Startup 3.5x faster

Pull Request resolved: #2099

Fixing various simple things that claude helped find.

1.7 seconds --> .5 seconds.

* Don't start the exponential backoff for initial connections at 50ms. We hit these like 20 times in a row and it adds up to 1/2 second!

* We need to rethink the batched aggregation interval. We probably need a simple predictor that assumes the average time between messages in `inf` for the first message, and just sends it, that way we do not add another 50 ms on every process start.

* Do not import ipython to check if ipython is active. If it isn't active this adds 1/2 second to startup time because it happens in three processes.
ghstack-source-id: 328496691
@exported-using-ghexport

Differential Revision: [D88810033](https://our.internmc.facebook.com/intern/diff/D88810033/)

**NOTE FOR REVIEWERS**: This PR has internal Meta-specific changes or comments, please review them on [Phabricator](https://our.internmc.facebook.com/intern/diff/D88810033/)!

Author: zdevito

hyperactor/src/channel/net/client.rs

@@ -543,7 +543,7 @@ impl<S: Stream> Conn<S> {
     fn reconnect_with_default() -> Self {
         Self::Disconnected(Box::new(
             ExponentialBackoffBuilder::new()
-                .with_initial_interval(Duration::from_millis(50))
+                .with_initial_interval(Duration::from_millis(1))
                 .with_multiplier(2.0)
                 .with_randomization_factor(0.1)
                 .with_max_interval(Duration::from_millis(1000))

hyperactor/src/config.rs

@@ -74,7 +74,7 @@ declare_attrs! {
         env_name: Some("HYPERACTOR_SPLIT_MAX_BUFFER_AGE".to_string()),
         py_name: None,
     })
-    pub attr SPLIT_MAX_BUFFER_AGE: Duration = Duration::from_millis(50);
+    pub attr SPLIT_MAX_BUFFER_AGE: Duration = Duration::from_millis(1);
 
     /// Timeout used by proc mesh for stopping an actor.
     @meta(CONFIG = ConfigAttr {

python/monarch/_src/actor/__init__.py

@@ -11,16 +11,10 @@
 """
 
 from monarch.config import configure
+from monarch._src.ipython_check import is_ipython
 
 # Detect if we're running in IPython/Jupyter
-_in_ipython = False
-try:
-    # pyre-ignore[21]
-    from IPython import get_ipython
-
-    _in_ipython = get_ipython() is not None
-except ImportError:
-    pass
+_in_ipython = is_ipython()
 
 # Set notebook-friendly defaults for stdio piping when spawning procs.
 # These config is read by:

python/monarch/_src/actor/logging.py

@@ -21,19 +21,9 @@
 )
 from monarch._src.actor.actor_mesh import context
 from monarch._src.actor.future import Future
+from monarch._src.ipython_check import is_ipython
 
-IN_IPYTHON = False
-try:
-    # Check if we are in ipython environment
-    # pyre-ignore[21]
-    from IPython import get_ipython
-
-    # pyre-ignore[21]
-    from IPython.core.interactiveshell import ExecutionResult  # noqa: F401
-
-    IN_IPYTHON = get_ipython() is not None
-except ImportError:
-    pass
+IN_IPYTHON = is_ipython()
 
 logger: logging.Logger = logging.getLogger(__name__)
 
@@ -88,7 +78,12 @@ def register_flusher_if_in_ipython(self) -> None:
             with _global_flush_lock:
                 global _global_flush_registered
                 if not _global_flush_registered:
-                    get_ipython().events.register(
+                    # pyre-ignore[21]: IPython is already loaded if IN_IPYTHON is True
+                    from IPython import get_ipython
+
+                    ipython = get_ipython()
+                    assert ipython is not None
+                    ipython.events.register(
                         "post_run_cell",
                         lambda _: flush_all_proc_mesh_logs(
                             self._logging_mesh_client is not None

python/monarch/_src/ipython_check.py

@@ -0,0 +1,26 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+# pyre-strict
+
+"""
+Fast IPython detection by checking sys.modules instead of importing.
+"""
+
+import sys
+
+
+def is_ipython() -> bool:
+    """
+    Check if code is running in an IPython/Jupyter environment.
+    Avoids slow IPython import by checking sys.modules first.
+    """
+    if "IPython" not in sys.modules:
+        return False
+
+    # pyre-ignore[16]: get_ipython exists in IPython module
+    get_ipython = sys.modules["IPython"].get_ipython
+    return get_ipython() is not None

python/monarch/controller/debugger.py

@@ -6,22 +6,8 @@
 
 # pyre-strict
 import sys
-from typing import Optional
 
-_is_ipython: Optional[bool] = None
-
-
-def is_ipython() -> bool:
-    global _is_ipython
-    if _is_ipython is not None:
-        return _is_ipython
-    try:
-        from IPython import get_ipython
-
-        _is_ipython = get_ipython() is not None
-    except ImportError:
-        _is_ipython = False
-    return _is_ipython
+from monarch._src.ipython_check import is_ipython
 
 
 def write(msg: str) -> None: