python/pyspark/daemon.py - master - githubmirror/spark

python/pyspark/daemon.py
292 строки12 KB
Dongjoon Hyun
[SPARK-57978][PYTHON] Use explicit UTF-8 encoding in text-mode I/O to prepare for PEP 686
07 июл 2026, 07:43
07 июл 2026, 07:43aeadc06
#
# 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 uuid
import os
import signal
import select
import socket
import sys
import traceback
import time
import gc
import faulthandler
from errno import EINTR, EAGAIN
from socket import AF_INET, AF_INET6, SOCK_STREAM, SOMAXCONN
from signal import SIGHUP, SIGTERM, SIGCHLD, SIG_DFL, SIG_IGN, SIGINT
from types import FrameType
from typing import Any, Optional

from pyspark.serializers import read_int, write_int, write_with_length, UTF8Deserializer
from pyspark.util import enable_faulthandler
from pyspark.errors import PySparkRuntimeError


def compute_real_exit_code(exit_code: Any) -> int:
    # SystemExit's code can be anything, but os._exit only accepts integer
    if isinstance(exit_code, int):
        return exit_code
    else:
        return 1


def worker(sock: socket.socket, authenticated: bool) -> int:
    """
    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.
    buffer_size = int(os.environ.get("SPARK_BUFFER_SIZE", 65536))
    infile = os.fdopen(os.dup(sock.fileno()), "rb", buffer_size)
    outfile = os.fdopen(os.dup(sock.fileno()), "wb", buffer_size)

    if not authenticated:
        client_secret = UTF8Deserializer().loads(infile)
        if os.environ["PYTHON_WORKER_FACTORY_SECRET"] == client_secret:
            write_with_length("ok".encode("utf-8"), outfile)
            outfile.flush()
        else:
            write_with_length("err".encode("utf-8"), outfile)
            outfile.flush()
            sock.close()
            return 1

    exit_code = 0

    # We don't know what could happen when we import the worker module. We have to
    # guarantee that no thread is spawned before we fork, so we have to import the
    # worker module after fork. For example, both pandas and pyarrow starts some
    # threads when they are imported.
    if len(sys.argv) > 1 and sys.argv[1].startswith("pyspark"):
        import importlib

        worker_module = importlib.import_module(sys.argv[1])
        worker_main = worker_module.main
    else:
        from pyspark.worker import main as worker_main

    try:
        worker_main(infile, outfile)
    except SystemExit as exc:
        exit_code = compute_real_exit_code(exc.code)
    finally:
        try:
            outfile.flush()
        except Exception:
            if os.environ.get("PYTHON_DAEMON_KILL_WORKER_ON_FLUSH_FAILURE", False):
                faulthandler_log_path = os.environ.get("PYTHON_FAULTHANDLER_DIR", None)
                if faulthandler_log_path:
                    faulthandler_log_path = os.path.join(faulthandler_log_path, str(os.getpid()))
                    with open(
                        faulthandler_log_path, "w", encoding="utf-8"
                    ) as faulthandler_log_file:
                        faulthandler.dump_traceback(file=faulthandler_log_file)
                raise
            else:
                print(
                    "PySpark daemon failed to flush the output to the worker process:\n"
                    + traceback.format_exc(),
                    file=sys.stderr,
                )
    return exit_code


def manager() -> None:
    # Create a new process group to corral our children
    os.setpgid(0, 0)

    is_unix_domain_sock = os.environ.get("PYTHON_UNIX_DOMAIN_ENABLED", "false").lower() == "true"
    socket_path = None

    # Create a listening socket on the loopback interface
    if is_unix_domain_sock:
        assert "PYTHON_WORKER_FACTORY_SOCK_DIR" in os.environ
        socket_path = os.path.join(
            os.environ["PYTHON_WORKER_FACTORY_SOCK_DIR"], f".{uuid.uuid4()}.sock"
        )
        listen_sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
        listen_sock.bind(socket_path)
        listen_sock.listen(max(1024, SOMAXCONN))
        listen_port = socket_path
    elif os.environ.get("SPARK_PREFER_IPV6", "false").lower() == "true":
        listen_sock = socket.socket(AF_INET6, SOCK_STREAM)
        listen_sock.bind(("::1", 0, 0, 0))
        listen_sock.listen(max(1024, SOMAXCONN))
        listen_host, listen_port, _, _ = listen_sock.getsockname()
    else:
        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()

    # re-open stdin/stdout in 'wb' mode
    stdin_bin = os.fdopen(sys.stdin.fileno(), "rb", 4)
    stdout_bin = os.fdopen(sys.stdout.fileno(), "wb", 4)
    if is_unix_domain_sock:
        write_with_length(listen_port.encode("utf-8"), stdout_bin)
    else:
        write_int(listen_port, stdout_bin)
    stdout_bin.flush()

    def shutdown(code: int) -> None:
        if socket_path is not None and os.path.exists(socket_path):
            os.remove(socket_path)
        signal.signal(SIGTERM, SIG_DFL)
        # Send SIGHUP to notify workers of shutdown
        os.kill(0, SIGHUP)
        sys.exit(code)

    def handle_sigterm(signal_number: int, frame: Optional[FrameType]) -> None:
        shutdown(1)

    signal.signal(SIGTERM, handle_sigterm)  # Gracefully exit on SIGTERM
    signal.signal(SIGHUP, SIG_IGN)  # Don't die on SIGHUP
    signal.signal(SIGCHLD, SIG_IGN)

    reuse = os.environ.get("SPARK_REUSE_WORKER")

    # Initialization complete
    try:
        poller = None
        if os.name == "posix":
            # select.select has a known limit on the number of file descriptors
            # it can handle. We use select.poll instead to avoid this limit.
            poller = select.poll()
            fd_reverse_map = {0: 0, listen_sock.fileno(): listen_sock}
            poller.register(0, select.POLLIN)
            poller.register(listen_sock, select.POLLIN)

        while True:
            if poller is not None:
                ready_fds = []
                # Unlike select, poll timeout is in millis.
                for fd, event in poller.poll(1000):
                    if event & (select.POLLIN | select.POLLHUP):
                        # Data can be read (for POLLHUP peer hang up, so reads will return
                        # 0 bytes, in which case we want to break out - this is consistent
                        # with how select behaves).
                        ready_fds.append(fd_reverse_map[fd])
                    else:
                        # Could be POLLERR or POLLNVAL (select would raise in this case).
                        raise PySparkRuntimeError(f"Polling error - event {event} on fd {fd}")
            else:
                # If poll is not available, use select.
                ready_fds = select.select([0, listen_sock], [], [], 1)[0]

            if 0 in ready_fds:
                try:
                    worker_pid = read_int(stdin_bin)
                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(mode="wb")
                        write_int(e.errno, outfile)  # Signal that the fork failed
                        outfile.flush()
                        outfile.close()
                        sock.close()
                        continue

                if pid == 0:
                    # in child process
                    with enable_faulthandler():
                        if poller is not None:
                            poller.unregister(0)
                            poller.unregister(listen_sock)
                        listen_sock.close()

                        # It should close the standard input in the child process so that
                        # Python native function executions stay intact.
                        #
                        # Note that if we just close the standard input (file descriptor 0),
                        # the lowest file descriptor (file descriptor 0) will be allocated,
                        # later when other file descriptors should happen to open.
                        #
                        # Therefore, here we redirects it to '/dev/null' by duplicating
                        # another file descriptor for '/dev/null' to the standard input (0).
                        # See SPARK-26175.
                        devnull = open(os.devnull, "r", encoding="utf-8")
                        os.dup2(devnull.fileno(), 0)
                        devnull.close()

                        try:
                            # Acknowledge that the fork was successful
                            outfile = sock.makefile(mode="wb")
                            write_int(os.getpid(), outfile)
                            outfile.flush()
                            outfile.close()
                            authenticated = (
                                os.environ.get("PYTHON_UNIX_DOMAIN_ENABLED", "false").lower()
                                == "true"
                            )
                            while True:
                                code = worker(sock, authenticated)
                                if code == 0:
                                    authenticated = True
                                if not reuse or code:
                                    # wait for closing
                                    try:
                                        while sock.recv(1024):
                                            pass
                                    except Exception:
                                        pass
                                    break
                                gc.collect()
                        except BaseException:
                            traceback.print_exc()
                            os._exit(1)
                        else:
                            os._exit(0)
                else:
                    sock.close()

    finally:
        if poller is not None:
            poller.unregister(0)
            poller.unregister(listen_sock)
        shutdown(1)


if __name__ == "__main__":
    manager()