fix(apps): give new deployed endpoints unique template names

runpod/api/queries/apps.py

@@ -19,7 +19,7 @@
             name
             state
             activeBuildId
-            endpoints { id name }
+            endpoints { id name template { name } }
         }
     }
 }
@@ -60,7 +60,7 @@
         state
         activeBuildId
         createdAt
-        endpoints { id name }
+        endpoints { id name template { name } }
         networkVolumes { id name }
     }
 }

runpod/apps/deploy.py

@@ -3,6 +3,7 @@
 import json
 import logging
 import os
+import secrets
 import tarfile
 import tempfile
 from dataclasses import dataclass, field
@@ -29,6 +30,8 @@
 log = logging.getLogger(__name__)
 
 MANIFEST_VERSION = 1
+# hex bytes in the suffix that keeps new template names unique (6 chars)
+TEMPLATE_TOKEN_BYTES = 3
 
 DEFAULT_IGNORES = [
     ".venv/",
@@ -363,6 +366,22 @@ def _container_disk_gb(spec) -> int:
     return CPU_CONTAINER_DISK_GB if spec.is_cpu else GPU_CONTAINER_DISK_GB
 
 
+def _legacy_template_name(app: App, spec) -> str:
+    """the fixed template name used before names carried a suffix."""
+    return f"{app.name}-{spec.name}-template"
+
+
+def _new_template_name(app: App, spec, token: str) -> str:
+    """a template name for a newly created endpoint.
+
+    template names are unique per owner, and a deleted endpoint keeps
+    its template for a while before the platform removes it, so a fixed
+    name fails when an app is redeployed soon after being deleted, or
+    deployed to a second environment.
+    """
+    return f"{_legacy_template_name(app, spec)}-{token}"
+
+
 def _deployed_endpoint_input(
     app: App,
     spec,
@@ -372,9 +391,13 @@ def _deployed_endpoint_input(
     *,
     env_name: Optional[str] = None,
     pool_members: Optional[Dict[str, Set[str]]] = None,
+    template_name: Optional[str] = None,
 ) -> Dict[str, Any]:
     """saveEndpoint payload for one deployed queue/api resource.
 
+    template_name names the bound template; it defaults to the legacy
+    fixed name (see reconcile_endpoints for why new endpoints vary it).
+
     the endpoint name must equal the resource name: sentinel
     resolution matches X-Flash-Endpoint against endpoint.name within
     the environment. binding via flashEnvironmentId is also what makes
@@ -421,7 +444,7 @@ def _deployed_endpoint_input(
         "scalerValue": spec.scaler_value,
         "executionTimeoutMs": spec.execution_timeout_ms,
         "template": {
-            "name": f"{app.name}-{spec.name}-template",
+            "name": template_name or _legacy_template_name(app, spec),
             "imageName": image_for_spec(spec, python_version=python_version),
             "containerDiskInGb": _container_disk_gb(spec),
             "dockerArgs": "",
@@ -483,7 +506,14 @@ async def reconcile_endpoints(
     """
     for handle in app.resources.values():
         handle.spec.validate()
-    existing = {e["name"]: e["id"] for e in environment.get("endpoints") or []}
+    existing_endpoints = environment.get("endpoints") or []
+    existing = {e["name"]: e["id"] for e in existing_endpoints}
+    # updates send the bound template's current name, so redeploying an
+    # existing endpoint sends the same template name as before.
+    existing_templates = {
+        e["name"]: (e.get("template") or {}).get("name") for e in existing_endpoints
+    }
+    template_token = secrets.token_hex(TEMPLATE_TOKEN_BYTES)
     provisionable = {
         h.spec.name: h
         for h in app.resources.values()
@@ -500,6 +530,10 @@ async def reconcile_endpoints(
         members = None
         if needs_gpu_catalog(handle.spec.gpu):
             members = await client.gpu_pool_members()
+        if name in existing:
+            template_name = existing_templates.get(name)
+        else:
+            template_name = _new_template_name(app, handle.spec, template_token)
         payload = _deployed_endpoint_input(
             app,
             handle.spec,
@@ -508,6 +542,7 @@ async def reconcile_endpoints(
             python_version,
             env_name=environment["name"],
             pool_members=members or {},
+            template_name=template_name,
         )
         await attach_endpoint_volumes(payload, handle.spec, resolver, app)
         auth_id = await resolve_registry_auth(handle.spec.registry_auth, api=client)

tests/test_apps/test_deploy.py

@@ -418,3 +418,84 @@ def test_import_time_invocation_message(self, tmp_path):
         client.write_text("from main import que1\nque1.remote(1)\n")
         with pytest.raises(DiscoveryError, match="invoked at import time"):
             discover_apps(client)
+
+
+class TestDeployedTemplateNames:
+    """Template names are unique per owner, and a deleted endpoint holds its
+    template for a while, so new endpoints must not reuse a fixed name."""
+
+    @staticmethod
+    def _app():
+        app = App("demo-app")
+
+        @app.queue(name="que1", cpu="cpu5c-2-4")
+        def que1(x: int):
+            return x
+
+        return app
+
+    @staticmethod
+    def _client():
+        client = AsyncMock()
+        client.save_endpoint.side_effect = lambda payload: {
+            "id": payload.get("id", "ep-new")
+        }
+        return client
+
+    @staticmethod
+    def _saved_template_name(client) -> str:
+        (payload,), _ = client.save_endpoint.await_args
+        return payload["template"]["name"]
+
+    async def _reconcile(self, environment):
+        from runpod.apps.deploy import reconcile_endpoints
+
+        client = self._client()
+        await reconcile_endpoints(client, self._app(), environment, "b-1", "3.12")
+        return client
+
+    async def test_new_endpoint_template_name_has_unique_suffix(self):
+        environment = {"id": "env-1", "name": "default", "endpoints": []}
+
+        first = self._saved_template_name(await self._reconcile(environment))
+        second = self._saved_template_name(await self._reconcile(environment))
+
+        assert first.startswith("demo-app-que1-template-")
+        assert second.startswith("demo-app-que1-template-")
+        assert first != second
+
+    async def test_update_keeps_the_existing_template_name(self):
+        environment = {
+            "id": "env-1",
+            "name": "default",
+            "endpoints": [
+                {"id": "ep-1", "name": "que1", "template": {"name": "kept-name"}}
+            ],
+        }
+
+        client = await self._reconcile(environment)
+
+        (payload,), _ = client.save_endpoint.await_args
+        assert payload["id"] == "ep-1"
+        assert payload["template"]["name"] == "kept-name"
+
+    async def test_update_without_template_info_uses_the_legacy_name(self):
+        environment = {
+            "id": "env-1",
+            "name": "default",
+            "endpoints": [{"id": "ep-1", "name": "que1", "template": None}],
+        }
+
+        client = await self._reconcile(environment)
+
+        assert self._saved_template_name(client) == "demo-app-que1-template"
+
+    @pytest.mark.parametrize(
+        "query_name", ["QUERY_FLASH_APP_BY_NAME", "QUERY_FLASH_ENVIRONMENT_BY_NAME"]
+    )
+    def test_environment_queries_select_the_bound_template_name(self, query_name):
+        from runpod.api.queries import apps as app_queries
+
+        query = " ".join(getattr(app_queries, query_name).split())
+
+        assert "endpoints { id name template { name } }" in query