diff --git a/python/ray/util/spark/cluster_init.py b/python/ray/util/spark/cluster_init.py
index 129f1677b6b8..ee0bcadf1ecc 100644
--- a/python/ray/util/spark/cluster_init.py
+++ b/python/ray/util/spark/cluster_init.py
@@ -703,6 +703,7 @@ def _setup_ray_cluster(
             extra_env={
                 RAY_ON_SPARK_COLLECT_LOG_TO_PATH: collect_log_to_path or "",
                 RAY_ON_SPARK_START_RAY_PARENT_PID: str(os.getpid()),
+                **start_hook.custom_environment_variables(),
             },
         )
         spark_job_server = None
@@ -1511,10 +1512,13 @@ def ray_cluster_job_mapper(_):
         if ray_temp_dir is not None:
             ray_worker_node_cmd.append(f"--temp-dir={ray_temp_dir}")
 
+        hook_entry = _create_hook_entry(is_global=(ray_temp_dir is None))
+
         ray_worker_node_extra_envs = {
             RAY_ON_SPARK_COLLECT_LOG_TO_PATH: collect_log_to_path or "",
             RAY_ON_SPARK_START_RAY_PARENT_PID: str(os.getpid()),
             "RAY_ENABLE_WINDOWS_OR_OSX_CLUSTER": "1",
+            **hook_entry.custom_environment_variables(),
         }
 
         if num_gpus_per_node > 0:
@@ -1780,10 +1784,13 @@ def start(
         )
         ray_head_node_cmd.extend(_convert_ray_node_options(head_node_options))
 
+        hook_entry = _create_hook_entry(is_global=(ray_temp_dir is None))
+
         extra_env = {
             "AUTOSCALER_UPDATE_INTERVAL_S": "1",
             RAY_ON_SPARK_COLLECT_LOG_TO_PATH: collect_log_to_path or "",
             RAY_ON_SPARK_START_RAY_PARENT_PID: str(os.getpid()),
+            **hook_entry.custom_environment_variables(),
         }
 
         self.ray_head_node_cmd = ray_head_node_cmd
diff --git a/python/ray/util/spark/databricks_hook.py b/python/ray/util/spark/databricks_hook.py
index 750ffb13d9b4..97aa120ac7a0 100644
--- a/python/ray/util/spark/databricks_hook.py
+++ b/python/ray/util/spark/databricks_hook.py
@@ -167,3 +167,13 @@ def auto_shutdown_watcher():
     def on_spark_job_created(self, job_group_id):
         db_api_entry = get_db_entry_point()
         db_api_entry.registerBackgroundSparkJobGroup("job_group_id")
+
+    def custom_environment_variables(self):
+        """Hardcode `GLOO_SOCKET_IFNAME` to `eth0` for Databricks runtime.
+
+        Torch on DBR does not reliably detect the correct interface to use,
+        and ends up selecting the loopback interface, breaking cross-node
+        commnication."""
+        return {
+            "GLOO_SOCKET_IFNAME": "eth0",
+        }
diff --git a/python/ray/util/spark/start_hook_base.py b/python/ray/util/spark/start_hook_base.py
index a45c216d6659..d51dbec3a02b 100644
--- a/python/ray/util/spark/start_hook_base.py
+++ b/python/ray/util/spark/start_hook_base.py
@@ -13,3 +13,6 @@ def on_cluster_created(self, ray_cluster_handler):
 
     def on_spark_job_created(self, job_group_id):
         pass
+
+    def custom_environment_variables(self):
+        return {}