACHE_MAY_BREAK_DEPENDENCY_LOCK_FILE` flag was introduced in
v1.4. This is an escape hatch to revert to the v1.3 equivalent of the
global cache behavior change in v1.4.
https://github.com/hashicorp/terraform/pull/32726

This behavior change has already been addressed in the previous commit
using a local file system mirror, so activating this flag does not seem
to make any sense. Even though I have no other reasonable solutions now,
please let me know if anyone finds a better solution.
---
 docker-compose.yml | 1 +
 1 file changed, 1 insertion(+)

diff --git a/docker-compose.yml b/docker-compose.yml
index 8186805..1657c6b 100644
--- a/docker-compose.yml
+++ b/docker-compose.yml
@@ -14,6 +14,7 @@ services:
       # Use the same filesystem to avoid a checksum mismatch error
       # or a file busy error caused by asynchronous IO.
       TF_PLUGIN_CACHE_DIR: "/tmp/plugin-cache"
+      TF_PLUGIN_CACHE_MAY_BREAK_DEPENDENCY_LOCK_FILE: "true"
       # From observation, although we don’t have complete confidence in the root cause,
       # it appears that localstack sometimes misses API requests when run in parallel.
       TF_CLI_ARGS_apply: "--parallelism=1"