[CI] Retry flaky tests (#53961) (#56060)

Jenkinsfile

@@ -4,11 +4,12 @@ library 'kibana-pipeline-library'
 kibanaLibrary.load()
 
 stage("Kibana Pipeline") { // This stage is just here to help the BlueOcean UI a little bit
-  timeout(time: 120, unit: 'MINUTES') {
+  timeout(time: 135, unit: 'MINUTES') {
     timestamps {
       ansiColor('xterm') {
         githubPr.withDefaultPrComments {
           catchError {
+            retryable.enable()
             parallel([
               'kibana-intake-agent': kibanaPipeline.legacyJobRunner('kibana-intake'),
               'x-pack-intake-agent': kibanaPipeline.legacyJobRunner('x-pack-intake'),
@@ -37,6 +38,8 @@ stage("Kibana Pipeline") { // This stage is just here to help the BlueOcean UI a
             ])
           }
         }
+
+        retryable.printFlakyFailures()
         kibanaPipeline.sendMail()
       }
     }

vars/githubPr.groovy

@@ -88,6 +88,8 @@ def getHistoryText(builds) {
     .collect { build ->
       if (build.status == "SUCCESS") {
         return "* :green_heart: [Build #${build.number}](${build.url}) succeeded ${build.commit}"
+      } else if(build.status == "UNSTABLE") {
+        return "* :yellow_heart: [Build #${build.number}](${build.url}) was flaky ${build.commit}"
       } else {
         return "* :broken_heart: [Build #${build.number}](${build.url}) failed ${build.commit}"
       }
@@ -97,18 +99,66 @@ def getHistoryText(builds) {
   return "### History\n${list}"
 }
 
-def getNextCommentMessage(previousCommentInfo = [:]) {
-  info = previousCommentInfo ?: [:]
-  info.builds = previousCommentInfo.builds ?: []
+def getTestFailuresMessage() {
+  def failures = testUtils.getFailures()
+  if (!failures) {
+    return ""
+  }
 
   def messages = []
 
-  if (buildUtils.getBuildStatus() == 'SUCCESS') {
+  failures.take(5).each { failure ->
+    messages << """
+---
+
+### [Test Failures](${env.BUILD_URL}testReport)
+<details><summary>${failure.fullDisplayName}</summary>
+
+[Link to Jenkins](${failure.url})
+
+```
+${failure.stdOut}
+```
+</details>
+
+---
+    """
+  }
+
+  if (failures.size() > 3) {
+    messages << "and ${failures.size() - 3} more failures, only showing the first 3."
+  }
+
+  return messages.join("\n")
+}
+
+def getNextCommentMessage(previousCommentInfo = [:]) {
+  def info = previousCommentInfo ?: [:]
+  info.builds = previousCommentInfo.builds ?: []
+
+  def messages = []
+  def status = buildUtils.getBuildStatus()
+
+  if (status == 'SUCCESS') {
     messages << """
       ## :green_heart: Build Succeeded
       * [continuous-integration/kibana-ci/pull-request](${env.BUILD_URL})
       * Commit: ${getCommitHash()}
     """
+  } else if(status == 'UNSTABLE') {
+    def message = """
+      ## :yellow_heart: Build succeeded, but was flaky
+      * [continuous-integration/kibana-ci/pull-request](${env.BUILD_URL})
+      * Commit: ${getCommitHash()}
+    """.stripIndent()
+
+    def failures = retryable.getFlakyFailures()
+    if (failures && failures.size() > 0) {
+      def list = failures.collect { "  * ${it.label}" }.join("\n")
+      message += "* Flaky suites:\n${list}"
+    }
+
+    messages << message
   } else {
     messages << """
       ## :broken_heart: Build Failed
@@ -117,6 +167,8 @@ def getNextCommentMessage(previousCommentInfo = [:]) {
     """
   }
 
+  messages << getTestFailuresMessage()
+
   if (info.builds && info.builds.size() > 0) {
     messages << getHistoryText(info.builds)
   }
@@ -133,7 +185,7 @@ def getNextCommentMessage(previousCommentInfo = [:]) {
 
   return messages
     .findAll { !!it } // No blank strings
-    .collect { it.stripIndent().trim() }
+    .collect { it.stripIndent().trim() } // This just allows us to indent various strings above, but leaves them un-indented in the comment
     .join("\n\n")
 }
 

vars/kibanaPipeline.groovy

@@ -72,7 +72,9 @@ def getOssCiGroupWorker(ciGroup) {
       "CI_GROUP=${ciGroup}",
       "JOB=kibana-ciGroup${ciGroup}",
     ]) {
-      runbld("./test/scripts/jenkins_ci_group.sh", "Execute kibana-ciGroup${ciGroup}")
+      retryable("kibana-ciGroup${ciGroup}") {
+        runbld("./test/scripts/jenkins_ci_group.sh", "Execute kibana-ciGroup${ciGroup}")
+      }
     }
   })
 }
@@ -83,7 +85,9 @@ def getXpackCiGroupWorker(ciGroup) {
       "CI_GROUP=${ciGroup}",
       "JOB=xpack-kibana-ciGroup${ciGroup}",
     ]) {
-      runbld("./test/scripts/jenkins_xpack_ci_group.sh", "Execute xpack-kibana-ciGroup${ciGroup}")
+      retryable("xpack-kibana-ciGroup${ciGroup}") {
+        runbld("./test/scripts/jenkins_xpack_ci_group.sh", "Execute xpack-kibana-ciGroup${ciGroup}")
+      }
     }
   })
 }

vars/retryable.groovy

@@ -0,0 +1,75 @@
+import groovy.transform.Field
+
+public static @Field GLOBAL_RETRIES_ENABLED = false
+public static @Field MAX_GLOBAL_RETRIES = 1
+public static @Field CURRENT_GLOBAL_RETRIES = 0
+public static @Field FLAKY_FAILURES = []
+
+def setMax(max) {
+  retryable.MAX_GLOBAL_RETRIES = max
+}
+
+def enable() {
+  retryable.GLOBAL_RETRIES_ENABLED = true
+}
+
+def enable(max) {
+  enable()
+  setMax(max)
+}
+
+def haveReachedMaxRetries() {
+  return retryable.CURRENT_GLOBAL_RETRIES >= retryable.MAX_GLOBAL_RETRIES
+}
+
+def getFlakyFailures() {
+  return retryable.FLAKY_FAILURES
+}
+
+def printFlakyFailures() {
+  catchError {
+    def failures = getFlakyFailures()
+
+    if (failures && failures.size() > 0) {
+      print "This build had the following flaky failures:"
+      failures.each {
+        print "\n${it.label}"
+        buildUtils.printStacktrace(it.exception)
+      }
+    }
+  }
+}
+
+def call(label, Closure closure) {
+  if (!retryable.GLOBAL_RETRIES_ENABLED) {
+    closure()
+    return
+  }
+
+  try {
+    closure()
+  } catch (ex) {
+    if (haveReachedMaxRetries()) {
+      print "Couldn't retry '${label}', have already reached the max number of retries for this build."
+      throw ex
+    }
+
+    retryable.CURRENT_GLOBAL_RETRIES++
+    buildUtils.printStacktrace(ex)
+    unstable "${label} failed but is retryable, trying a second time..."
+
+    def JOB = env.JOB ? "${env.JOB}-retry" : ""
+    withEnv([
+      "JOB=${JOB}",
+    ]) {
+      closure()
+    }
+
+    retryable.FLAKY_FAILURES << [
+      label: label,
+      exception: ex,
+    ]
+
+    unstable "${label} failed on the first attempt, but succeeded on the second. Marking it as flaky."
+  }
+}