Object dispatch not working #13

Activity

1. thorwhalen commented on Feb 22, 2024

   

   thorwhalen

   on Feb 22, 2024

   MemberAuthor

   More actions

   The problem may not be with object dispatch but something else, since web services that worked before don't any more.

   For example, getting this:

   Bottle v0.12.25 server starting up (using GunicornServer(certfile=None, keyfile=None))...
   Listening on http://localhost:3030/
   Hit Ctrl-C to quit.
   
   [2024-02-22 08:50:37 +0100] [29952] [INFO] Starting gunicorn 21.2.0
   [2024-02-22 08:50:37 +0100] [29952] [INFO] Listening at: http://127.0.0.1:3030 (29952)
   [2024-02-22 08:50:37 +0100] [29952] [INFO] Using worker: sync
   [2024-02-22 08:50:37 +0100] [30000] [INFO] Booting worker with pid: 30000
   [2024-02-22 08:51:24 +0100] [29952] [CRITICAL] WORKER TIMEOUT (pid:30000)
   [2024-02-22 08:51:24 +0100] [30000] [INFO] Worker exiting (pid: 30000)
   [2024-02-22 08:51:25 +0100] [29952] [ERROR] Worker (pid:30000) exited with code 1
   [2024-02-22 08:51:25 +0100] [29952] [ERROR] Worker (pid:30000) exited with code 1.
   [2024-02-22 08:51:25 +0100] [30003] [INFO] Booting worker with pid: 30003
   

   with another one.
2. thorwhalen commented on Feb 22, 2024

   

   thorwhalen

   on Feb 22, 2024

   MemberAuthor

   More actions

   Okay, so there's layers of things happening here, some that I don't understand.

   What I don't understand

   unpredictable backend server choice

   Sometimes py2http chooses to use WSGIRefServer instead of GunicornServer -- unpredictably, for me.

   Bottle v0.12.25 server starting up (using WSGIRefServer())...
   Listening on http://127.0.0.1:8080/
   Hit Ctrl-C to quit.
   

   Swagger works... sometimes

   Swagger works... and the doesn't anymore (I start getting those "CORS or Network failure or URL scheme" errors, and then they stick, even after relaunching the server.

   Now, even while swagger doesn't work, I still get the curl and requests to work.

   This is not always the case, but that's because Swagger's message is catchall. The errors I was getting before, I was getting them simply because there was a network error (see later).

   What I DO understand

   Swagger curl pointing to wrong url

   So, when (for reasons I do not know) py2http chooses to use WSGIRefServer, it listens on 8080.
   To go to swagger, I also use 8080 -- but when I execute, swagger displays a curl that uses 3030.
   So of course, it fails with the catchall (unhelpful) error message.

   Argument type

   I used:

     "_attr_name": "__iter__",
     "kwargs": {}
   

   because it's the example I found, and I assumed that one was supposed to specify all arguments in a kwargs dict,
   which is empty here since __iter__ has no arguments. This is not how it works.
   Instead, I should be specifying the keyword arguments at the same level as the "_attr_name" field.
   So I should be doing:

   url = 'http://localhost:3030/backend_store'
   add_args = {
     "_attr_name": '__iter__',
   }
   requests.post(url, json=add_args).json()

   or

   curl -X 'POST' \
     'http://localhost:3030/backend_store' \
     -H 'accept: application/json' \
     -H 'Content-Type: application/json' \
     -d '{
     "_attr_name": "__iter__"
   }'
   

   Argument type

   But we still have a problem.
   We'd like to do

   url = 'http://localhost:3030/backend_store'
   add_args = {
     "_attr_name": '__getitem__',
     "key": 'some_key'
   }
   requests.post(url, json=add_args).json()
   

   but actually dict.__getitem__ has only one, positional-only, argument.

   So: Handling position-only arguments issue.

   Output types

   There were two more problems with py2http/examples/object_dispatch.py code.
   One is that the values of the backend_store were bytes: This causes problems when doing a __getitem__ since it's not JSON serializable.
   The second is that the __iter__ return type is a generator, also not serializable.

   Both of these things can be handled at the egress level (output_trans), but since we also have the problem of the positional-only, we can also just wrap the dict.
   So we can change our previous code to this new code:

   from typing import MutableMapping
   
   class BackendStore(MutableMapping):
       """MutableMapping wrapper that assumes that values are given as strings,
       but stored as bytes"""
   
       def __init__(self, store):
           self.store = store
   
       def __iter__(self):
           return list(self.store)
   
       def __getitem__(self, key: str) -> str:
           return self.store[key].decode()
   
       def __setitem__(self, key: str, value: str):
           self.store[key] = value.encode()
   
       def __len__(self) -> int:
           return len(self.store)
   
       def __delitem__(self, key: str):
           del self.store[key]
   
   
   backend_store = BackendStore(
       {
           'some_key': b'some_value',
           'some_other_key': b'some_other_value',
       }
   )
   
   handlers = [
       dict(
           endpoint=backend_store,
           name='backend_store',
           attr_names=['__iter__', '__getitem__', '__setitem__'],
       ),
   ]
   
   from py2http import mk_app
   
   app = mk_app(handlers, publish_openapi=True, publish_swagger=True)
   
   if __name__ == '__main__':
       app.run()

   Of course, if not explicitly specified, we still can't predict what the root url will be (depends on the unpredictable server that will be chosen), but here's some code to test the service above:

   def test_service(rooturl='http://localhost:8080'):
       import requests
   
       url = f'{rooturl}/backend_store'
   
       add_args = {
           "_attr_name": '__iter__',
       }
       assert requests.post(url, json=add_args).json() == ['some_key', 'some_other_key']
   
       add_args = {"_attr_name": '__getitem__', "key": 'some_key'}
       assert requests.post(url, json=add_args).json() == 'some_value'
3. thorwhalen

   added

   bugSomething isn't workingSomething isn't working

   on Sep 22, 2026
4. thorwhalen commented on Sep 22, 2026

   

   thorwhalen

   on Sep 22, 2026

   MemberAuthor

   More actions

   Verified still live against current master (ee5b962). Root cause: in py2http/service.py, _get_func_to_dispatch_handler builds the object-dispatch wrapper as func(_attr_name=None, **kwargs) and _get_attr_value finally calls attr(**kwargs) (service.py:589-632). Since the request body is always turned into keyword arguments, any dunder/builtin method whose parameters are positional-only (e.g. dict.__getitem__(self, key, /), __setitem__) fails with TypeError: __getitem__() takes no keyword arguments. This is the same root cause as #14 — fixing one fixes both. A fix would need _get_attr_value (or the wrapper) to inspect the target callable's signature and pass POSITIONAL_ONLY parameters positionally (py2http already has the POSITIONAL_ONLY/KEYWORD_ONLY plumbing in py2http/types.py and py2http/util.py for the opposite direction — py→http — so the introspection primitives exist, just not wired into the http→object-attr call path).
5. thorwhalen

   added a commit that references this issue on Oct 13, 2026

   Call positional-only parameters positionally, by name (#14) (#33)

   ...

   3866d3a
6. thorwhalen commented on Oct 13, 2026

   

   thorwhalen

   on Oct 13, 2026

   MemberAuthor

   More actions

   The positional-only half of this (attr(**kwargs) on __getitem__/__setitem__) is fixed by #33 (closes #14). The gunicorn worker timeout reported here was not reproduced and is not addressed, so this stays open.
7. thorwhalen

   added a commit that references this issue on Oct 20, 2026

   object_dispatch example: serve with run_app so the port matches the O…

   ...

   c7f5b81
8. thorwhalen commented on Oct 20, 2026

   

   thorwhalen

   on Oct 20, 2026

   MemberAuthor

   More actions

   Diagnosis. Two of the reported symptoms were one cause: the example ran the app with bottle's own app.run() (wsgiref, port 8080), while the generated OpenAPI spec names http://localhost:3030 (where run_app serves, with gunicorn). So Swagger's "Try it out" and its curl hit a port nothing listened on. Fixed in #48: the example uses run_app, and the spec's default server URL is derived from the default config so the two cannot drift. Positional-only dunders were fixed by #33. Still open: the gunicorn WORKER TIMEOUT on one of the reporter's other services, which I could not reproduce and which has no repro in the thread; leaving this open for a repro (the service and the server= setting would help).