From 276a9b4c974099a4825d92abb61c9db92e722f2b Mon Sep 17 00:00:00 2001
From: wei <wei@xiaobuou.local>
Date: Sun, 26 Jul 2026 13:43:26 +0800
Subject: [PATCH] fix(etcd): fail role probe on status errors

---
 addons/etcd/scripts-ut-spec/roleprobe_spec.sh | 285 +++++++++++-------
 addons/etcd/scripts/roleprobe.sh              |  64 +++-
 2 files changed, 237 insertions(+), 112 deletions(-)

diff --git a/addons/etcd/scripts-ut-spec/roleprobe_spec.sh b/addons/etcd/scripts-ut-spec/roleprobe_spec.sh
index 03174efb04..fc49a88ceb 100644
--- a/addons/etcd/scripts-ut-spec/roleprobe_spec.sh
+++ b/addons/etcd/scripts-ut-spec/roleprobe_spec.sh
@@ -1,132 +1,215 @@
 # shellcheck shell=bash
-# shellcheck disable=SC2034
+# shellcheck disable=SC2034,SC2329
 
-# validate_shell_type_and_version defined in shellspec/spec_helper.sh used to validate the expected shell type and version this script needs to run.
 if ! validate_shell_type_and_version "bash" 4 &>/dev/null; then
   echo "roleprobe_spec.sh skip cases because dependency bash version 4 or higher is not installed."
   exit 0
 fi
 
-source ./utils.sh
-
-# The unit test needs to rely on the common library functions defined in kblib.
-# Therefore, we first dynamically generate the required common library files from the kblib library chart.
-common_library_file="./common.sh"
-generate_common_library $common_library_file
-
 Describe "Etcd Role Probe Script Tests"
-  # load the scripts to be tested and dependencies
-  Include $common_library_file
+  Include ../scripts/roleprobe.sh
+
+  setup() {
+    TEST_DIR=$(mktemp -d)
+    export TEST_DIR
+    export ROLEPROBE_ENTRYPOINT_DIR="$TEST_DIR/scripts"
+    export ROLEPROBE_ETCDCTL_RC=0
+    export ROLEPROBE_ETCDCTL_STDOUT=
+    export ROLEPROBE_ETCDCTL_STDERR=
+
+    mkdir -p "$ROLEPROBE_ENTRYPOINT_DIR"
+    cp ../scripts/roleprobe.sh "$ROLEPROBE_ENTRYPOINT_DIR/roleprobe.sh"
+    cat > "$ROLEPROBE_ENTRYPOINT_DIR/common.sh" <<'COMMON'
+setup_shellspec() { :; }
+load_common_library() { :; }
+exec_etcdctl() {
+  if [ "$*" != "127.0.0.1:2379 endpoint status -w fields --command-timeout=300ms --dial-timeout=100ms" ]; then
+    printf 'unexpected etcdctl arguments: %s\n' "$*" >&2
+    return 64
+  fi
+  printf '%s' "${ROLEPROBE_ETCDCTL_STDOUT:-}"
+  printf '%s' "${ROLEPROBE_ETCDCTL_STDERR:-}" >&2
+  return "${ROLEPROBE_ETCDCTL_RC:-0}"
+}
+COMMON
 
-  init() {
-    # set ut_mode to true to hack control flow in the script
-    ut_mode="true"
-
-    # Mock exec_etcdctl function
     exec_etcdctl() {
-      local endpoint="$1"
-      shift
-      case "$*" in
-        *"endpoint status -w fields"*)
-          echo '"MemberID" : 1002'
-          echo '"Leader" : 1002'
-          echo '"IsLearner" : false'
-          return 0
-          ;;
-        *)
-          echo "MOCK: exec_etcdctl $endpoint $*"
-          return 0
-          ;;
-      esac
-    }
-
-    # Define get_etcd_role function based on real script logic
-    get_etcd_role() {
-      local status member_id leader_id is_learner
-
-      if ! status=$(exec_etcdctl 127.0.0.1:2379 endpoint status -w fields --command-timeout=300ms --dial-timeout=100ms); then
-        echo "ERROR: Failed to get endpoint status" >&2
-        return 1
-      fi
-
-      member_id=$(echo "$status" | grep -o '"MemberID" : [0-9]*' | awk '{print $3}')
-      leader_id=$(echo "$status" | grep -o '"Leader" : [0-9]*' | awk '{print $3}')
-      is_learner=$(echo "$status" | grep -o '"IsLearner" : [a-z]*' | awk '{print $3}')
-
-      # Check if required fields are present
-      if [ -z "$member_id" ] || [ -z "$leader_id" ]; then
-        echo "follower"  # Default to follower when fields are missing
-        return 0
-      fi
-
-      if [ "$member_id" = "$leader_id" ]; then
-        echo "leader"
-      elif [ "$is_learner" = "true" ]; then
-        echo "learner"
-      else
-        echo "follower"
-      fi
+      printf '%s' "${ROLEPROBE_ETCDCTL_STDOUT:-}"
+      printf '%s' "${ROLEPROBE_ETCDCTL_STDERR:-}" >&2
+      return "${ROLEPROBE_ETCDCTL_RC:-0}"
     }
   }
-  BeforeAll "init"
+  BeforeEach "setup"
 
   cleanup() {
-    rm -f $common_library_file
-    unset ut_mode
-    unset -f exec_etcdctl get_etcd_role
+    rm -rf "$TEST_DIR"
+    unset TEST_DIR ROLEPROBE_ENTRYPOINT_DIR ROLEPROBE_ETCDCTL_RC
+    unset ROLEPROBE_ETCDCTL_STDOUT ROLEPROBE_ETCDCTL_STDERR
+    unset -f exec_etcdctl
+  }
+  AfterEach "cleanup"
+
+  endpoint_status() {
+    printf '"MemberID" : %s\n"Leader" : %s\n"IsLearner" : %s\n' "$1" "$2" "$3"
+  }
+
+  run_roleprobe_entrypoint() {
+    cmp -s ../scripts/roleprobe.sh "$ROLEPROBE_ENTRYPOINT_DIR/roleprobe.sh" ||
+      return 99
+    bash "$ROLEPROBE_ENTRYPOINT_DIR/roleprobe.sh"
+  }
+
+  assert_role_failure() {
+    ROLEPROBE_ETCDCTL_STDOUT="$1"
+    export ROLEPROBE_ETCDCTL_STDOUT
+    roleprobe_main
   }
-  AfterAll 'cleanup'
 
-  Describe "get_etcd_role() function"
-    It "detects leader role correctly"
+  Describe "production get_etcd_role()"
+    It "detects an official-shape leader"
+      ROLEPROBE_ETCDCTL_STDOUT=$'"Endpoint" : "127.0.0.1:2379"\n"ClusterID" : 9\n'
+      ROLEPROBE_ETCDCTL_STDOUT+="$(endpoint_status 1002 1002 false)"
       When call get_etcd_role
       The status should be success
-      The output should equal "leader"
-    End
-
-    It "detects follower role correctly"
-      # Override exec_etcdctl to return follower status
-      exec_etcdctl() {
-        case "$*" in
-          *"endpoint status -w fields"*)
-            echo '"MemberID" : 1001'
-            echo '"Leader" : 1002'
-            echo '"IsLearner" : false'
-            return 0
-            ;;
-        esac
-      }
+      The output should eq "leader"
+    End
 
+    It "detects an official-shape follower"
+      ROLEPROBE_ETCDCTL_STDOUT=$(endpoint_status 1001 1002 false)
       When call get_etcd_role
       The status should be success
-      The output should equal "follower"
-    End
-
-    It "detects learner role correctly"
-      # Override exec_etcdctl to return learner status
-      exec_etcdctl() {
-        case "$*" in
-          *"endpoint status -w fields"*)
-            echo '"MemberID" : 1003'
-            echo '"Leader" : 1002'
-            echo '"IsLearner" : true'
-            return 0
-            ;;
-        esac
-      }
+      The output should eq "follower"
+    End
 
+    It "detects an official-shape learner"
+      ROLEPROBE_ETCDCTL_STDOUT=$(endpoint_status 1003 1002 true)
       When call get_etcd_role
       The status should be success
-      The output should equal "learner"
+      The output should eq "learner"
     End
 
-    It "handles etcdctl failure"
-      # Override exec_etcdctl to fail
-      exec_etcdctl() { return 1; }
-
+    It "propagates an etcdctl failure"
+      ROLEPROBE_ETCDCTL_RC=23
+      ROLEPROBE_ETCDCTL_STDERR="controlled etcdctl failure"
       When call get_etcd_role
       The status should be failure
+      The output should eq ""
+      The stderr should include "controlled etcdctl failure"
       The stderr should include "Failed to get endpoint status"
     End
+
+    It "rejects a missing MemberID"
+      When call assert_role_failure $'"Leader" : 1002\n"IsLearner" : false'
+      The status should be failure
+      The output should eq ""
+      The stderr should include "Failed to extract role fields"
+    End
+
+    It "rejects a missing Leader"
+      When call assert_role_failure $'"MemberID" : 1002\n"IsLearner" : false'
+      The status should be failure
+      The output should eq ""
+      The stderr should include "Failed to extract role fields"
+    End
+
+    It "rejects a missing IsLearner"
+      When call assert_role_failure $'"MemberID" : 1002\n"Leader" : 1002'
+      The status should be failure
+      The output should eq ""
+      The stderr should include "Failed to extract role fields"
+    End
+
+    It "rejects a MemberID suffix"
+      When call assert_role_failure "$(endpoint_status 1002garbage 1002 false)"
+      The status should be failure
+      The output should eq ""
+      The stderr should include "Failed to extract role fields"
+    End
+
+    It "rejects a Leader suffix"
+      When call assert_role_failure "$(endpoint_status 1002 1002garbage false)"
+      The status should be failure
+      The output should eq ""
+      The stderr should include "Failed to extract role fields"
+    End
+
+    It "rejects an invalid learner boolean"
+      When call assert_role_failure "$(endpoint_status 1002 1002 falsegarbage)"
+      The status should be failure
+      The output should eq ""
+      The stderr should include "Failed to extract role fields"
+    End
+
+    It "rejects duplicate MemberID fields"
+      When call assert_role_failure $'"MemberID" : 1002\n"MemberID" : 1003\n"Leader" : 1002\n"IsLearner" : false'
+      The status should be failure
+      The output should eq ""
+      The stderr should include "Failed to extract role fields"
+    End
+
+    It "rejects a repeated identical Leader field"
+      When call assert_role_failure $'"MemberID" : 1002\n"Leader" : 1002\n"Leader" : 1002\n"IsLearner" : false'
+      The status should be failure
+      The output should eq ""
+      The stderr should include "Failed to extract role fields"
+    End
+
+    It "rejects conflicting duplicate Leader fields"
+      When call assert_role_failure $'"MemberID" : 1002\n"Leader" : 1002\n"Leader" : 1003\n"IsLearner" : false'
+      The status should be failure
+      The output should eq ""
+      The stderr should include "Failed to extract role fields"
+    End
+
+    It "rejects conflicting duplicate IsLearner fields"
+      When call assert_role_failure $'"MemberID" : 1002\n"Leader" : 1002\n"IsLearner" : false\n"IsLearner" : true'
+      The status should be failure
+      The output should eq ""
+      The stderr should include "Failed to extract role fields"
+    End
+  End
+
+  Describe "production roleprobe entrypoint"
+    It "emits the classified role"
+      ROLEPROBE_ETCDCTL_STDOUT=$(endpoint_status 1002 1002 false)
+      When call run_roleprobe_entrypoint
+      The status should be success
+      The output should eq "leader"
+      The stderr should eq ""
+    End
+
+    It "emits follower from the official field shape"
+      ROLEPROBE_ETCDCTL_STDOUT=$(endpoint_status 1001 1002 false)
+      When call run_roleprobe_entrypoint
+      The status should be success
+      The output should eq "follower"
+      The stderr should eq ""
+    End
+
+    It "emits learner from the official field shape"
+      ROLEPROBE_ETCDCTL_STDOUT=$(endpoint_status 1003 1002 true)
+      When call run_roleprobe_entrypoint
+      The status should be success
+      The output should eq "learner"
+      The stderr should eq ""
+    End
+
+    It "propagates etcdctl rc and keeps stdout empty"
+      ROLEPROBE_ETCDCTL_RC=23
+      ROLEPROBE_ETCDCTL_STDERR="controlled entrypoint failure"
+      When call run_roleprobe_entrypoint
+      The status should be failure
+      The output should eq ""
+      The stderr should include "controlled entrypoint failure"
+      The stderr should include "Failed to get endpoint status"
+    End
+
+    It "fails closed on malformed rc0 status"
+      ROLEPROBE_ETCDCTL_STDOUT=$'"MemberID" : 1002garbage\n"Leader" : 1002\n"IsLearner" : false'
+      When call run_roleprobe_entrypoint
+      The status should be failure
+      The output should eq ""
+      The stderr should include "Failed to extract role fields"
+    End
   End
-End
\ No newline at end of file
+End
diff --git a/addons/etcd/scripts/roleprobe.sh b/addons/etcd/scripts/roleprobe.sh
index bc9b069667..6afb53f9f7 100644
--- a/addons/etcd/scripts/roleprobe.sh
+++ b/addons/etcd/scripts/roleprobe.sh
@@ -1,7 +1,41 @@
 #!/bin/bash
 
-# shellcheck disable=SC1091
-. "/scripts/common.sh"
+# shellcheck disable=SC1090,SC1091
+roleprobe_script_dir=$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)
+. "${roleprobe_script_dir}/common.sh"
+
+parse_role_field() {
+  local status="$1"
+  local field_name="$2"
+  local value_type="$3"
+
+  printf '%s\n' "$status" | awk -v field_name="$field_name" -v value_type="$value_type" '
+    BEGIN {
+      key = "^[[:space:]]*\"" field_name "\"[[:space:]]*:"
+      if (value_type == "id") {
+        value = "[0-9]+"
+      } else {
+        value = "(true|false)"
+      }
+      valid = key "[[:space:]]*" value "[[:space:]]*$"
+    }
+    $0 ~ key {
+      seen++
+      if ($0 ~ valid) {
+        parsed = $0
+        sub(key "[[:space:]]*", "", parsed)
+        sub("[[:space:]]*$", "", parsed)
+        valid_count++
+      }
+    }
+    END {
+      if (seen != 1 || valid_count != 1) {
+        exit 1
+      }
+      print parsed
+    }
+  '
+}
 
 get_etcd_role() {
   local status member_id leader_id is_learner
@@ -10,9 +44,12 @@ get_etcd_role() {
     return 1
   fi
 
-  member_id=$(echo "$status" | grep -o '"MemberID" : [0-9]*' | awk '{print $3}')
-  leader_id=$(echo "$status" | grep -o '"Leader" : [0-9]*' | awk '{print $3}')
-  is_learner=$(echo "$status" | grep -o '"IsLearner" : [a-z]*' | awk '{print $3}')
+  if ! member_id=$(parse_role_field "$status" "MemberID" "id") ||
+    ! leader_id=$(parse_role_field "$status" "Leader" "id") ||
+    ! is_learner=$(parse_role_field "$status" "IsLearner" "bool"); then
+    echo "ERROR: Failed to extract role fields from endpoint status" >&2
+    return 1
+  fi
 
   if [ "$member_id" = "$leader_id" ]; then
     echo "leader"
@@ -23,10 +60,15 @@ get_etcd_role() {
   fi
 }
 
-# Shellspec magic
-setup_shellspec
+roleprobe_main() {
+  local etcd_role
+  if ! etcd_role=$(get_etcd_role); then
+    return 1
+  fi
+  printf '%s' "$etcd_role"
+}
 
-# main
-load_common_library
-etcd_role=$(get_etcd_role)
-echo -n "$etcd_role"
+if [[ "${BASH_SOURCE[0]}" == "$0" ]]; then
+  load_common_library
+  roleprobe_main "$@"
+fi