fix(spark): propagate exit code of the entrypoint (#1595)

* fix(spark): propagate exit code of the entrypoint

* chore: add changelog entry

* fix(spark): rewrite run-spark.sh to fit into existing script pattern

Author: sweb

CHANGELOG.md

@@ -4,11 +4,16 @@ All notable changes to this project will be documented in this file.
 
 ## [Unreleased]
 
+### Fixed
+
+- spark: Propagate the entrypoint's exit code so failed applications are no longer reported as successful ([#1595]).
+
 ### Removed
 
 - omid: remove 1.1.2 ([#1593]).
 
 [#1593]: https://github.com/stackabletech/docker-images/pull/1593
+[#1595]: https://github.com/stackabletech/docker-images/pull/1595
 
 ## [26.7.0] - 2026-07-21
 

spark-k8s/stackable/run-spark.sh

@@ -1,24 +1,52 @@
 #!/bin/bash
 
-eval "$_STACKABLE_PRE_HOOK"
-
-# Forward SIGTERM to the Spark entrypoint to support spark JVM's gracefully shutdown.
-/stackable/spark/kubernetes/dockerfiles/spark/entrypoint.sh "$@" &
-child_pid=$!
+prepare_signal_handlers() {
+  unset term_child_pid
+  unset term_kill_needed
+  trap handle_term_signal TERM INT
+}
 
+# Only ever invoked via the trap installed above.
 # shellcheck disable=SC2329
-_handle_term() {
-  kill -TERM "$child_pid" 2>/dev/null || true
+handle_term_signal() {
+  if [[ -v term_child_pid ]]; then
+    kill -TERM "${term_child_pid}" 2>/dev/null || true
+  else
+    term_kill_needed="yes"
+  fi
+}
+
+# Waits for the given child and returns *its* exit status.
+# A trapped signal makes `wait` return immediately with a status > 128 while the child keeps
+# running its shutdown sequence, so in that case wait again instead of taking the interrupted
+# status.
+wait_for_termination() {
+  term_child_pid=$1
+  if [[ -v term_kill_needed ]]; then
+    kill -TERM "${term_child_pid}" 2>/dev/null || true
+  fi
+  while true; do
+    wait "${term_child_pid}"
+    term_child_status=$?
+    if [[ "${term_child_status}" -gt 128 ]] && kill -0 "${term_child_pid}" 2>/dev/null; then
+      continue
+    fi
+    return "${term_child_status}"
+  done
 }
-trap _handle_term TERM INT
 
-# `wait` returns immediately when the trap fires; loop until the child is actually gone.
-wait "$child_pid"
-while kill -0 "$child_pid" 2>/dev/null; do
-  wait "$child_pid" 2>/dev/null || true
-done
-result=$?
+eval "$_STACKABLE_PRE_HOOK"
+
+# The signal handlers are installed before the child is started, so that a signal arriving in
+# between is not lost. SIGTERM is forwarded to the Spark entrypoint to let the Spark JVM shut
+# down gracefully.
+prepare_signal_handlers
+
+/stackable/spark/kubernetes/dockerfiles/spark/entrypoint.sh "$@" &
+
+result=0
+wait_for_termination $! || result=$?
 
 eval "$_STACKABLE_POST_HOOK"
 
-exit $result
+exit "${result}"