#!/usr/bin/env python
+# arv-web enables you to run a custom web service from the contents of an Arvados collection.
+#
+# See http://doc.arvados.org/user/topics/arv-web.html
+
import arvados
import subprocess
from arvados_fuse import Operations, SafeApi, CollectionDirectory
import sys
import functools
+# Run an arvados_fuse mount under the control of the local process. This lets
+# us switch out the contents of the directory without having to unmount and
+# remount.
def run_fuse_mount(api, collection):
mountdir = tempfile.mkdtemp()
return (mountdir, cdir)
+# Handle messages from Arvados event bus.
def on_message(project, evqueue, ev):
if 'event_type' in ev:
old_attr = None
parser = argparse.ArgumentParser()
parser.add_argument('--project-uuid', type=str, required=True, help="Project uuid to watch")
- parser.add_argument('--port', type=int, default=8080, help="Port to listen on (default 8080)")
+ parser.add_argument('--port', type=int, default=8080, help="Host port to listen on (default 8080)")
parser.add_argument('--image', type=str, help="Docker image to run")
args = parser.parse_args(argv)
loop = True
cid = None
+ docker_proc = None
prev_docker_image = None
mountdir = None
with llfuse.lock:
cdir.clear()
if collection:
+ # Switch the FUSE directory object so that it stores
+ # the newly selected collection
logger.info("Mounting %s", collection)
cdir.collection_locator = collection
cdir.collection_object = None
- cdir.clear()
+ cdir.update()
try:
try:
if collection:
if not args.image:
docker_image = None
+
+ # FUSE is asynchronous, so there is a race between
+ # the directory being updated above and the kernel
+ # cache being refreshed. This manifests as the
+ # bizare behavior where os.path.exists() returns
+ # True, but open() raises "file not found". The
+ # workaround is to keep trying until the kernel
+ # catches up.
while not docker_image and os.path.exists(os.path.join(mountdir, "docker_image")):
try:
with open(os.path.join(mountdir, "docker_image")) as di:
if not docker_image:
logger.error("Collection must contain a file 'docker_image' or must specify --image on the command line.")
- if docker_image != prev_docker_image:
+ if docker_image and ((docker_image != prev_docker_image) or cid is None):
if cid:
- logger.info("Stopping docker container")
+ logger.info("Stopping Docker container")
subprocess.check_call(["docker", "stop", cid])
+ cid = None
+ docker_proc = None
if docker_image:
- logger.info("Starting docker container %s", docker_image)
- cid = subprocess.check_output(["docker", "run",
- "--detach=true",
- "--publish=%i:80" % (port),
- "--volume=%s:/mnt:ro" % mountdir,
- docker_image])
- cid = cid.rstrip()
+ logger.info("Starting Docker container %s", docker_image)
+ ciddir = tempfile.mkdtemp()
+ cidfilepath = os.path.join(ciddir, "cidfile")
+ docker_proc = subprocess.Popen(["docker", "run",
+ "--cidfile=%s" % (cidfilepath),
+ "--publish=%i:80" % (port),
+ "--volume=%s:/mnt:ro" % mountdir,
+ docker_image])
+ cid = None
+ while not cid and docker_proc.poll() is None:
+ try:
+ with open(cidfilepath) as cidfile:
+ cid = cidfile.read().strip()
+ except IOError:
+ pass
+ try:
+ os.unlink(cidfilepath)
+ os.rmdir(ciddir)
+ except OSError:
+ pass
+
prev_docker_image = docker_image
- logger.info("Container id is %s", cid)
+ logger.info("Container id %s", cid)
elif cid:
- subprocess.check_call(["docker", "kill", "--signal=HUP", cid])
+ logger.info("Sending refresh signal to container")
+ # Send SIGHUP to all the processes inside the
+ # container. By convention, services are expected
+ # to reload their configuration. If they die
+ # instead, that's okay, because then we'll just
+ # start a new container.
+ #
+ # Getting the services inside the container to
+ # refresh turned out to be really hard. Here are
+ # some of the other things I tried:
+ #
+ # docker kill --signal=HUP # no effect
+ # docker_proc.send_signal(signal.SIGHUP) # no effect
+ # os.killpg(os.getpgid(docker_proc.pid), signal.SIGHUP) # docker-proxy dies as collatoral damage
+ # docker exec apache2ctl restart # only works if service is using apache.
+ # Sending HUP directly to the processes inside the container: permission denied
+
+ subprocess.check_call(["docker", "exec", cid, "killall", "--regexp", ".*", "--signal", "HUP"])
elif cid:
logger.info("Stopping docker container")
- subprocess.call(["docker", "stop", cid])
+ subprocess.check_call(["docker", "stop", cid])
except subprocess.CalledProcessError:
cid = None
logger.warning("No service running! Will wait for a new collection to appear in the project.")
else:
logger.info("Waiting for events")
+
running = True
loop = True
while running:
+ # Main run loop. Wait on project events, signals, or the
+ # Docker container stopping.
+
try:
+ # Poll the queue with a 1 second timeout, if we have no
+ # timeout the Python runtime doesn't have a chance to
+ # process SIGINT or SIGTERM.
eq = evqueue.get(True, 1)
logger.info("%s %s", eq[1], eq[2])
newcollection = collection
except Queue.Empty:
pass
+ if docker_proc and docker_proc.poll() is not None:
+ logger.warning("Service has terminated. Will try to restart.")
+ cid = None
+ docker_proc = None
+ running = False
+
except (KeyboardInterrupt):
logger.info("Got keyboard interrupt")
ws.close()
finally:
if cid:
logger.info("Stopping docker container")
- cid = subprocess.call(["docker", "stop", cid])
+ subprocess.check_call(["docker", "stop", cid])
if mountdir:
logger.info("Unmounting")