Merge pull request #1245 from buildkite/pdp-1828-take-over-terminate-instance-pr

Author: triarius

packer/linux/conf/bin/bk-install-elastic-stack.sh

@@ -131,6 +131,10 @@ echo Writing Phase 2/2 for /var/lib/buildkite-agent/cfn-env helper function...
 cat <<EOF >>/var/lib/buildkite-agent/cfn-env
 
 set_always         "BUILDKITE_AGENTS_PER_INSTANCE" "$BUILDKITE_AGENTS_PER_INSTANCE"
+
+# also set via /etc/systemd/system/buildkite-agent.service.d/environment.conf
+set_always         "BUILDKITE_TERMINATE_INSTANCE_AFTER_JOB" "$BUILDKITE_TERMINATE_INSTANCE_AFTER_JOB"
+
 set_always         "BUILDKITE_ECR_POLICY" "${BUILDKITE_ECR_POLICY:-none}"
 set_always         "BUILDKITE_SECRETS_BUCKET" "$BUILDKITE_SECRETS_BUCKET"
 set_always         "BUILDKITE_SECRETS_BUCKET_REGION" "$BUILDKITE_SECRETS_BUCKET_REGION"
@@ -357,6 +361,17 @@ done
 echo "Waited $next_wait_time times for docker to start. We will exit if it still has not started."
 check_docker
 
+echo Writing buildkite-agent systemd environment override...
+# also set in /var/lib/buildkite-agent/cfn-env so that it's shown in the job logs
+mkdir -p /etc/systemd/system/buildkite-agent.service.d
+cat <<EOF | tee /etc/systemd/system/buildkite-agent.service.d/environment.conf
+[Service]
+Environment="BUILDKITE_TERMINATE_INSTANCE_AFTER_JOB=${BUILDKITE_TERMINATE_INSTANCE_AFTER_JOB}"
+EOF
+
+echo Reloading systemctl services...
+systemctl daemon-reload
+
 echo Starting buildkite-agent...
 systemctl enable --now buildkite-agent
 

packer/linux/conf/buildkite-agent/scripts/terminate-instance

@@ -2,14 +2,51 @@
 
 set -euo pipefail
 
-echo "sleeping for 10 seconds before terminating instance to allow agent logs to drain to cloudwatch..."
+terminate() {
+  aws autoscaling terminate-instance-in-auto-scaling-group \
+    --region "$1" \
+    --instance-id "$2" \
+    --should-decrement-desired-capacity
+}
+
+mark_as_unhealthy() {
+  aws autoscaling set-instance-health \
+    --region "$1" \
+    --instance-id "$2" \
+    --health-status Unhealthy
+}
 
+echo "sleeping for 10 seconds before terminating instance to allow agent logs to drain to cloudwatch..."
 sleep 10
 
-token=$(curl -X PUT -H "X-aws-ec2-metadata-token-ttl-seconds: 60" --fail --silent --show-error --location "http://169.254.169.254/latest/api/token")
-instance_id=$(curl -H "X-aws-ec2-metadata-token: $token" --fail --silent --show-error --location "http://169.254.169.254/latest/meta-data/instance-id")
-region=$(curl -H "X-aws-ec2-metadata-token: $token" --fail --silent --show-error --location "http://169.254.169.254/latest/meta-data/placement/region")
+token=$(
+  curl \
+    --fail --silent --show-error \
+    -X PUT \
+    -H "X-aws-ec2-metadata-token-ttl-seconds: 60" \
+    --location "http://169.254.169.254/latest/api/token"
+)
+instance_id=$(
+  curl \
+    --fail --silent --show-error \
+    -H "X-aws-ec2-metadata-token: $token" \
+    --location "http://169.254.169.254/latest/meta-data/instance-id"
+)
+region=$(
+  curl \
+    --fail --silent --show-error \
+    -H "X-aws-ec2-metadata-token: $token" \
+    --location "http://169.254.169.254/latest/meta-data/placement/region"
+)
 
 echo "requesting instance termination..."
-
-aws autoscaling terminate-instance-in-auto-scaling-group --region "$region" --instance-id "$instance_id" "--should-decrement-desired-capacity"
+if [[ $BUILDKITE_TERMINATE_INSTANCE_AFTER_JOB == "true" ]]; then
+  # If the ASG is at the min capacity, the call to terminate-instance-in-autoscaling-group
+  # In this case, we mark the instance as unhealthy, then the ASG will spin up a new instance
+  # to replace it.
+  terminate "$region" "$instance_id" || mark_as_unhealthy "$region" "$instance_id"
+else
+  # If we're not in terminate-after-job mode, then it's fine for this to fail. Systemd will restart
+  # the agent and it'll be as if the instance never got shut down.
+  terminate "$region" "$instance_id"
+fi

packer/windows/conf/bin/bk-install-elastic-stack.ps1

@@ -72,6 +72,10 @@ function set_always() {
 Add-Content -Path C:\buildkite-agent\cfn-env -Value @"
 
 set_always         "BUILDKITE_AGENTS_PER_INSTANCE" "$Env:BUILDKITE_AGENTS_PER_INSTANCE"
+
+# also set via nssm
+set_always         "BUILDKITE_TERMINATE_INSTANCE_AFTER_JOB" "$Env:BUILDKITE_TERMINATE_INSTANCE_AFTER_JOB"
+
 set_always         "BUILDKITE_ECR_POLICY" "$Env:BUILDKITE_ECR_POLICY"
 set_always         "BUILDKITE_SECRETS_BUCKET" "$Env:BUILDKITE_SECRETS_BUCKET"
 set_always         "BUILDKITE_SECRETS_BUCKET_REGION" "$Env:BUILDKITE_SECRETS_BUCKET_REGION"
@@ -146,7 +150,7 @@ tracing-backend=${Env:BUILDKITE_AGENT_TRACING_BACKEND}
 "@
 $OFS=" "
 
-nssm set lifecycled AppEnvironmentExtra :AWS_REGION=$Env:AWS_REGION
+nssm set lifecycled AppEnvironmentExtra +AWS_REGION=$Env:AWS_REGION
 nssm set lifecycled AppEnvironmentExtra +LIFECYCLED_HANDLER="C:\buildkite-agent\bin\stop-agent-gracefully.ps1"
 Restart-Service lifecycled
 
@@ -212,26 +216,35 @@ Write-Output "Starting the Buildkite Agent"
 
 nssm install buildkite-agent C:\buildkite-agent\bin\buildkite-agent.exe start
 If ($lastexitcode -ne 0) { Exit $lastexitcode }
+
 nssm set buildkite-agent ObjectName .\$UserName $Password
 If ($lastexitcode -ne 0) { Exit $lastexitcode }
+
 nssm set buildkite-agent AppStdout C:\buildkite-agent\buildkite-agent.log
 If ($lastexitcode -ne 0) { Exit $lastexitcode }
+
 nssm set buildkite-agent AppStderr C:\buildkite-agent\buildkite-agent.log
 If ($lastexitcode -ne 0) { Exit $lastexitcode }
-nssm set buildkite-agent AppEnvironmentExtra :HOME=C:\buildkite-agent
+
+nssm set buildkite-agent AppEnvironmentExtra +HOME=C:\buildkite-agent
 
 If ((![string]::IsNullOrEmpty($Env:BUILDKITE_ENV_FILE_URL)) -And (Test-Path -Path C:\buildkite-agent\env -PathType leaf)) {
   foreach ($var in Get-Content C:\buildkite-agent\env) {
-    nssm set buildkite-agent AppEnvironmentExtra $var
+    nssm set buildkite-agent AppEnvironmentExtra "+$var"
     If ($lastexitcode -ne 0) { Exit $lastexitcode }
   }
 }
 
+# also set in cfn so it's show in job logs
+nssm set buildkite-agent AppEnvironmentExtra +BUILDKITE_TERMINATE_INSTANCE_AFTER_JOB=$Env:BUILDKITE_TERMINATE_INSTANCE_AFTER_JOB
 If ($lastexitcode -ne 0) { Exit $lastexitcode }
+
 nssm set buildkite-agent AppExit Default Restart
 If ($lastexitcode -ne 0) { Exit $lastexitcode }
+
 nssm set buildkite-agent AppRestartDelay 10000
 If ($lastexitcode -ne 0) { Exit $lastexitcode }
+
 nssm set buildkite-agent AppEvents Exit/Post "powershell C:\buildkite-agent\bin\terminate-instance.ps1"
 If ($lastexitcode -ne 0) { Exit $lastexitcode }
 

packer/windows/conf/buildkite-agent/scripts/terminate-instance.ps1

@@ -3,13 +3,28 @@ $Token = (Invoke-WebRequest -UseBasicParsing -Method Put -Headers @{'X-aws-ec2-m
 $InstanceId = (Invoke-WebRequest -UseBasicParsing -Headers @{'X-aws-ec2-metadata-token' = $Token} http://169.254.169.254/latest/meta-data/instance-id).content
 $Region = (Invoke-WebRequest -UseBasicParsing -Headers @{'X-aws-ec2-metadata-token' = $Token} http://169.254.169.254/latest/meta-data/placement/region).content
 
+Write-Output "terminate-instance: disconnecting agent..."
+nssm stop buildkite-agent
+
 Write-Output "terminate-instance: requesting instance termination..."
 aws autoscaling terminate-instance-in-auto-scaling-group --region "$Region" --instance-id "$InstanceId" "--should-decrement-desired-capacity" 2> $null
 
-if ($lastexitcode -eq 0) { # If autoscaling request was successful, we will terminate
-  Write-Output "terminate-instance: disabling buildkite-agent service"
-  nssm stop buildkite-agent
-}
-else {
+# If autoscaling request was successful, we will terminate the instance, otherwise, if
+# BuildkiteTerminateInstanceAfterJob is set to true, we will mark the instance as unhealthy
+# so that the ASG will terminate it despite scale-in protection. Otherwise, we should not
+# terminate the instance, so we need to retart the agent.
+if ($lastexitcode -eq 0) {
+  Write-Output "terminate-instance: terminating instance..."
+} else {
   Write-Output "terminate-instance: ASG could not decrement (we're already at minSize)"
+  if ($Env:BUILDKITE_TERMINATE_INSTANCE_AFTER_JOB -eq "true") {
+    Write-Output "terminate-instance: marking instance as unhealthy"
+    aws autoscaling set-instance-health `
+      --instance-id "$InstanceId" `
+      --region "$Region" `
+      --health-status Unhealthy
+  } else {
+    Write-Output "terminate-instance: restarting agent..."
+    nssm start buildkite-agent
+  }
 }