diff options
Diffstat (limited to 'python/pyspark/daemon.py')
-rw-r--r-- | python/pyspark/daemon.py | 42 |
1 files changed, 24 insertions, 18 deletions
diff --git a/python/pyspark/daemon.py b/python/pyspark/daemon.py index 78a2da1e18..78c9457b84 100644 --- a/python/pyspark/daemon.py +++ b/python/pyspark/daemon.py @@ -1,10 +1,13 @@ import os +import signal +import socket import sys +import traceback import multiprocessing from ctypes import c_bool from errno import EINTR, ECHILD -from socket import socket, AF_INET, SOCK_STREAM, SOMAXCONN -from signal import signal, SIGHUP, SIGTERM, SIGCHLD, SIG_DFL, SIG_IGN +from socket import AF_INET, SOCK_STREAM, SOMAXCONN +from signal import SIGHUP, SIGTERM, SIGCHLD, SIG_DFL, SIG_IGN from pyspark.worker import main as worker_main from pyspark.serializers import write_int @@ -33,11 +36,12 @@ def compute_real_exit_code(exit_code): def worker(listen_sock): # Redirect stdout to stderr os.dup2(2, 1) + sys.stdout = sys.stderr # The sys.stdout object is different from file descriptor 1 # Manager sends SIGHUP to request termination of workers in the pool def handle_sighup(*args): assert should_exit() - signal(SIGHUP, handle_sighup) + signal.signal(SIGHUP, handle_sighup) # Cleanup zombie children def handle_sigchld(*args): @@ -51,7 +55,7 @@ def worker(listen_sock): handle_sigchld() elif err.errno != ECHILD: raise - signal(SIGCHLD, handle_sigchld) + signal.signal(SIGCHLD, handle_sigchld) # Handle clients while not should_exit(): @@ -70,19 +74,22 @@ def worker(listen_sock): # never receives SIGCHLD unless a worker crashes. if os.fork() == 0: # Leave the worker pool - signal(SIGHUP, SIG_DFL) + signal.signal(SIGHUP, SIG_DFL) listen_sock.close() - # Handle the client then exit - sockfile = sock.makefile() + # Read the socket using fdopen instead of socket.makefile() because the latter + # seems to be very slow; note that we need to dup() the file descriptor because + # otherwise writes also cause a seek that makes us miss data on the read side. + infile = os.fdopen(os.dup(sock.fileno()), "a+", 65536) + outfile = os.fdopen(os.dup(sock.fileno()), "a+", 65536) exit_code = 0 try: - worker_main(sockfile, sockfile) + worker_main(infile, outfile) except SystemExit as exc: - exit_code = exc.code + exit_code = exc.code finally: - sockfile.close() - sock.close() - os._exit(compute_real_exit_code(exit_code)) + outfile.flush() + sock.close() + os._exit(compute_real_exit_code(exit_code)) else: sock.close() @@ -92,7 +99,6 @@ def launch_worker(listen_sock): try: worker(listen_sock) except Exception as err: - import traceback traceback.print_exc() os._exit(1) else: @@ -105,7 +111,7 @@ def manager(): os.setpgid(0, 0) # Create a listening socket on the AF_INET loopback interface - listen_sock = socket(AF_INET, SOCK_STREAM) + listen_sock = socket.socket(AF_INET, SOCK_STREAM) listen_sock.bind(('127.0.0.1', 0)) listen_sock.listen(max(1024, 2 * POOLSIZE, SOMAXCONN)) listen_host, listen_port = listen_sock.getsockname() @@ -121,8 +127,8 @@ def manager(): exit_flag.value = True # Gracefully exit on SIGTERM, don't die on SIGHUP - signal(SIGTERM, lambda signum, frame: shutdown()) - signal(SIGHUP, SIG_IGN) + signal.signal(SIGTERM, lambda signum, frame: shutdown()) + signal.signal(SIGHUP, SIG_IGN) # Cleanup zombie children def handle_sigchld(*args): @@ -133,7 +139,7 @@ def manager(): except EnvironmentError as err: if err.errno not in (ECHILD, EINTR): raise - signal(SIGCHLD, handle_sigchld) + signal.signal(SIGCHLD, handle_sigchld) # Initialization complete sys.stdout.close() @@ -148,7 +154,7 @@ def manager(): shutdown() raise finally: - signal(SIGTERM, SIG_DFL) + signal.signal(SIGTERM, SIG_DFL) exit_flag.value = True # Send SIGHUP to notify workers of shutdown os.kill(0, SIGHUP) |