Add back uuid

src/manager.rs

@@ -37,14 +37,19 @@ use log::{info, warn};
 #[cfg(test)]
 use std::{println as info, println as warn};
 
+// The replica_id string is of the form {replica_name}:{uuid} or just {uuid} (see torchft/manager.py)
+// We can parse the replica_id if it exists, otherwise we just use the uuid
 macro_rules! info_with_replica {
-    ($replica_id:expr, $($arg:tt)*) => {
-        info!(
-            "[Replica {}] {}",
-            $replica_id,
-            format!($($arg)*)
-        );
-    };
+    ($replica_id:expr, $($arg:tt)*) => {{
+        let parts: Vec<&str> = $replica_id.splitn(2, ':').collect();
+        let formatted_message = if parts.len() == 2 {
+            // If there are two parts, use the replica name
+            info!("[Replica {}] {}", parts[0], format!($($arg)*))
+        } else {
+            // Otherwise, just use the UUID
+            info!("[Replica {}] {}", $replica_id, format!($($arg)*))
+        };
+    }};
 }
 
 struct ManagerState {

torchft/manager.py

@@ -29,6 +29,7 @@
 import logging
 import os
 import socket
+import uuid
 from concurrent.futures import ThreadPoolExecutor
 from datetime import timedelta
 from enum import Enum
@@ -188,9 +189,13 @@ def __init__(
             bind = f"[::]:{port}"
             lighthouse_addr = lighthouse_addr or os.environ["TORCHFT_LIGHTHOUSE"]
 
-            if replica_id is None:
-                replica_id = ""
-            replica_id = replica_id
+            # We need a unique identifier in the case that a worker restarts quickly and
+            # replaces the previous worker with the same ID.
+            new_uuid = str(uuid.uuid4())
+            if replica_id is None or replica_id == "":
+                replica_id = new_uuid
+            else:
+                replica_id = f"{replica_id}:{new_uuid}"
             self._manager = ManagerServer(
                 replica_id=replica_id,
                 lighthouse_addr=lighthouse_addr,