Merge pull request #1283 from OCR-D/mets-server-kill-zombies

add endpoint DELETE /kill_mets_server_zombies to kill -SIGTERM METS servers with ctime > 60mins ago

src/ocrd/mets_server.py

@@ -1,8 +1,10 @@
 """
 # METS server functionality
 """
+import os
 import re
 from os import _exit, chmod
+import signal
 from typing import Dict, Optional, Union, List, Tuple
 from time import sleep
 from pathlib import Path
@@ -428,8 +430,12 @@ def create_process(mets_server_url: str, ws_dir_path: str, log_file: str) -> int
 
     @staticmethod
     def kill_process(mets_server_pid: int):
-        subprocess_run(args=["kill", "-s", "SIGINT", f"{mets_server_pid}"], shell=False, universal_newlines=True)
-        return
+        os.kill(mets_server_pid, signal.SIGINT)
+        sleep(3)
+        try:
+            os.kill(mets_server_pid, signal.SIGKILL)
+        except ProcessLookupError as e:
+            pass
 
     def shutdown(self):
         if self.is_uds:

src/ocrd_network/processing_server.py

@@ -48,6 +48,7 @@
     get_workflow_content,
     get_from_database_workspace,
     get_from_database_workflow_job,
+    kill_mets_server_zombies,
     parse_workflow_tasks,
     raise_http_exception,
     request_processor_server_tool_json,
@@ -200,6 +201,14 @@ def add_api_routes_others(self):
             tags=[ServerApiTags.WORKSPACE],
             summary="Forward a TCP request to UDS mets server"
         )
+        others_router.add_api_route(
+            path="/kill_mets_server_zombies",
+            endpoint=self.kill_mets_server_zombies,
+            methods=["DELETE"],
+            tags=[ServerApiTags.WORKFLOW, ServerApiTags.PROCESSING],
+            status_code=status.HTTP_200_OK,
+            summary="!! Workaround Do Not Use Unless You Have A Reason !! Kill all METS servers on this machine that have been created more than 60 minutes ago."
+        )
         self.include_router(others_router)
 
     def add_api_routes_processing(self):
@@ -817,6 +826,10 @@ async def get_workflow_info(self, workflow_job_id) -> Dict:
         response = self._produce_workflow_status_response(processing_jobs=jobs)
         return response
 
+    async def kill_mets_server_zombies(self) -> List[int]:
+        pids_killed = kill_mets_server_zombies(minutes_ago=60)
+        return pids_killed
+
     async def get_workflow_info_simple(self, workflow_job_id) -> Dict[str, JobState]:
         """
         Simplified version of the `get_workflow_info` that returns a single state for the entire workflow.

src/ocrd_network/server_utils.py

@@ -1,12 +1,18 @@
+import os
+import re
+import signal
+from pathlib import Path
+from json import dumps, loads
+from urllib.parse import urljoin
+from typing import Dict, List, Union
+from time import time
+
 from fastapi import HTTPException, status, UploadFile
 from fastapi.responses import FileResponse
 from httpx import AsyncClient, Timeout
-from json import dumps, loads
 from logging import Logger
-from pathlib import Path
 from requests import get as requests_get
-from typing import Dict, List, Union
-from urllib.parse import urljoin
+from requests_unixsocket import sys
 
 from ocrd.resolver import Resolver
 from ocrd.task_sequence import ProcessorTask
@@ -241,3 +247,25 @@ def validate_first_task_input_file_groups_existence(logger: Logger, mets_path: s
         if group not in available_groups:
             message = f"Input file group '{group}' of the first processor not found: {input_file_grps}"
             raise_http_exception(logger, status.HTTP_422_UNPROCESSABLE_ENTITY, message)
+
+
+def kill_mets_server_zombies(minutes_ago=60) -> List[int]:
+    now = time()
+    cmdline_pat = r'.*ocrd workspace -U.*server start $'
+    ret = []
+    for procdir in sorted(Path('/proc').glob('*'), key=os.path.getctime):
+        if not procdir.is_dir():
+            continue
+        cmdline_file = procdir.joinpath('cmdline')
+        if not cmdline_file.is_file():
+            continue
+        ctime_ago = int((now - procdir.stat().st_ctime) / 60)
+        if ctime_ago < minutes_ago:
+            continue
+        cmdline = cmdline_file.read_text().replace('\x00', ' ')
+        if re.match(cmdline_pat, cmdline):
+            pid = procdir.name
+            ret.append(pid)
+            print(f'METS Server with PID {pid} was created {ctime_ago} minutes ago, more than {minutes_ago}, so killing (cmdline="{cmdline})', file=sys.stderr)
+            os.kill(int(pid), signal.SIGTERM)
+    return ret