/** Used to update the existing deployment of a job. */
@Override
public void updateDeployment(final String host, final Deployment deployment)
throws HostNotFoundException, JobNotDeployedException {
log.info("updating deployment {}: {}", deployment, host);
final ZooKeeperClient client = provider.get("updateDeployment");
final JobId jobId = deployment.getJobId();
final Job job = getJob(client, jobId);
if (job == null) {
throw new JobNotDeployedException(host, jobId);
}
assertHostExists(client, host);
assertTaskExists(client, host, deployment.getJobId());
final String path = Paths.configHostJob(host, jobId);
final Task task = new Task(job, deployment.getGoal());
try {
client.setData(path, task.toJsonBytes());
} catch (Exception e) {
throw new HeliosRuntimeException(
"updating deployment " + deployment + " on host " + host + " failed", e);
}
}
/** Undeploys the job specified by {@code jobId} on {@code host}. */
@Override
public Deployment undeployJob(final String host, final JobId jobId)
throws HostNotFoundException, JobNotDeployedException {
log.info("undeploying {}: {}", jobId, host);
final ZooKeeperClient client = provider.get("undeployJob");
assertHostExists(client, host);
final Deployment deployment = getDeployment(host, jobId);
if (deployment == null) {
throw new JobNotDeployedException(host, jobId);
}
// TODO (dano): Is this safe? can the ports of an undeployed job collide with a new deployment?
// TODO (drewc): If it's still in UNDEPLOY, that means the agent hasn't gotten to it
// yet, which means it probably won't see the new job yet either. However, it may spin up
// a new supervisor for the new job before the old one is done being torn down. So it can
// race and lose. With a little change to the Agent where we manage the supervisors, with
// some coordination, we could remove the race, such that this race goes away. Specifically,
// we're creating new Supervisors before updating existing ones. If we swap that, part of
// the problem goes away, but we'd need some coordination between the Supervisor and the
// agent such that the Agent could wait until the Supervisor had handled the Goal change.
// Additionally, since ZK guarantees we'll see the writes in the proper order, we wouldn't
// need to deal with seeing the new job before the UNDEPLOY.
final Job job = getJob(client, jobId);
final String path = Paths.configHostJob(host, jobId);
final Task task = new Task(job, UNDEPLOY);
final List<ZooKeeperOperation> operations =
Lists.newArrayList(set(path, task.toJsonBytes()), delete(Paths.configJobHost(jobId, host)));
final List<Integer> staticPorts = staticPorts(job);
for (int port : staticPorts) {
operations.add(delete(Paths.configHostPort(host, port)));
}
try {
client.transaction(operations);
} catch (NoNodeException e) {
if (e.getPath().equals(path)) {
// NoNodeException on updating the deployment node may happen due to retry failures.
// If the deployment isn't there anymore, we're done.
return deployment;
} else {
// The relation node deletes should not fail unless there is a programming error.
throw new HeliosRuntimeException("Removing deployment failed", e);
}
} catch (KeeperException e) {
throw new HeliosRuntimeException("Removing deployment failed", e);
}
return deployment;
}
