diff --git a/tests/templates/kuttl/graceful-shutdown/10-assert.yaml b/tests/templates/kuttl/graceful-shutdown/10-assert.yaml
index f30e7b3f..abaa736c 100644
--- a/tests/templates/kuttl/graceful-shutdown/10-assert.yaml
+++ b/tests/templates/kuttl/graceful-shutdown/10-assert.yaml
@@ -5,15 +5,50 @@ timeout: 900
 commands:
   # Both executors have to be registered and working: an executor that is still registering is
   # not in the driver's list yet and would never be asked to shut down.
-  - script: |
+  #
+  # Pods that are terminating are still listed by `kubectl get pods` and keep the log lines below
+  # for as long as they linger, so they would satisfy this check without being able to do any
+  # work. Only executors that are alive are counted.
+  #
+  # The polling happens here and not by letting kuttl retry the assert, because kuttl reprints the
+  # whole script on every attempt and would bury the rest of the test log.
+  - timeout: 900
+    script: |
       set -eu
 
-      registered=0
-      for pod in $(kubectl -n "$NAMESPACE" get pods -o name \
-        -l spark-role=executor,app.kubernetes.io/instance=graceful-shutdown | cut -d/ -f2); do
-        log=$(kubectl -n "$NAMESPACE" logs --tail=-1 "$pod" -c spark)
-        echo "$log" | grep -q 'Successfully registered with driver'
-        echo "$log" | grep -q 'Running task'
-        registered=$(( registered + 1 ))
+      selector="spark-role=executor,app.kubernetes.io/instance=graceful-shutdown"
+      budget_seconds=870
+      poll_seconds=5
+
+      start=$(date +%s)
+      while :; do
+        executors=$(
+          kubectl -n "$NAMESPACE" get pods -l "$selector" -o json | jq -r '
+            .items[]
+            | select(.metadata.deletionTimestamp == null and .status.phase == "Running")
+            | .metadata.name'
+        )
+
+        registered=0
+        for pod in $executors; do
+          # The Pod can disappear between the listing and the read, which is not a failure.
+          logs=$(kubectl -n "$NAMESPACE" logs --tail=-1 "$pod" -c spark 2>/dev/null || true)
+          echo "$logs" | grep -q 'Successfully registered with driver' || continue
+          echo "$logs" | grep -q 'Running task' || continue
+          registered=$(( registered + 1 ))
+        done
+
+        if [ "$registered" -eq 2 ]; then
+          echo "OK: 2 executors are registered with the driver and running a task"
+          exit 0
+        fi
+
+        if [ $(( $(date +%s) - start )) -ge "$budget_seconds" ]; then
+          echo "FAIL: ${budget_seconds}s elapsed with $registered of 2 executors registered and"
+          echo "running a task, so the driver's list is not what the rest of the test assumes"
+          kubectl -n "$NAMESPACE" get pods -l "$selector"
+          exit 1
+        fi
+
+        sleep "$poll_seconds"
       done
-      test "$registered" -eq 2
diff --git a/tests/templates/kuttl/graceful-shutdown/12-assert.yaml b/tests/templates/kuttl/graceful-shutdown/12-assert.yaml
index bf511321..7533aa2d 100644
--- a/tests/templates/kuttl/graceful-shutdown/12-assert.yaml
+++ b/tests/templates/kuttl/graceful-shutdown/12-assert.yaml
@@ -3,17 +3,45 @@ apiVersion: kuttl.dev/v1beta1
 kind: TestAssert
 timeout: 600
 commands:
-  # Same check as in step 10: the replacement for the stopped executor has to be registered and
-  # working before the driver is stopped, otherwise it is not in the driver's list yet.
-  - script: |
+  # The polling happens here and not by letting kuttl retry the assert, because kuttl reprints the
+  # whole script on every attempt and would bury the rest of the test log.
+  - timeout: 600
+    script: |
       set -eu
 
-      registered=0
-      for pod in $(kubectl -n "$NAMESPACE" get pods -o name \
-        -l spark-role=executor,app.kubernetes.io/instance=graceful-shutdown | cut -d/ -f2); do
-        log=$(kubectl -n "$NAMESPACE" logs --tail=-1 "$pod" -c spark)
-        echo "$log" | grep -q 'Successfully registered with driver'
-        echo "$log" | grep -q 'Running task'
-        registered=$(( registered + 1 ))
+      selector="spark-role=executor,app.kubernetes.io/instance=graceful-shutdown"
+      budget_seconds=570
+      poll_seconds=5
+
+      start=$(date +%s)
+      while :; do
+        executors=$(
+          kubectl -n "$NAMESPACE" get pods -l "$selector" -o json | jq -r '
+            .items[]
+            | select(.metadata.deletionTimestamp == null and .status.phase == "Running")
+            | .metadata.name'
+        )
+
+        registered=0
+        for pod in $executors; do
+          # The Pod can disappear between the listing and the read, which is not a failure.
+          logs=$(kubectl -n "$NAMESPACE" logs --tail=-1 "$pod" -c spark 2>/dev/null || true)
+          echo "$logs" | grep -q 'Successfully registered with driver' || continue
+          echo "$logs" | grep -q 'Running task' || continue
+          registered=$(( registered + 1 ))
+        done
+
+        if [ "$registered" -eq 2 ]; then
+          echo "OK: the replacement executor is registered with the driver and running a task"
+          exit 0
+        fi
+
+        if [ $(( $(date +%s) - start )) -ge "$budget_seconds" ]; then
+          echo "FAIL: ${budget_seconds}s after the executor was stopped, $registered of 2"
+          echo "executors are registered and running a task, so no replacement came back"
+          kubectl -n "$NAMESPACE" get pods -l "$selector"
+          exit 1
+        fi
+
+        sleep "$poll_seconds"
       done
-      test "$registered" -eq 2
diff --git a/tests/templates/kuttl/graceful-shutdown/13-check-driver-shutdown-propagation.yaml b/tests/templates/kuttl/graceful-shutdown/13-check-driver-shutdown-propagation.yaml
index 683d94e5..4de4738e 100644
--- a/tests/templates/kuttl/graceful-shutdown/13-check-driver-shutdown-propagation.yaml
+++ b/tests/templates/kuttl/graceful-shutdown/13-check-driver-shutdown-propagation.yaml
@@ -17,8 +17,19 @@ commands:
       capture=$(mktemp -d)
       driver=$(kubectl -n "$NAMESPACE" get pods -o name \
         -l "spark-role=driver,app.kubernetes.io/instance=$app" | head -1 | cut -d/ -f2)
-      executors=$(kubectl -n "$NAMESPACE" get pods -o name \
-        -l "spark-role=executor,app.kubernetes.io/instance=$app" | cut -d/ -f2)
+      # Terminating Pods are still listed by `kubectl get pods`, so an executor left over from
+      # step 11 would be waited on for a shutdown it has already performed and can never log
+      # again. Only executors that are alive can still be asked to stop.
+      live=$(
+        kubectl -n "$NAMESPACE" get pods \
+          -l "spark-role=executor,app.kubernetes.io/instance=$app" \
+          -o json | jq '
+            [ .items[]
+              | select(.metadata.deletionTimestamp == null and .status.phase == "Running")
+              | .metadata.name ]'
+      )
+      executors=$(echo "$live" | jq -r '.[]')
+      running=$(echo "$live" | jq 'length')
 
       # Without these the loop below has nothing to wait for and would report success while having
       # checked nothing at all.
@@ -26,8 +37,10 @@ commands:
         echo "FAIL: no driver Pod found for $app"
         exit 1
       fi
-      if [ -z "$executors" ]; then
-        echo "FAIL: no executor Pods found for $app, so no shutdown can be propagated to any"
+      if [ "$running" -ne 2 ]; then
+        echo "FAIL: expected 2 running executor Pods for $app, found $running, so no shutdown can"
+        echo "be propagated to the replicas the application asked for"
+        kubectl -n "$NAMESPACE" get pods -l "spark-role=executor,app.kubernetes.io/instance=$app"
         exit 1
       fi