fixing final SPDK deployment

serving/docker-compose.yml

@@ -285,13 +285,18 @@ services:
       TARGET_PORT: ${TARGET_PORT}
       SPDK_CPUMASK: ${SPDK_CPUMASK}
       NVMEOF_SVC: nvmeof
-    # Healthy once the RPC socket answers — gates the spdk-loader's startup.
+    # Healthy once spdk-target has written its readiness stamp (transport up +
+    # releases reconciled). A FILE check, not a python rpc.py probe: nvmf_tgt does
+    # a slow DPDK init (~50s observed) and busy-polls its reactor core, so a python
+    # healthcheck starves/times out and the container sticks at "unhealthy" even
+    # though the target is up. `test -f` is instant and contention-free. The stamp
+    # is removed on (re)start + exit, so it never falsely reports a dead target.
     healthcheck:
-      test: ["CMD-SHELL", "rpc.py -s /var/run/spdk/spdk.sock framework_get_config >/dev/null 2>&1 || exit 1"]
-      interval: 5s
-      timeout: 3s
-      retries: 24
-      start_period: 10s
+      test: ["CMD-SHELL", "test -f /var/run/spdk/ready || exit 1"]
+      interval: 10s
+      timeout: 5s
+      retries: 6
+      start_period: 120s
     # Persistent target: started ONCE, never restarted for a version change.
     # Rolling updates add/remove versions online (spdk-loader, or spdk-ctl.sh).
     entrypoint: ["/bin/bash", "/usr/local/bin/spdk-target"]