spark-instrumented-optimizer/python/pyspark/daemon.py
Davies Liu e595c8d08a [SPARK-3993] [PySpark] fix bug while reuse worker after take()
After take(), maybe there are some garbage left in the socket, then next task assigned to this worker will hang because of corrupted data.

We should make sure the socket is clean before reuse it, write END_OF_STREAM at the end, and check it after read out all result from python.

Author: Davies Liu <davies.liu@gmail.com>
Author: Davies Liu <davies@databricks.com>

Closes #2838 from davies/fix_reuse and squashes the following commits:

8872914 [Davies Liu] fix tests
660875b [Davies Liu] fix bug while reuse worker after take()
2014-10-23 17:20:00 -07:00

184 lines
5.9 KiB
Python

#
# Licensed to the Apache Software Foundation (ASF) under one or more
# contributor license agreements. See the NOTICE file distributed with
# this work for additional information regarding copyright ownership.
# The ASF licenses this file to You under the Apache License, Version 2.0
# (the "License"); you may not use this file except in compliance with
# the License. You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
import numbers
import os
import signal
import select
import socket
import sys
import traceback
import time
import gc
from errno import EINTR, ECHILD, EAGAIN
from socket import AF_INET, SOCK_STREAM, SOMAXCONN
from signal import SIGHUP, SIGTERM, SIGCHLD, SIG_DFL, SIG_IGN, SIGINT
from pyspark.worker import main as worker_main
from pyspark.serializers import read_int, write_int
def compute_real_exit_code(exit_code):
# SystemExit's code can be integer or string, but os._exit only accepts integers
if isinstance(exit_code, numbers.Integral):
return exit_code
else:
return 1
def worker(sock):
"""
Called by a worker process after the fork().
"""
signal.signal(SIGHUP, SIG_DFL)
signal.signal(SIGCHLD, SIG_DFL)
signal.signal(SIGTERM, SIG_DFL)
# restore the handler for SIGINT,
# it's useful for debugging (show the stacktrace before exit)
signal.signal(SIGINT, signal.default_int_handler)
# 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(infile, outfile)
except SystemExit as exc:
exit_code = compute_real_exit_code(exc.code)
finally:
outfile.flush()
if exit_code:
os._exit(exit_code)
# Cleanup zombie children
def cleanup_dead_children():
try:
while True:
pid, _ = os.waitpid(0, os.WNOHANG)
if not pid:
break
except:
pass
def manager():
# Create a new process group to corral our children
os.setpgid(0, 0)
# Create a listening socket on the AF_INET loopback interface
listen_sock = socket.socket(AF_INET, SOCK_STREAM)
listen_sock.bind(('127.0.0.1', 0))
listen_sock.listen(max(1024, SOMAXCONN))
listen_host, listen_port = listen_sock.getsockname()
write_int(listen_port, sys.stdout)
sys.stdout.flush()
def shutdown(code):
signal.signal(SIGTERM, SIG_DFL)
# Send SIGHUP to notify workers of shutdown
os.kill(0, SIGHUP)
exit(code)
def handle_sigterm(*args):
shutdown(1)
signal.signal(SIGTERM, handle_sigterm) # Gracefully exit on SIGTERM
signal.signal(SIGHUP, SIG_IGN) # Don't die on SIGHUP
reuse = os.environ.get("SPARK_REUSE_WORKER")
# Initialization complete
try:
while True:
try:
ready_fds = select.select([0, listen_sock], [], [], 1)[0]
except select.error as ex:
if ex[0] == EINTR:
continue
else:
raise
# cleanup in signal handler will cause deadlock
cleanup_dead_children()
if 0 in ready_fds:
try:
worker_pid = read_int(sys.stdin)
except EOFError:
# Spark told us to exit by closing stdin
shutdown(0)
try:
os.kill(worker_pid, signal.SIGKILL)
except OSError:
pass # process already died
if listen_sock in ready_fds:
try:
sock, _ = listen_sock.accept()
except OSError as e:
if e.errno == EINTR:
continue
raise
# Launch a worker process
try:
pid = os.fork()
except OSError as e:
if e.errno in (EAGAIN, EINTR):
time.sleep(1)
pid = os.fork() # error here will shutdown daemon
else:
outfile = sock.makefile('w')
write_int(e.errno, outfile) # Signal that the fork failed
outfile.flush()
outfile.close()
sock.close()
continue
if pid == 0:
# in child process
listen_sock.close()
try:
# Acknowledge that the fork was successful
outfile = sock.makefile("w")
write_int(os.getpid(), outfile)
outfile.flush()
outfile.close()
while True:
worker(sock)
if not reuse:
# wait for closing
while sock.recv(1024):
pass
break
gc.collect()
except:
traceback.print_exc()
os._exit(1)
else:
os._exit(0)
else:
sock.close()
finally:
shutdown(1)
if __name__ == '__main__':
manager()