/** Returns the current status of the host named by {@code host}. */
@Override
public HostStatus getHostStatus(final String host) {
final Stat stat;
final ZooKeeperClient client = provider.get("getHostStatus");
try {
stat = client.exists(Paths.configHostId(host));
} catch (KeeperException e) {
throw new HeliosRuntimeException("Failed to check host status", e);
}
if (stat == null) {
return null;
}
final boolean up = checkHostUp(client, host);
final HostInfo hostInfo = getHostInfo(client, host);
final AgentInfo agentInfo = getAgentInfo(client, host);
final Map<JobId, Deployment> tasks = getTasks(client, host);
final Map<JobId, TaskStatus> statuses = getTaskStatuses(client, host);
final Map<String, String> environment = getEnvironment(client, host);
return HostStatus.newBuilder()
.setJobs(tasks)
.setStatuses(fromNullable(statuses).or(EMPTY_STATUSES))
.setHostInfo(hostInfo)
.setAgentInfo(agentInfo)
.setStatus(up ? UP : DOWN)
.setEnvironment(environment)
.build();
}
/** 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);
}
}
/** Deletes a job from ZooKeeper. Ensures that job is not currently running anywhere. */
@Override
public Job removeJob(final JobId id) throws JobDoesNotExistException, JobStillDeployedException {
log.info("removing job: id={}", id);
final ZooKeeperClient client = provider.get("removeJob");
final Job job = getJob(client, id);
if (job == null) {
throw new JobDoesNotExistException(id);
}
// TODO (dano): handle retry failures
try {
final ImmutableList.Builder<ZooKeeperOperation> operations = ImmutableList.builder();
final UUID jobCreationOperationId = getJobCreation(client, id);
if (jobCreationOperationId != null) {
operations.add(delete(Paths.configJobCreation(id, jobCreationOperationId)));
}
operations.add(
delete(Paths.configJobHosts(id)),
delete(Paths.configJobRefShort(id)),
delete(Paths.configJob(id)));
client.transaction(operations.build());
} catch (final NoNodeException e) {
throw new JobDoesNotExistException(id);
} catch (final NotEmptyException e) {
throw new JobStillDeployedException(id, listJobHosts(client, id));
} catch (final KeeperException e) {
throw new HeliosRuntimeException("removing job " + id + " failed", e);
}
return job;
}
/** Returns a {@link Map} of {@link JobId} to {@link Job} objects for all of the jobs known. */
@Override
public Map<JobId, Job> getJobs() {
log.debug("getting jobs");
final String folder = Paths.configJobs();
final ZooKeeperClient client = provider.get("getJobs");
try {
final List<String> ids;
try {
ids = client.getChildren(folder);
} catch (NoNodeException e) {
return Maps.newHashMap();
}
final Map<JobId, Job> descriptors = Maps.newHashMap();
for (final String id : ids) {
final JobId jobId = JobId.fromString(id);
final String path = Paths.configJob(jobId);
final byte[] data = client.getData(path);
final Job descriptor = parse(data, Job.class);
descriptors.put(descriptor.getId(), descriptor);
}
return descriptors;
} catch (KeeperException | IOException e) {
throw new HeliosRuntimeException("getting jobs failed", e);
}
}
/** Adds a job into the configuration. */
@Override
public void addJob(final Job job) throws JobExistsException {
log.info("adding job: {}", job);
final JobId id = job.getId();
final UUID operationId = UUID.randomUUID();
final String creationPath = Paths.configJobCreation(id, operationId);
final ZooKeeperClient client = provider.get("addJob");
try {
try {
client.ensurePath(Paths.historyJob(id));
client.transaction(
create(Paths.configJob(id), job),
create(Paths.configJobRefShort(id), id),
create(Paths.configJobHosts(id)),
create(creationPath));
} catch (final NodeExistsException e) {
if (client.exists(creationPath) != null) {
// The job was created, we're done here
return;
}
throw new JobExistsException(id.toString());
}
} catch (final KeeperException e) {
throw new HeliosRuntimeException("adding job " + job + " failed", e);
}
}
private Map<JobId, Deployment> getTasks(final ZooKeeperClient client, final String host) {
final Map<JobId, Deployment> jobs = Maps.newHashMap();
try {
final String folder = Paths.configHostJobs(host);
final List<String> jobIds;
try {
jobIds = client.getChildren(folder);
} catch (KeeperException.NoNodeException e) {
return null;
}
for (final String jobIdString : jobIds) {
final JobId jobId = JobId.fromString(jobIdString);
final String containerPath = Paths.configHostJob(host, jobId);
try {
final byte[] data = client.getData(containerPath);
final Task task = parse(data, Task.class);
jobs.put(jobId, Deployment.of(jobId, task.getGoal()));
} catch (KeeperException.NoNodeException ignored) {
log.debug("deployment config node disappeared: {}", jobIdString);
}
}
} catch (KeeperException | IOException e) {
throw new HeliosRuntimeException("getting deployment config failed", e);
}
return jobs;
}
/** 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;
}
/** Returns the current deployment state of {@code jobId} on {@code host}. */
@Override
public Deployment getDeployment(final String host, final JobId jobId) {
final String path = Paths.configHostJob(host, jobId);
final ZooKeeperClient client = provider.get("getDeployment");
try {
final byte[] data = client.getData(path);
final Task task = parse(data, Task.class);
return Deployment.of(jobId, task.getGoal());
} catch (KeeperException.NoNodeException e) {
return null;
} catch (KeeperException | IOException e) {
throw new HeliosRuntimeException("getting deployment failed", e);
}
}
/** Returns a list of the host names of the currently running masters. */
@Override
public List<String> getRunningMasters() {
final ZooKeeperClient client = provider.get("getRunningMasters");
try {
final List<String> masters = client.getChildren(Paths.statusMaster());
final ImmutableList.Builder<String> upMasters = ImmutableList.builder();
for (final String master : masters) {
if (client.exists(Paths.statusMasterUp(master)) != null) {
upMasters.add(master);
}
}
return upMasters.build();
} catch (KeeperException e) {
throw new HeliosRuntimeException("listing masters failed", e);
}
}
private boolean checkHostUp(final ZooKeeperClient client, final String host) {
try {
final Stat stat = client.exists(Paths.statusHostUp(host));
return stat != null;
} catch (KeeperException e) {
throw new HeliosRuntimeException("getting host " + host + " up status failed", e);
}
}
private void assertTaskExists(final ZooKeeperClient client, final String host, final JobId jobId)
throws JobNotDeployedException {
try {
client.getData(Paths.configHostJob(host, jobId));
} catch (NoNodeException e) {
throw new JobNotDeployedException(host, jobId);
} catch (KeeperException e) {
throw new HeliosRuntimeException(e);
}
}
private UUID getJobCreation(final ZooKeeperClient client, final JobId id) throws KeeperException {
final String parent = Paths.configHostJobCreationParent(id);
final List<String> children = client.getChildren(parent);
for (final String child : children) {
if (Paths.isConfigJobCreation(id, parent, child)) {
return Paths.configJobCreationId(id, parent, child);
}
}
return null;
}
private void assertHostExists(final ZooKeeperClient client, final String host)
throws HostNotFoundException {
try {
client.getData(Paths.configHost(host));
} catch (NoNodeException e) {
throw new HostNotFoundException(host, e);
} catch (KeeperException e) {
throw new HeliosRuntimeException(e);
}
}
private void assertJobExists(final ZooKeeperClient client, final JobId id)
throws JobDoesNotExistException {
try {
final String path = Paths.configJob(id);
if (client.stat(path) == null) {
throw new JobDoesNotExistException(id);
}
} catch (KeeperException e) {
throw new HeliosRuntimeException("checking job existence failed", e);
}
}
private <T> T tryGetEntity(
final ZooKeeperClient client, String path, TypeReference<T> type, String name) {
try {
final byte[] data = client.getData(path);
return Json.read(data, type);
} catch (NoNodeException e) {
return null;
} catch (KeeperException | IOException e) {
throw new HeliosRuntimeException("reading " + name + " info failed", e);
}
}
/** Given a jobId, returns the N most recent events in it's history in the cluster. */
@Override
public List<TaskStatusEvent> getJobHistory(final JobId jobId) throws JobDoesNotExistException {
final Job descriptor = getJob(jobId);
if (descriptor == null) {
throw new JobDoesNotExistException(jobId);
}
final ZooKeeperClient client = provider.get("getJobHistory");
final List<String> hosts;
try {
hosts = client.getChildren(Paths.historyJobHosts(jobId));
} catch (NoNodeException e) {
return emptyList();
} catch (KeeperException e) {
throw Throwables.propagate(e);
}
final List<TaskStatusEvent> jsEvents = Lists.newArrayList();
for (String host : hosts) {
final List<String> events;
try {
events = client.getChildren(Paths.historyJobHostEvents(jobId, host));
} catch (KeeperException e) {
throw Throwables.propagate(e);
}
for (String event : events) {
try {
byte[] data =
client.getData(Paths.historyJobHostEventsTimestamp(jobId, host, Long.valueOf(event)));
final TaskStatus status = Json.read(data, TaskStatus.class);
jsEvents.add(new TaskStatusEvent(status, Long.valueOf(event), host));
} catch (NoNodeException e) { // ignore, it went away before we read it
} catch (KeeperException | IOException e) {
throw Throwables.propagate(e);
}
}
}
return Ordering.from(EVENT_COMPARATOR).sortedCopy(jsEvents);
}
private Job getJob(final ZooKeeperClient client, final JobId id) {
final String path = Paths.configJob(id);
try {
final byte[] data = client.getData(path);
return Json.read(data, Job.class);
} catch (NoNodeException e) {
// Return null to indicate that the job does not exist
return null;
} catch (KeeperException | IOException e) {
throw new HeliosRuntimeException("getting job " + id + " failed", e);
}
}
private List<String> listJobHosts(final ZooKeeperClient client, final JobId jobId)
throws JobDoesNotExistException {
final List<String> hosts;
try {
hosts = client.getChildren(Paths.configJobHosts(jobId));
} catch (NoNodeException e) {
throw new JobDoesNotExistException(jobId);
} catch (KeeperException e) {
throw new HeliosRuntimeException("failed to list hosts for job: " + jobId, e);
}
return hosts;
}
@Nullable
private TaskStatus getTaskStatus(
final ZooKeeperClient client, final String host, final JobId jobId) {
final String containerPath = Paths.statusHostJob(host, jobId);
try {
final byte[] data = client.getData(containerPath);
return parse(data, TaskStatus.class);
} catch (NoNodeException ignored) {
return null;
} catch (KeeperException | IOException e) {
throw new HeliosRuntimeException(
"Getting task " + jobId + " status " + "for host " + host + " failed", e);
}
}
private List<JobId> listHostJobs(final ZooKeeperClient client, final String host) {
final List<String> jobIdStrings;
final String folder = Paths.statusHostJobs(host);
try {
jobIdStrings = client.getChildren(folder);
} catch (KeeperException.NoNodeException e) {
return null;
} catch (KeeperException e) {
throw new HeliosRuntimeException("List tasks for host failed: " + host, e);
}
final ImmutableList.Builder<JobId> jobIds = ImmutableList.builder();
for (String jobIdString : jobIdStrings) {
jobIds.add(JobId.fromString(jobIdString));
}
return jobIds.build();
}
/**
* Registers a host into ZooKeeper. The {@code id} is initially generated randomly by the Agent
* and persisted on disk. This way, in the event that you have two agents attempting to register
* with the same value of @{code host}, the first one will win.
*/
@Override
public void registerHost(final String host, final String id) {
log.info("registering host: {}", host);
final ZooKeeperClient client = provider.get("registerHost");
try {
// TODO (dano): this code is replicated in AgentZooKeeperRegistrar
// This would've been nice to do in a transaction but PathChildrenCache ensures paths
// so we can't know what paths already exist so assembling a suitable transaction is too
// painful.
client.ensurePath(Paths.configHost(host));
client.ensurePath(Paths.configHostJobs(host));
client.ensurePath(Paths.configHostPorts(host));
client.ensurePath(Paths.statusHost(host));
client.ensurePath(Paths.statusHostJobs(host));
// Finish registration by creating the id node last
client.createAndSetData(Paths.configHostId(host), id.getBytes(UTF_8));
} catch (Exception e) {
throw new HeliosRuntimeException("registering host " + host + " failed", e);
}
}
/**
* Undoes the effect of {@link ZooKeeperMasterModel#registerHost(String, String)}. Cleans up any
* leftover host-related things.
*/
@Override
public void deregisterHost(final String host)
throws HostNotFoundException, HostStillInUseException {
log.info("deregistering host: {}", host);
final ZooKeeperClient client = provider.get("deregisterHost");
// TODO (dano): handle retry failures
try {
final List<ZooKeeperOperation> operations = Lists.newArrayList();
// Remove all jobs deployed to this host
final List<JobId> jobs = listHostJobs(client, host);
if (jobs == null) {
if (client.exists(Paths.configHost(host)) == null) {
throw new HostNotFoundException("host [" + host + "] does not exist");
}
}
for (JobId job : jobs) {
final String hostJobPath = Paths.configHostJob(host, job);
final List<String> nodes = client.listRecursive(hostJobPath);
for (final String node : reverse(nodes)) {
operations.add(delete(node));
}
if (client.exists(Paths.configJobHost(job, host)) != null) {
operations.add(delete(Paths.configJobHost(job, host)));
}
// Clean out the history for each job
try {
final List<String> history = client.listRecursive(Paths.historyJobHost(job, host));
for (String s : reverse(history)) {
operations.add(delete(s));
}
} catch (NoNodeException ignore) {
}
}
operations.add(delete(Paths.configHostJobs(host)));
// Remove the host status
try {
final List<String> nodes = client.listRecursive(Paths.statusHost(host));
for (final String node : reverse(nodes)) {
operations.add(delete(node));
}
} catch (NoNodeException ignore) {
}
// Remove port allocations
try {
final List<String> ports = client.getChildren(Paths.configHostPorts(host));
for (final String port : ports) {
operations.add(delete(Paths.configHostPort(host, Integer.valueOf(port))));
}
operations.add(delete(Paths.configHostPorts(host)));
} catch (NoNodeException ignore) {
}
// Remove host id
String idPath = Paths.configHostId(host);
if (client.exists(idPath) != null) {
operations.add(delete(idPath));
}
// Remove host config root
operations.add(delete(Paths.configHost(host)));
client.transaction(operations);
} catch (NotEmptyException e) {
final HostStatus hostStatus = getHostStatus(host);
final List<JobId> jobs =
hostStatus != null
? ImmutableList.copyOf(hostStatus.getJobs().keySet())
: Collections.<JobId>emptyList();
throw new HostStillInUseException(host, jobs);
} catch (NoNodeException e) {
throw new HostNotFoundException(host);
} catch (KeeperException e) {
throw new HeliosRuntimeException(e);
}
}
// TODO(drewc): this kinda screams "long method"
private void deployJobRetry(
final ZooKeeperClient client, final String host, final Deployment deployment, int count)
throws JobDoesNotExistException, JobAlreadyDeployedException, HostNotFoundException,
JobPortAllocationConflictException {
if (count == 3) {
throw new HeliosRuntimeException(
"3 failures (possibly concurrent modifications) while " + "deploying. Giving up.");
}
log.info("deploying {}: {} (retry={})", deployment, host, count);
final JobId id = deployment.getJobId();
final Job job = getJob(id);
if (job == null) {
throw new JobDoesNotExistException(id);
}
final UUID operationId = UUID.randomUUID();
final String jobPath = Paths.configJob(id);
final String taskPath = Paths.configHostJob(host, id);
final String taskCreationPath = Paths.configHostJobCreation(host, id, operationId);
final List<Integer> staticPorts = staticPorts(job);
final Map<String, byte[]> portNodes = Maps.newHashMap();
final byte[] idJson = id.toJsonBytes();
for (final int port : staticPorts) {
final String path = Paths.configHostPort(host, port);
portNodes.put(path, idJson);
}
final Task task = new Task(job, deployment.getGoal());
final List<ZooKeeperOperation> operations =
Lists.newArrayList(
check(jobPath), create(portNodes), create(Paths.configJobHost(id, host)));
// Attempt to read a task here. If it's goal is UNDEPLOY, it's as good as not existing
try {
final Node existing = client.getNode(taskPath);
byte[] bytes = existing.getBytes();
Task readTask = Json.read(bytes, Task.class);
if (readTask.getGoal() != Goal.UNDEPLOY) {
throw new JobAlreadyDeployedException(host, id);
}
operations.add(check(taskPath, existing.getStat().getVersion()));
operations.add(set(taskPath, task));
} catch (NoNodeException e) {
operations.add(create(taskPath, task));
operations.add(create(taskCreationPath));
} catch (IOException | KeeperException e) {
throw new HeliosRuntimeException("reading existing task description failed", e);
}
// TODO (dano): Failure handling is racy wrt agent and job modifications.
try {
client.transaction(operations);
log.info("deployed {}: {} (retry={})", deployment, host, count);
} catch (NoNodeException e) {
// Either the job, the host or the task went away
assertJobExists(client, id);
assertHostExists(client, host);
// If the job and host still exists, we likely tried to redeploy a job that had an UNDEPLOY
// goal and lost the race with the agent removing the task before we could set it. Retry.
deployJobRetry(client, host, deployment, count + 1);
} catch (NodeExistsException e) {
// Check for conflict due to transaction retry
try {
if (client.exists(taskCreationPath) != null) {
// Our creation operation node existed, we're done here
return;
}
} catch (KeeperException ex) {
throw new HeliosRuntimeException("checking job deployment failed", ex);
}
try {
// Check if the job was already deployed
if (client.stat(taskPath) != null) {
throw new JobAlreadyDeployedException(host, id);
}
} catch (KeeperException ex) {
throw new HeliosRuntimeException("checking job deployment failed", e);
}
// Check for static port collisions
for (final int port : staticPorts) {
final String path = Paths.configHostPort(host, port);
try {
if (client.stat(path) == null) {
continue;
}
final byte[] b = client.getData(path);
final JobId existingJobId = parse(b, JobId.class);
throw new JobPortAllocationConflictException(id, existingJobId, host, port);
} catch (KeeperException | IOException ex) {
throw new HeliosRuntimeException("checking port allocations failed", e);
}
}
// Catch all for logic and ephemeral issues
throw new HeliosRuntimeException("deploying job failed", e);
} catch (KeeperException e) {
throw new HeliosRuntimeException("deploying job failed", e);
}
}