Add Dockerfile for self-hosted runner bolierplate

Pull Request: #5 (main)

.github/workflows/ci.yml

@@ -7,8 +7,10 @@ on:
   push:
     branches:
       - main
+
 jobs:
-  test:
+  # Tests ci-storage tool itself.
+  ci-storage-test:
     runs-on: ubuntu-latest
     steps:
       - name: Checkout
@@ -32,3 +34,38 @@ jobs:
           set -e
           ls -la ~/ci-storage/dimikot/ci-storage
           [ "$(cat dummy.txt)" = "dummy" ] || { echo "dummy.txt was not restored"; exit 1; }
+
+  # Builds and boots a self-hosted runner inside GitHub's infra. Once it's
+  # settled, there is a container with one self-hosted runner running and
+  # waiting for jobs with "ci-storage-test" tag to pick up (based on
+  # docker/Dockerfile image).
+  self-hosted-runner-boot-docker:
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    permissions:
+      actions: write
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v2
+      - name: Build Docker image
+        run: cd docker && docker-compose build
+      - name: Start Docker container
+        run: cd docker && docker-compose up
+        env:
+          GH_REPOSITORY: ${{ github.repository }}
+          GH_LABELS: ci-storage-test
+          GH_TOKEN: ${{ secrets.CI_PAT  }}
+
+  # The test job with ci-storage-test tag which is initially queued, but then is
+  # picked up by the self-hosted runner booted in the previous job. In the end,
+  # the test job sends SIGINT to the container entrypoint.sh PID, so the
+  # container (based on docker/Dockerfile image) shuts down gracefully.
+  self-hosted-runner-spawn-job-test:
+    runs-on: ["self-hosted", "ci-storage-test"]
+    steps:
+      - name: Run Hello World job and then terminate run.sh
+        run: |
+          set -e -o xtrace
+          echo "Hello, world!"
+          cd /home/ubuntu/actions-runner
+          kill -SIGINT $(cat runner.pid)

docker/Dockerfile

@@ -0,0 +1,45 @@
+FROM ubuntu:22.04
+
+ARG RUNNER_VERSION=2.314.1
+
+ENV GH_REPOSITORY=""
+ENV GH_LABELS=""
+ENV GH_TOKEN=""
+
+ENV DEBIAN_FRONTEND=noninteractive
+RUN true \
+    && apt-get update -y \
+    && apt-get upgrade -y \
+    && apt-get install -y --no-install-recommends \
+      awscli jq gh \
+      mc gcc git curl wget pv psmisc unzip vim nano telnet net-tools bash-completion \
+      libssl-dev apt-transport-https build-essential ca-certificates locales pkg-config
+
+RUN true \
+    && useradd -m ubuntu
+
+USER ubuntu
+RUN true \
+    && mkdir /home/ubuntu/actions-runner \
+    && cd /home/ubuntu/actions-runner \
+    && arch=$(dpkg --print-architecture) \
+    && case "$arch" in \
+      x86_64|amd64) arch=linux-x64 ;; \
+      aarch64|arm64) arch=linux-arm64 ;; \
+      *) echo >&2 "unsupported architecture: $arch"; exit 1 ;; \
+    esac \
+    && curl --no-progress-meter -L https://github.com/actions/runner/releases/download/v$RUNNER_VERSION/actions-runner-$arch-$RUNNER_VERSION.tar.gz | tar xz
+
+USER root
+RUN /home/ubuntu/actions-runner/bin/installdependencies.sh \
+    apt-get autoremove \
+    && apt-get clean \
+    && apt-get autoclean \
+    && rm -rf /var/lib/apt/lists/*
+
+USER ubuntu
+COPY --chmod=755 --chown=ubuntu:ubuntu entrypoint.sh /home/ubuntu
+
+WORKDIR /home/ubuntu
+ENTRYPOINT ["./entrypoint.sh"]
+CMD ["./run.sh"]

docker/docker-compose.yml

@@ -0,0 +1,10 @@
+version: "3.4"
+services:
+  ci-storage:
+    build:
+      context: .
+      dockerfile: ./Dockerfile
+    environment:
+      - GH_REPOSITORY
+      - GH_LABELS
+      - GH_TOKEN

docker/entrypoint.sh

@@ -0,0 +1,55 @@
+#!/bin/bash
+#
+# Here we make an opinionated decision to NOT use ephemeral or jit acton
+# runners. Reasons:
+# - We WANT to reuse the work directory across job runs, that's the whole point
+#   of ci-storage architecture and its speedup benefits. So once the runner
+#   finishes some job, we do NOT want it to terminate (as it does in ephemeral
+#   or jit mode), we want it to CONTINUE listening for more jobs to run.
+# - GitHub doesn't allow to remove busy runners via API, which is very good for
+#   us: in case the container shuts down externaly due to downscaling, we just
+#   enter the graceful retry loop to delete the corresponding runner via API.
+# - One downside happens when a runner container dies unexpectedly (rare). In
+#   this case, regular "offline" long-living runners are auto-removed by GitHub
+#   itself once in 2 weeks, whilst ephemeral (or jit) "offline" runners are
+#   auto-removed in 1 day. But we anyways need to implement some manual removal
+#   cycle exernally, since even 1 day is way too much for garbage accumulation.
+#
+set -u -e -o xtrace
+
+: $GH_REPOSITORY # {owner}/{repo}
+: $GH_LABELS
+: $GH_TOKEN # used by gh cli
+
+cd ./actions-runner
+
+name="ci-storage-$(hostname)"
+
+token=$(gh api -X POST --jq .token "repos/$GH_REPOSITORY/actions/runners/registration-token")
+./config.sh \
+  --unattended \
+  --url https://github.com/$GH_REPOSITORY \
+  --token "$token" \
+  --name "$name" \
+  --labels "$GH_LABELS"
+
+cleanup() {
+  # Retry deleting the runner until it succeeds.
+  # - Busy runner fails in deletion, so we can retry safely until it becomes
+  #   idle and is successfully deleted.
+  # - The extrnal orchestrator will eventually kill the container after a large
+  #   timeout (say, 15 minutes or so) needed for a running job to finish.
+  while :; do
+    token=$(gh api -X POST --jq .token "repos/$GH_REPOSITORY/actions/runners/remove-token")
+    ./config.sh remove --token "$token" && break
+    sleep 5
+    : "Retrying deletion till the runner becomes idle and succeeds..."
+  done
+}
+
+trap "cleanup; exit 130" INT
+trap "cleanup; exit 143" TERM
+
+echo $$ > runner.pid
+
+eval "$@" & wait $!