PYTHON-5536 Avoid clearing the connection pool when the server connection rate limiter triggers

.evergreen/run-tests.sh

@@ -26,12 +26,9 @@ else
 fi
 
 # List the packages.
-uv sync ${UV_ARGS} --reinstall
+uv sync ${UV_ARGS} --reinstall --quiet
 uv pip list
 
-# Ensure we go back to base environment after the test.
-trap "uv sync" EXIT HUP
-
 # Start the test runner.
 uv run ${UV_ARGS} .evergreen/scripts/run_tests.py "$@"
 

justfile

@@ -1,10 +1,11 @@
 # See https://just.systems/man/en/ for instructions
 set shell := ["bash", "-c"]
+# Do not modify the lock file when running justfile commands.
+export UV_FROZEN := "1"
 
 # Commonly used command segments.
-uv_run := "uv run --isolated --frozen "
-typing_run := uv_run + "--group typing --extra aws --extra encryption --extra ocsp --extra snappy --extra test --extra zstd"
-docs_run := uv_run + "--extra docs"
+typing_run := "uv run --group typing --extra aws --extra encryption --extra ocsp --extra snappy --extra test --extra zstd"
+docs_run := "uv run --extra docs"
 doc_build := "./doc/_build"
 mypy_args := "--install-types --non-interactive"
 
@@ -13,51 +14,55 @@ mypy_args := "--install-types --non-interactive"
 default:
   @just --list
 
+[private]
+resync:
+ @uv sync --quiet --frozen
+
 install:
    bash .evergreen/scripts/setup-dev-env.sh
 
 [group('docs')]
-docs:
+docs: && resync
     {{docs_run}} sphinx-build -W -b html doc {{doc_build}}/html
 
 [group('docs')]
-docs-serve:
+docs-serve: && resync
     {{docs_run}} sphinx-autobuild -W -b html doc --watch ./pymongo --watch ./bson --watch ./gridfs {{doc_build}}/serve
 
 [group('docs')]
-docs-linkcheck:
+docs-linkcheck: && resync
     {{docs_run}} sphinx-build -E -b linkcheck doc {{doc_build}}/linkcheck
 
 [group('typing')]
-typing:
+typing: && resync
     just typing-mypy
     just typing-pyright
 
 [group('typing')]
-typing-mypy:
+typing-mypy: && resync
     {{typing_run}} mypy {{mypy_args}} bson gridfs tools pymongo
     {{typing_run}} mypy {{mypy_args}} --config-file mypy_test.ini test
     {{typing_run}} mypy {{mypy_args}} test/test_typing.py test/test_typing_strict.py
 
 [group('typing')]
-typing-pyright:
+typing-pyright: && resync
     {{typing_run}} pyright test/test_typing.py test/test_typing_strict.py
     {{typing_run}} pyright -p strict_pyrightconfig.json test/test_typing_strict.py
 
 [group('lint')]
-lint:
-    {{uv_run}} pre-commit run --all-files
+lint: && resync
+    uv run pre-commit run --all-files
 
 [group('lint')]
-lint-manual:
-    {{uv_run}} pre-commit run --all-files --hook-stage manual
+lint-manual: && resync
+    uv run pre-commit run --all-files --hook-stage manual
 
 [group('test')]
-test *args="-v --durations=5 --maxfail=10":
-    {{uv_run}} --extra test pytest {{args}}
+test *args="-v --durations=5 --maxfail=10": && resync
+    uv run --extra test pytest {{args}}
 
 [group('test')]
-run-tests *args:
+run-tests *args: && resync
     bash ./.evergreen/run-tests.sh {{args}}
 
 [group('test')]

pymongo/asynchronous/pool.py

@@ -37,7 +37,7 @@
 from bson import DEFAULT_CODEC_OPTIONS
 from pymongo import _csot, helpers_shared
 from pymongo.asynchronous.client_session import _validate_session_write_concern
-from pymongo.asynchronous.helpers import _handle_reauth
+from pymongo.asynchronous.helpers import _backoff, _handle_reauth
 from pymongo.asynchronous.network import command
 from pymongo.common import (
     MAX_BSON_SIZE,
@@ -788,9 +788,9 @@ def __init__(
         # Enforces: maxConnecting
         # Also used for: clearing the wait queue
         self._max_connecting_cond = _async_create_condition(self.lock)
-        self._max_connecting = self.opts.max_connecting
         self._pending = 0
         self._client_id = client_id
+        self._backoff = 0
         if self.enabled_for_cmap:
             assert self.opts._event_listeners is not None
             self.opts._event_listeners.publish_pool_created(
@@ -846,6 +846,8 @@ async def _reset(
         async with self.size_cond:
             if self.closed:
                 return
+            # Clear the backoff state.
+            self._backoff = 0
             if self.opts.pause_enabled and pause and not self.opts.load_balanced:
                 old_state, self.state = self.state, PoolState.PAUSED
             self.gen.inc(service_id)
@@ -928,6 +930,11 @@ async def _reset(
                 for conn in sockets:
                     await conn.close_conn(ConnectionClosedReason.STALE)
 
+    @property
+    def max_connecting(self) -> int:
+        """The current max connecting limit for the pool."""
+        return 1 if self._backoff else self.opts.max_connecting
+
     async def update_is_writable(self, is_writable: Optional[bool]) -> None:
         """Updates the is_writable attribute on all sockets currently in the
         Pool.
@@ -994,7 +1001,7 @@ async def remove_stale_sockets(self, reference_generation: int) -> None:
                 async with self._max_connecting_cond:
                     # If maxConnecting connections are already being created
                     # by this pool then try again later instead of waiting.
-                    if self._pending >= self._max_connecting:
+                    if self._pending >= self.max_connecting:
                         return
                     self._pending += 1
                     incremented = True
@@ -1022,6 +1029,19 @@ async def remove_stale_sockets(self, reference_generation: int) -> None:
                     self.requests -= 1
                     self.size_cond.notify()
 
+    def _handle_connection_error(self, error: BaseException, phase: str) -> None:
+        # Handle system overload condition for non-sdam pools.
+        # Look for an AutoReconnect error raised from a ConnectionResetError with
+        # errno == errno.ECONNRESET or raised from an OSError that we've created due to
+        # a closed connection.
+        # If found, set backoff and add error labels.
+        if self.is_sdam or type(error) != AutoReconnect:
+            return
+        self._backoff += 1
+        error._add_error_label("SystemOverloadedError")
+        error._add_error_label("RetryableError")
+        print(f"Setting backoff in {phase}:", self._backoff)  # noqa: T201
+
     async def connect(self, handler: Optional[_MongoClientErrorHandler] = None) -> AsyncConnection:
         """Connect to Mongo and return a new AsyncConnection.
 
@@ -1051,8 +1071,17 @@ async def connect(self, handler: Optional[_MongoClientErrorHandler] = None) -> A
                 driverConnectionId=conn_id,
             )
 
+        # Apply backoff if applicable.
+        if self._backoff:
+            await asyncio.sleep(_backoff(self._backoff))
+
+        # Pass a context to determine if we successfully create a configured socket.
+        context = dict(has_created_socket=False)
+
         try:
-            networking_interface = await _configured_protocol_interface(self.address, self.opts)
+            networking_interface = await _configured_protocol_interface(
+                self.address, self.opts, context=context
+            )
         # Catch KeyboardInterrupt, CancelledError, etc. and cleanup.
         except BaseException as error:
             async with self.lock:
@@ -1073,10 +1102,11 @@ async def connect(self, handler: Optional[_MongoClientErrorHandler] = None) -> A
                     reason=_verbose_connection_error_reason(ConnectionClosedReason.ERROR),
                     error=ConnectionClosedReason.ERROR,
                 )
+            if context["has_created_socket"]:
+                self._handle_connection_error(error, "handshake")
             if isinstance(error, (IOError, OSError, *SSLErrors)):
                 details = _get_timeout_details(self.opts)
                 _raise_connection_failure(self.address, error, timeout_details=details)
-
             raise
 
         conn = AsyncConnection(networking_interface, self, self.address, conn_id, self.is_sdam)  # type: ignore[arg-type]
@@ -1094,15 +1124,18 @@ async def connect(self, handler: Optional[_MongoClientErrorHandler] = None) -> A
 
             await conn.authenticate()
         # Catch KeyboardInterrupt, CancelledError, etc. and cleanup.
-        except BaseException:
+        except BaseException as e:
             async with self.lock:
                 self.active_contexts.discard(conn.cancel_context)
+            self._handle_connection_error(e, "hello")
             await conn.close_conn(ConnectionClosedReason.ERROR)
             raise
 
         if handler:
             await handler.client._topology.receive_cluster_time(conn._cluster_time)
 
+        # Clear the backoff state.
+        self._backoff = 0
         return conn
 
     @contextlib.asynccontextmanager
@@ -1279,12 +1312,12 @@ async def _get_conn(
                 # to be checked back into the pool.
                 async with self._max_connecting_cond:
                     self._raise_if_not_ready(checkout_started_time, emit_event=False)
-                    while not (self.conns or self._pending < self._max_connecting):
+                    while not (self.conns or self._pending < self.max_connecting):
                         timeout = deadline - time.monotonic() if deadline else None
                         if not await _async_cond_wait(self._max_connecting_cond, timeout):
                             # Timed out, notify the next thread to ensure a
                             # timeout doesn't consume the condition.
-                            if self.conns or self._pending < self._max_connecting:
+                            if self.conns or self._pending < self.max_connecting:
                                 self._max_connecting_cond.notify()
                             emitted_event = True
                             self._raise_wait_queue_timeout(checkout_started_time)
@@ -1425,8 +1458,8 @@ async def _perished(self, conn: AsyncConnection) -> bool:
         :class:`~pymongo.errors.AutoReconnect` exceptions on server
         hiccups, etc. We only check if the socket was closed by an external
         error if it has been > 1 second since the socket was checked into the
-        pool, to keep performance reasonable - we can't avoid AutoReconnects
-        completely anyway.
+        pool, or we are in backoff mode, to keep performance reasonable -
+        we can't avoid AutoReconnects completely anyway.
         """
         idle_time_seconds = conn.idle_time_seconds()
         # If socket is idle, open a new one.
@@ -1437,8 +1470,11 @@ async def _perished(self, conn: AsyncConnection) -> bool:
             await conn.close_conn(ConnectionClosedReason.IDLE)
             return True
 
-        if self._check_interval_seconds is not None and (
-            self._check_interval_seconds == 0 or idle_time_seconds > self._check_interval_seconds
+        check_interval_seconds = self._check_interval_seconds
+        if self._backoff:
+            check_interval_seconds = 0
+        if check_interval_seconds is not None and (
+            check_interval_seconds == 0 or idle_time_seconds > check_interval_seconds
         ):
             if conn.conn_closed():
                 await conn.close_conn(ConnectionClosedReason.ERROR)

pymongo/asynchronous/topology.py

@@ -890,7 +890,9 @@ async def _handle_error(self, address: _Address, err_ctx: _ErrorContext) -> None
                 # Clear the pool.
                 await server.reset(service_id)
         elif isinstance(error, ConnectionFailure):
-            if isinstance(error, WaitQueueTimeoutError):
+            if isinstance(error, WaitQueueTimeoutError) or error.has_error_label(
+                "SystemOverloadedError"
+            ):
                 return
             # "Client MUST replace the server's description with type Unknown
             # ... MUST NOT request an immediate check of the server."

pymongo/network_layer.py

@@ -256,6 +256,7 @@ def __init__(self, timeout: Optional[float] = None):
         self._timeout = timeout
         self._closed = asyncio.get_running_loop().create_future()
         self._connection_lost = False
+        self._closing_exception = None
 
     def settimeout(self, timeout: float | None) -> None:
         self._timeout = timeout
@@ -269,9 +270,11 @@ def close(self, exc: Optional[Exception] = None) -> None:
         self.transport.abort()
         self._resolve_pending(exc)
         self._connection_lost = True
+        self._closing_exception = exc  # type:ignore[assignment]
 
     def connection_lost(self, exc: Optional[Exception] = None) -> None:
         self._resolve_pending(exc)
+        self._closing_exception = exc  # type:ignore[assignment]
         if not self._closed.done():
             self._closed.set_result(None)
 
@@ -335,8 +338,11 @@ async def read(self, request_id: Optional[int], max_message_size: int) -> tuple[
         if self._done_messages:
             message = await self._done_messages.popleft()
         else:
-            if self.transport and self.transport.is_closing():
-                raise OSError("connection is already closed")
+            if self._closed.done():
+                if self._closing_exception:
+                    raise self._closing_exception
+                else:
+                    raise OSError("connection closed")
             read_waiter = asyncio.get_running_loop().create_future()
             self._pending_messages.append(read_waiter)
             try:
@@ -474,6 +480,7 @@ def _resolve_pending(self, exc: Optional[Exception] = None) -> None:
                 else:
                     msg.set_exception(exc)
             self._done_messages.append(msg)
+        self._pending_messages.clear()
 
 
 class PyMongoKMSProtocol(PyMongoBaseProtocol):

pymongo/pool_shared.py

@@ -250,6 +250,7 @@ async def _configured_protocol_interface(
     address: _Address,
     options: PoolOptions,
     protocol_kls: type[PyMongoBaseProtocol] = PyMongoProtocol,
+    context: dict[str, bool] | None = None,
 ) -> AsyncNetworkingInterface:
     """Given (host, port) and PoolOptions, return a configured AsyncNetworkingInterface.
 
@@ -261,6 +262,10 @@ async def _configured_protocol_interface(
     ssl_context = options._ssl_context
     timeout = options.socket_timeout
 
+    # Signal that we have created the socket successfully.
+    if context:
+        context["has_created_socket"] = True
+
     if ssl_context is None:
         return AsyncNetworkingInterface(
             await asyncio.get_running_loop().create_connection(
@@ -374,7 +379,7 @@ def _create_connection(address: _Address, options: PoolOptions) -> socket.socket
 
 
 def _configured_socket_interface(
-    address: _Address, options: PoolOptions, *args: Any
+    address: _Address, options: PoolOptions, *args: Any, context: dict[str, bool] | None = None
 ) -> NetworkingInterface:
     """Given (host, port) and PoolOptions, return a NetworkingInterface wrapping a configured socket.
 
@@ -385,6 +390,10 @@ def _configured_socket_interface(
     sock = _create_connection(address, options)
     ssl_context = options._ssl_context
 
+    # Signal that we have created the socket successfully.
+    if context:
+        context["has_created_socket"] = True
+
     if ssl_context is None:
         sock.settimeout(options.socket_timeout)
         return NetworkingInterface(sock)