1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
|
import os
import sys
import multiprocessing
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 pyspark.worker import main as worker_main
from pyspark.serializers import write_int
try:
POOLSIZE = multiprocessing.cpu_count()
except NotImplementedError:
POOLSIZE = 4
should_exit = False
def worker(listen_sock):
# Redirect stdout to stderr
os.dup2(2, 1)
# Manager sends SIGHUP to request termination of workers in the pool
def handle_sighup(signum, frame):
global should_exit
should_exit = True
signal(SIGHUP, handle_sighup)
while not should_exit:
# Wait until a client arrives or we have to exit
sock = None
while not should_exit and sock is None:
try:
sock, addr = listen_sock.accept()
except EnvironmentError as err:
if err.errno != EINTR:
raise
if sock is not None:
# Fork a child to handle the client
if os.fork() == 0:
# Leave the worker pool
signal(SIGHUP, SIG_DFL)
listen_sock.close()
# Handle the client then exit
sockfile = sock.makefile()
worker_main(sockfile, sockfile)
sockfile.close()
sock.close()
os._exit(0)
else:
sock.close()
assert should_exit
os._exit(0)
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(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()
write_int(listen_port, sys.stdout)
# Launch initial worker pool
for idx in range(POOLSIZE):
if os.fork() == 0:
worker(listen_sock)
raise RuntimeError("worker() unexpectedly returned")
listen_sock.close()
def shutdown():
global should_exit
os.kill(0, SIGHUP)
should_exit = True
# Gracefully exit on SIGTERM, don't die on SIGHUP
signal(SIGTERM, lambda signum, frame: shutdown())
signal(SIGHUP, SIG_IGN)
# Cleanup zombie children
def handle_sigchld(signum, frame):
try:
pid, status = os.waitpid(0, os.WNOHANG)
if (pid, status) != (0, 0) and not should_exit:
raise RuntimeError("pool member crashed: %s, %s" % (pid, status))
except EnvironmentError as err:
if err.errno not in (ECHILD, EINTR):
raise
signal(SIGCHLD, handle_sigchld)
# Initialization complete
sys.stdout.close()
while not should_exit:
try:
# Spark tells us to exit by closing stdin
if sys.stdin.read() == '':
shutdown()
except EnvironmentError as err:
if err.errno != EINTR:
shutdown()
raise
if __name__ == '__main__':
manager()
|
