speedy-utils 1.0.3__py3-none-any.whl → 1.0.5__py3-none-any.whl

llm_utils/scripts/vllm_load_balancer.py

@@ -0,0 +1,353 @@
+import asyncio
+import random
+from collections import defaultdict
+import time
+from tabulate import tabulate
+import logging
+import contextlib
+import aiohttp # <-- Import aiohttp
+
+# --- Configuration ---
+LOAD_BALANCER_HOST = '0.0.0.0'
+LOAD_BALANCER_PORT = 8008
+
+SCAN_TARGET_HOST = 'localhost'
+SCAN_PORT_START = 8150
+SCAN_PORT_END = 8170 # Inclusive
+SCAN_INTERVAL = 30
+# Timeout applies to the HTTP health check request now
+HEALTH_CHECK_TIMEOUT = 2.0 # Increased slightly for HTTP requests
+
+STATUS_PRINT_INTERVAL = 5
+BUFFER_SIZE = 4096
+
+# Setup basic logging
+logging.basicConfig(level=logging.INFO, format='%(asctime)s - %(levelname)s - %(message)s')
+
+# --- Global Shared State ---
+available_servers = []
+connection_counts = defaultdict(int)
+state_lock = asyncio.Lock()
+
+# --- Helper Functions --- (relay_data and safe_close_writer remain the same)
+async def relay_data(reader, writer, direction):
+    """Reads data from reader and writes to writer until EOF or error."""
+    try:
+        while True:
+            data = await reader.read(BUFFER_SIZE)
+            if not data:
+                logging.debug(f"EOF received on {direction} stream.")
+                break
+            writer.write(data)
+            await writer.drain()
+    except ConnectionResetError:
+        logging.warning(f"Connection reset on {direction} stream.")
+    except asyncio.CancelledError:
+        logging.debug(f"Relay task cancelled for {direction}.")
+        raise
+    except Exception as e:
+        logging.warning(f"Error during data relay ({direction}): {e}")
+    finally:
+        if not writer.is_closing():
+            try:
+                writer.close()
+                await writer.wait_closed()
+                logging.debug(f"Closed writer for {direction}")
+            except Exception as close_err:
+                logging.debug(f"Error closing writer for {direction} (might be expected): {close_err}")
+
+@contextlib.asynccontextmanager
+async def safe_close_writer(writer):
+    """Async context manager to safely close an asyncio StreamWriter."""
+    try:
+        yield writer
+    finally:
+        if writer and not writer.is_closing():
+            try:
+                writer.close()
+                await writer.wait_closed()
+            except Exception as e:
+                logging.debug(f"Error closing writer in context manager: {e}")
+
+# --- Server Scanning and Health Check (Modified) ---
+
+async def check_server_health(session, host, port):
+    """Performs an HTTP GET request to the /health endpoint."""
+    url = f"http://{host}:{port}/health"
+    try:
+        # Use the provided aiohttp session to make the GET request
+        async with session.get(url, timeout=HEALTH_CHECK_TIMEOUT) as response:
+            # Check for a successful status code (2xx range)
+            if 200 <= response.status < 300:
+                logging.debug(f"Health check success for {url} (Status: {response.status})")
+                # Ensure the connection is released back to the pool
+                await response.release()
+                return True
+            else:
+                logging.debug(f"Health check failed for {url} (Status: {response.status})")
+                await response.release()
+                return False
+    except asyncio.TimeoutError:
+        logging.debug(f"Health check HTTP request timeout for {url}")
+        return False
+    except aiohttp.ClientConnectorError as e:
+        # Handles connection refused, DNS errors etc. - server likely down
+        logging.debug(f"Health check connection error for {url}: {e}")
+        return False
+    except aiohttp.ClientError as e:
+        # Catch other potential client errors (e.g., invalid URL structure, too many redirects)
+        logging.warning(f"Health check client error for {url}: {e}")
+        return False
+    except Exception as e:
+        # Catch any other unexpected errors during the check
+        logging.error(f"Unexpected health check error for {url}: {e}")
+        return False
+
+async def scan_and_update_servers():
+    """Periodically scans ports using HTTP /health check and updates available servers."""
+    global available_servers
+    logging.debug(f"Starting server scan task (HTTP GET /health on Ports {SCAN_PORT_START}-{SCAN_PORT_END} every {SCAN_INTERVAL}s)")
+    while True:
+        try:
+            current_scan_results = []
+            scan_tasks = []
+            ports_to_scan = range(SCAN_PORT_START, SCAN_PORT_END + 1)
+
+            # Create ONE aiohttp session for all checks within this scan cycle for efficiency
+            async with aiohttp.ClientSession() as session:
+                # Create health check tasks for all ports, passing the shared session
+                for port in ports_to_scan:
+                    task = asyncio.create_task(check_server_health(session, SCAN_TARGET_HOST, port))
+                    scan_tasks.append((task, port))
+
+                # Wait for all health checks to complete
+                # return_exceptions=True prevents gather from stopping if one check fails
+                await asyncio.gather(*(task for task, port in scan_tasks), return_exceptions=True)
+
+                # Collect results from completed tasks
+                for task, port in scan_tasks:
+                    try:
+                        # Check if task finished, wasn't cancelled, and returned True
+                        if task.done() and not task.cancelled() and task.result() is True:
+                            current_scan_results.append((SCAN_TARGET_HOST, port))
+                    except Exception as e:
+                        # Log errors from the health check task itself if gather didn't catch them
+                        logging.error(f"Error retrieving health check result for port {port}: {e}")
+
+            # --- Update Shared State (Locked) ---
+            async with state_lock:
+                previous_servers = set(available_servers)
+                current_set = set(current_scan_results)
+
+                added = current_set - previous_servers
+                removed = previous_servers - current_set
+
+                if added:
+                    logging.info(f"Servers added (passed /health check): {sorted(list(added))}")
+                if removed:
+                    logging.info(f"Servers removed (failed /health check or stopped): {sorted(list(removed))}")
+                    for server in removed:
+                        if server in connection_counts:
+                            del connection_counts[server]
+                            logging.debug(f"Removed connection count entry for unavailable server {server}")
+
+                available_servers = sorted(list(current_set))
+                for server in available_servers:
+                    if server not in connection_counts:
+                        connection_counts[server] = 0
+
+            logging.debug(f"Scan complete. Active servers: {available_servers}")
+
+        except asyncio.CancelledError:
+             logging.info("Server scan task cancelled.")
+             break
+        except Exception as e:
+            logging.error(f"Error in scan_and_update_servers loop: {e}")
+            await asyncio.sleep(SCAN_INTERVAL / 2) # Avoid tight loop on error
+
+        await asyncio.sleep(SCAN_INTERVAL)
+
+
+# --- Core Load Balancer Logic (handle_client remains the same) ---
+async def handle_client(client_reader, client_writer):
+    """Handles a single client connection."""
+    client_addr = client_writer.get_extra_info('peername')
+    logging.info(f"Accepted connection from {client_addr}")
+
+    backend_server = None
+    backend_reader = None
+    backend_writer = None
+    server_selected = False
+
+    try:
+        # --- Select Backend Server (Least Connections from Available) ---
+        selected_server = None
+        async with state_lock: # Lock to safely access available_servers and connection_counts
+            if not available_servers:
+                logging.warning(f"No backend servers available (failed health checks?) for client {client_addr}. Closing connection.")
+                async with safe_close_writer(client_writer): pass
+                return
+
+            min_connections = float('inf')
+            least_used_available_servers = []
+            for server in available_servers: # Iterate only over servers that passed health check
+                count = connection_counts.get(server, 0)
+                if count < min_connections:
+                    min_connections = count
+                    least_used_available_servers = [server]
+                elif count == min_connections:
+                    least_used_available_servers.append(server)
+
+            if least_used_available_servers:
+                selected_server = random.choice(least_used_available_servers)
+                connection_counts[selected_server] += 1
+                backend_server = selected_server
+                server_selected = True
+                logging.info(f"Routing {client_addr} to {backend_server} (Current connections: {connection_counts[backend_server]})")
+            else:
+                 logging.error(f"Logic error: No server chosen despite available servers list not being empty for {client_addr}.")
+                 async with safe_close_writer(client_writer): pass
+                 return
+
+        # --- Connect to Backend Server ---
+        if not backend_server:
+             logging.error(f"No backend server selected for {client_addr} before connection attempt.")
+             async with safe_close_writer(client_writer): pass
+             server_selected = False
+             return
+
+        try:
+            logging.debug(f"Attempting connection to backend {backend_server} for {client_addr}")
+            backend_reader, backend_writer = await asyncio.open_connection(
+                backend_server[0], backend_server[1]
+            )
+            logging.debug(f"Successfully connected to backend {backend_server} for {client_addr}")
+
+        # Handle connection failure AFTER selection (server might go down between health check and selection)
+        except ConnectionRefusedError:
+            logging.error(f"Connection refused by selected backend server {backend_server} for {client_addr}")
+            async with state_lock: # Decrement count under lock
+                 if backend_server in connection_counts and connection_counts[backend_server] > 0: connection_counts[backend_server] -= 1
+            server_selected = False # Mark failure
+            async with safe_close_writer(client_writer): pass
+            return
+        except Exception as e:
+            logging.error(f"Failed to connect to selected backend server {backend_server} for {client_addr}: {e}")
+            async with state_lock: # Decrement count under lock
+                 if backend_server in connection_counts and connection_counts[backend_server] > 0: connection_counts[backend_server] -= 1
+            server_selected = False # Mark failure
+            async with safe_close_writer(client_writer): pass
+            return
+
+        # --- Relay Data Bidirectionally ---
+        async with safe_close_writer(backend_writer): # Ensure backend writer is closed
+            client_to_backend = asyncio.create_task(
+                relay_data(client_reader, backend_writer, f"{client_addr} -> {backend_server}")
+            )
+            backend_to_client = asyncio.create_task(
+                relay_data(backend_reader, client_writer, f"{backend_server} -> {client_addr}")
+            )
+            done, pending = await asyncio.wait(
+                [client_to_backend, backend_to_client], return_when=asyncio.FIRST_COMPLETED
+            )
+            for task in pending: task.cancel()
+            for task in done:
+                 with contextlib.suppress(asyncio.CancelledError):
+                     if task.exception(): logging.warning(f"Relay task finished with error: {task.exception()}")
+
+    except asyncio.CancelledError:
+        logging.info(f"Client handler for {client_addr} cancelled.")
+    except Exception as e:
+        logging.error(f"Error handling client {client_addr}: {e}")
+    finally:
+        logging.info(f"Closing connection for {client_addr}")
+        # Decrement connection count only if we successfully selected/incremented
+        if backend_server and server_selected:
+            async with state_lock:
+                if backend_server in connection_counts:
+                    if connection_counts[backend_server] > 0:
+                        connection_counts[backend_server] -= 1
+                        logging.info(f"Connection closed for {client_addr}. Backend {backend_server} connections: {connection_counts[backend_server]}")
+                    else:
+                        logging.warning(f"Attempted to decrement count below zero for {backend_server} on close")
+                        connection_counts[backend_server] = 0
+
+# --- Status Reporting Task (print_status_periodically remains the same) ---
+async def print_status_periodically():
+    """Periodically prints the connection status based on available servers."""
+    while True:
+        await asyncio.sleep(STATUS_PRINT_INTERVAL)
+        async with state_lock:
+            headers = ["Backend Server", "Host", "Port", "Active Connections", "Status"]
+            table_data = []
+            total_connections = 0
+            current_available = available_servers[:]
+            current_counts = connection_counts.copy()
+
+        if not current_available:
+            # clear terminal and print status
+            print("\033[H\033[J", end="")  # Clear terminal
+            print("\n----- Load Balancer Status -----")
+            print("No backend servers currently available (failed /health check).")
+            print("------------------------------\n")
+            continue
+
+        for server in current_available:
+            host, port = server
+            count = current_counts.get(server, 0)
+            table_data.append([f"{host}:{port}", host, port, count, "Available"])
+            total_connections += count
+
+        table_data.sort(key=lambda row: (row[1], row[2]))
+
+        try:
+            table = tabulate(table_data, headers=headers, tablefmt="grid")
+            print("\n----- Load Balancer Status -----")
+            print(f"Scanning Ports: {SCAN_PORT_START}-{SCAN_PORT_END} on {SCAN_TARGET_HOST} (using /health endpoint)")
+            print(f"Scan Interval: {SCAN_INTERVAL}s | Health Check Timeout: {HEALTH_CHECK_TIMEOUT}s")
+            print(table)
+            print(f"Total Active Connections (on available servers): {total_connections}")
+            print("------------------------------\n")
+        except Exception as e:
+            logging.error(f"Error printing status table: {e}")
+
+
+# --- Main Execution (main remains the same) ---
+async def main():
+    scan_task = asyncio.create_task(scan_and_update_servers())
+    status_task = asyncio.create_task(print_status_periodically())
+
+    server = await asyncio.start_server(
+        handle_client, LOAD_BALANCER_HOST, LOAD_BALANCER_PORT
+    )
+
+    addrs = ', '.join(str(sock.getsockname()) for sock in server.sockets)
+    logging.info(f'Load balancer serving on {addrs}')
+    logging.info(f'Dynamically discovering servers via HTTP /health on {SCAN_TARGET_HOST}:{SCAN_PORT_START}-{SCAN_PORT_END}')
+
+    async with server:
+        try:
+            await server.serve_forever()
+        except asyncio.CancelledError:
+            logging.info("Load balancer server shutting down.")
+        finally:
+            logging.info("Cancelling background tasks...")
+            scan_task.cancel()
+            status_task.cancel()
+            try:
+                await asyncio.gather(scan_task, status_task, return_exceptions=True)
+            except asyncio.CancelledError:
+                pass
+            logging.info("Background tasks finished.")
+
+def run_load_balancer():
+    # Make sure to install aiohttp: pip install aiohttp
+    try:
+        asyncio.run(main())
+    except KeyboardInterrupt:
+        logging.info("Shutdown requested by user.")
+    except Exception as e:
+         logging.critical(f"Critical error in main execution: {e}")
+
+if __name__ == "__main__":
+    run_load_balancer()