fix(tightly-coupled): pin k8s step IO manager to shared storage

src/distributed_execution/tightly_coupled/jobs.py

@@ -12,7 +12,7 @@ is the instance-level run launcher (see ``yaml/tightly-coupled/``).
 
 from __future__ import annotations
 
-from dagster import graph, in_process_executor, multiprocess_executor
+from dagster import fs_io_manager, graph, in_process_executor, multiprocess_executor
 from dagster_k8s import k8s_job_executor
 
 from distributed_execution.ops import (
@@ -28,15 +28,14 @@ COMMON_TAGS = {
     "business_operation": "DISTRIBUTED_EXECUTION_REFERENCE",
 }
 
+# RWX PVC mounted into the run worker and every step pod. Set as the I/O manager
+# base_dir rather than via DAGSTER_HOME: the Dagster chart already injects
+# DAGSTER_HOME=/tmp/dagster and the earlier duplicate wins in the step pod.
+SHARED_IO_BASE_DIR = "/dagster/shared/distributed-execution/storage"
+
 # Per-step pod shape. Only honoured by k8s_job_executor.
 STEP_K8S_CONFIG = {
     "container_config": {
-        "env": [
-            {
-                "name": "DAGSTER_HOME",
-                "value": "/dagster/shared/distributed-execution",
-            }
-        ],
         "resources": {
             "requests": {"cpu": "100m", "memory": "128Mi"},
             "limits": {"cpu": "500m", "memory": "512Mi"},
@@ -104,6 +103,8 @@ tightly_coupled_k8s_job = distributed_execution_reference.to_job(
     # config `execution:` block is accepted and then silently discarded - which
     # also makes job_image and job_namespace impossible to supply at launch time.
     executor_def=k8s_job_executor,
+    # Steps run in separate pods, so step outputs have to land on shared storage.
+    resource_defs={"io_manager": fs_io_manager.configured({"base_dir": SHARED_IO_BASE_DIR})},
     config={
         "execution": {
             "config": {