diff --git a/CHANGELOG.md b/CHANGELOG.md
index 268574d7b..77af9e6a6 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -21,6 +21,7 @@
 * Fix bug that crashed when using `view.flows.resolve` (#4916, @rbdixon)
 * Fix a bug where `running()` is invoked twice on startup (#3584, @mhils)
 * Correct documentation example for User-Agent header modification (#4997, @jamesyale)
+* Fix random connection stalls (#5040, @EndUser509)
 
 ## 28 September 2021: mitmproxy 7.0.4
 
diff --git a/mitmproxy/proxy/server.py b/mitmproxy/proxy/server.py
index 4b26b23c6..7af7f09d6 100644
--- a/mitmproxy/proxy/server.py
+++ b/mitmproxy/proxy/server.py
@@ -223,11 +223,14 @@ class ConnectionHandler(metaclass=abc.ABCMeta):
             except asyncio.CancelledError as e:
                 cancelled = e
                 break
-            else:
-                self.server_event(events.DataReceived(connection, data))
-                for transport in self.transports.values():
-                    if transport.writer is not None:
-                        await transport.writer.drain()
+
+            self.server_event(events.DataReceived(connection, data))
+
+            try:
+                await self.drain_writers()
+            except asyncio.CancelledError as e:
+                cancelled = e
+                break
 
         if cancelled is None:
             connection.state &= ~ConnectionState.CAN_READ
@@ -253,6 +256,19 @@ class ConnectionHandler(metaclass=abc.ABCMeta):
         if cancelled:
             raise cancelled
 
+    async def drain_writers(self):
+        """
+        Drain all writers to create some backpressure. We won't continue reading until there's space available in our
+        write buffers, so if we cannot write fast enough our own read buffers run full and the TCP recv stream is throttled.
+        """
+        for transport in self.transports.values():
+            if transport.writer is not None:
+                try:
+                    await transport.writer.drain()
+                except OSError as e:
+                    if transport.handler is not None:
+                        asyncio_utils.cancel_task(transport.handler, f"Error sending data: {e}")
+
     async def on_timeout(self) -> None:
         self.log(f"Closing connection due to inactivity: {self.client}")
         handler = self.transports[self.client].handler