From 0c91fb7d3d437933ff6dc2419276c639594eff54 Mon Sep 17 00:00:00 2001
From: Nick Volynkin <nick.volynkin@gmail.com>
Date: Thu, 10 Nov 2022 16:26:30 +0600
Subject: [PATCH] prepare-workspace: make actions/checkout more reliable

Make an empty detached commit for actions/checkout to properly
clean the project's workspace before checking out new code
revision. It makes the actions/checkout resilient to
various problems related to git submodules.

Part of #28
Related to tarantool/tarantool-qa#145
---
 prepare-checkout/README.md  | 56 +++++++++++++++++++++++++++++++++++++
 prepare-checkout/action.yml | 14 ++++++++++
 2 files changed, 70 insertions(+)
 create mode 100644 prepare-checkout/README.md
 create mode 100644 prepare-checkout/action.yml

diff --git a/prepare-checkout/README.md b/prepare-checkout/README.md
new file mode 100644
index 0000000..5fa5916
--- /dev/null
+++ b/prepare-checkout/README.md
@@ -0,0 +1,56 @@
+# Prepare workspace for checkout action
+
+This action creates and checks out an empty detached commit.
+It helps subsequent [actions/checkout](https://github.com/actions/checkout) action
+correctly clean the workspace.
+
+## Usage
+
+Add the following line to your workflow before the `actions/checkout` action:
+
+```diff
++ - uses: tarantool/actions/prepare-checkout@master
+  - uses: actions/checkout@v3
+    ...
+```
+
+## Explanation and rationale
+
+This actions is solution to 
+[tarantool/tarantool-qa#145](https://github.com/tarantool/tarantool-qa/issues/145).
+First attempt to solve this was the tarantool/actions/cleanup action.
+The tarantool/actions/prepare-checkout action is a softer alternative to it.
+
+When submodules change, actions/checkout can fail with git errors. 
+It's a well-known problem and the related issues are still open:
+actions/checkout#354,
+actions/checkout#385,
+actions/checkout#418, and
+actions/checkout#590.
+
+Before checking out a new revision, actions/checkout runs the following code:
+
+```bash
+# removes ignored and non-versioned files
+git clean -ffdx
+# resets workspace to the commit, on which it was left
+# after the last job run
+git reset --hard
+```
+
+The problem is that when a workflow fails because of a particular commit,
+the repository still stays on that commit. On the next job run, 
+actions/checkout will run the above code, restore that particular commit,
+and fail to make a proper code checkout.
+
+By creating a detached empty commit, this actions forces actions/checkout
+to clean up the project's workspace entirely, removing any files that could
+break checkout. Meanwhile, the `.git` directory stays intact, so full checkout
+isn't required and the workflow does not waste much time.
+
+If this is a first job run on a particular runner and there's no repository yet,
+the command in this action will silently fail, thanks to `|| :`,
+but the action itself will succeed.
+
+This action uses a solution proposed in a
+[comment](https://github.com/actions/checkout/issues/590#issuecomment-970586842) at actions/checkout#590.
diff --git a/prepare-checkout/action.yml b/prepare-checkout/action.yml
new file mode 100644
index 0000000..57d2409
--- /dev/null
+++ b/prepare-checkout/action.yml
@@ -0,0 +1,14 @@
+---
+name: 'Prepare workspace for checkout'
+description: 'Prepare workspace for the actions/checkout action'
+runs:
+  using: 'composite'
+  steps:
+    - run: |
+        set -e
+        shopt -s dotglob
+        git checkout -f \
+          $(git -c user.name=TarantoolBot -c user.email=bot@tarantool.io commit-tree \
+          $(git hash-object -t tree /dev/null) -m 'empty commit' \
+          < /dev/null) || :
+      shell: bash