class _QueuedOpCode(object):
- """Encasulates an opcode object.
+ """Encapsulates an opcode object.
@ivar log: holds the execution log and consists of tuples
of the form C{(log_serial, timestamp, level, message)}
@ivar stop_timestamp: timestamp for the end of the execution
"""
+ __slots__ = ["input", "status", "result", "log",
+ "start_timestamp", "end_timestamp",
+ "__weakref__"]
+
def __init__(self, op):
"""Constructor for the _QuededOpCode.
@ivar id: the job ID
@type ops: list
@ivar ops: the list of _QueuedOpCode that constitute the job
- @type run_op_index: int
- @ivar run_op_index: the currently executing opcode, or -1 if
- we didn't yet start executing
@type log_serial: int
@ivar log_serial: holds the index for the next log entry
@ivar received_timestamp: the timestamp for when the job was received
@ivar start_timestmap: the timestamp for start of execution
@ivar end_timestamp: the timestamp for end of execution
+ @ivar lock_status: In-memory locking information for debugging
@ivar change: a Condition variable we use for waiting for job changes
"""
+ __slots__ = ["queue", "id", "ops", "log_serial",
+ "received_timestamp", "start_timestamp", "end_timestamp",
+ "lock_status", "change",
+ "__weakref__"]
+
def __init__(self, queue, job_id, ops):
"""Constructor for the _QueuedJob.
self.queue = queue
self.id = job_id
self.ops = [_QueuedOpCode(op) for op in ops]
- self.run_op_index = -1
self.log_serial = 0
self.received_timestamp = TimeStampNow()
self.start_timestamp = None
self.end_timestamp = None
+ # In-memory attributes
+ self.lock_status = None
+
# Condition to wait for changes
self.change = threading.Condition(self.queue._lock)
obj = _QueuedJob.__new__(cls)
obj.queue = queue
obj.id = state["id"]
- obj.run_op_index = state["run_op_index"]
obj.received_timestamp = state.get("received_timestamp", None)
obj.start_timestamp = state.get("start_timestamp", None)
obj.end_timestamp = state.get("end_timestamp", None)
+ # In-memory attributes
+ obj.lock_status = None
+
obj.ops = []
obj.log_serial = 0
for op_state in state["ops"]:
return {
"id": self.id,
"ops": [op.Serialize() for op in self.ops],
- "run_op_index": self.run_op_index,
"start_timestamp": self.start_timestamp,
"end_timestamp": self.end_timestamp,
"received_timestamp": self.received_timestamp,
"""Selectively returns the log entries.
@type newer_than: None or int
- @param newer_than: if this is None, return all log enties,
+ @param newer_than: if this is None, return all log entries,
otherwise return only the log entries with serial higher
than this value
@rtype: list
return entries
+ def MarkUnfinishedOps(self, status, result):
+ """Mark unfinished opcodes with a given status and result.
-class _JobQueueWorker(workerpool.BaseWorker):
- """The actual job workers.
+ This is an utility function for marking all running or waiting to
+ be run opcodes with a given status. Opcodes which are already
+ finalised are not changed.
+
+ @param status: a given opcode status
+ @param result: the opcode result
+
+ """
+ not_marked = True
+ for op in self.ops:
+ if op.status in constants.OPS_FINALIZED:
+ assert not_marked, "Finalized opcodes found after non-finalized ones"
+ continue
+ op.status = status
+ op.result = result
+ not_marked = False
- """
- def _NotifyStart(self):
- """Mark the opcode as running, not lock-waiting.
- This is called from the mcpu code as a notifier function, when the
- LU is finally about to start the Exec() method. Of course, to have
- end-user visible results, the opcode must be initially (before
- calling into Processor.ExecOpCode) set to OP_STATUS_WAITLOCK.
+class _OpExecCallbacks(mcpu.OpExecCbBase):
+ def __init__(self, queue, job, op):
+ """Initializes this class.
+
+ @type queue: L{JobQueue}
+ @param queue: Job queue
+ @type job: L{_QueuedJob}
+ @param job: Job object
+ @type op: L{_QueuedOpCode}
+ @param op: OpCode
"""
- assert self.queue, "Queue attribute is missing"
- assert self.opcode, "Opcode attribute is missing"
+ assert queue, "Queue is missing"
+ assert job, "Job is missing"
+ assert op, "Opcode is missing"
+
+ self._queue = queue
+ self._job = job
+ self._op = op
+
+ def NotifyStart(self):
+ """Mark the opcode as running, not lock-waiting.
- self.queue.acquire()
+ This is called from the mcpu code as a notifier function, when the LU is
+ finally about to start the Exec() method. Of course, to have end-user
+ visible results, the opcode must be initially (before calling into
+ Processor.ExecOpCode) set to OP_STATUS_WAITLOCK.
+
+ """
+ self._queue.acquire()
try:
- assert self.opcode.status in (constants.OP_STATUS_WAITLOCK,
- constants.OP_STATUS_CANCELING)
+ assert self._op.status in (constants.OP_STATUS_WAITLOCK,
+ constants.OP_STATUS_CANCELING)
+
+ # All locks are acquired by now
+ self._job.lock_status = None
# Cancel here if we were asked to
- if self.opcode.status == constants.OP_STATUS_CANCELING:
+ if self._op.status == constants.OP_STATUS_CANCELING:
raise CancelJob()
- self.opcode.status = constants.OP_STATUS_RUNNING
+ self._op.status = constants.OP_STATUS_RUNNING
+ finally:
+ self._queue.release()
+
+ def Feedback(self, *args):
+ """Append a log entry.
+
+ """
+ assert len(args) < 3
+
+ if len(args) == 1:
+ log_type = constants.ELOG_MESSAGE
+ log_msg = args[0]
+ else:
+ (log_type, log_msg) = args
+
+ # The time is split to make serialization easier and not lose
+ # precision.
+ timestamp = utils.SplitTime(time.time())
+
+ self._queue.acquire()
+ try:
+ self._job.log_serial += 1
+ self._op.log.append((self._job.log_serial, timestamp, log_type, log_msg))
+
+ self._job.change.notifyAll()
finally:
- self.queue.release()
+ self._queue.release()
+
+ def ReportLocks(self, msg):
+ """Write locking information to the job.
+
+ Called whenever the LU processor is waiting for a lock or has acquired one.
+
+ """
+ # Not getting the queue lock because this is a single assignment
+ self._job.lock_status = msg
+
+
+class _JobQueueWorker(workerpool.BaseWorker):
+ """The actual job workers.
+ """
def RunTask(self, job):
"""Job executor.
@param job: the job to be processed
"""
- logging.debug("Worker %s processing job %s",
+ logging.info("Worker %s processing job %s",
self.worker_id, job.id)
proc = mcpu.Processor(self.pool.queue.context)
- self.queue = queue = job.queue
+ queue = job.queue
try:
try:
count = len(job.ops)
for idx, op in enumerate(job.ops):
+ op_summary = op.input.Summary()
+ if op.status == constants.OP_STATUS_SUCCESS:
+ # this is a job that was partially completed before master
+ # daemon shutdown, so it can be expected that some opcodes
+ # are already completed successfully (if any did error
+ # out, then the whole job should have been aborted and not
+ # resubmitted for processing)
+ logging.info("Op %s/%s: opcode %s already processed, skipping",
+ idx + 1, count, op_summary)
+ continue
try:
- logging.debug("Op %s/%s: Starting %s", idx + 1, count, op)
+ logging.info("Op %s/%s: Starting opcode %s", idx + 1, count,
+ op_summary)
queue.acquire()
try:
+ if op.status == constants.OP_STATUS_CANCELED:
+ raise CancelJob()
assert op.status == constants.OP_STATUS_QUEUED
- job.run_op_index = idx
op.status = constants.OP_STATUS_WAITLOCK
op.result = None
op.start_timestamp = TimeStampNow()
finally:
queue.release()
- def _Log(*args):
- """Append a log entry.
-
- """
- assert len(args) < 3
-
- if len(args) == 1:
- log_type = constants.ELOG_MESSAGE
- log_msg = args[0]
- else:
- log_type, log_msg = args
-
- # The time is split to make serialization easier and not lose
- # precision.
- timestamp = utils.SplitTime(time.time())
-
- queue.acquire()
- try:
- job.log_serial += 1
- op.log.append((job.log_serial, timestamp, log_type, log_msg))
-
- job.change.notifyAll()
- finally:
- queue.release()
-
- # Make sure not to hold lock while _Log is called
- self.opcode = op
- result = proc.ExecOpCode(input_opcode, _Log, self._NotifyStart)
+ # Make sure not to hold queue lock while calling ExecOpCode
+ result = proc.ExecOpCode(input_opcode,
+ _OpExecCallbacks(queue, job, op))
queue.acquire()
try:
finally:
queue.release()
- logging.debug("Op %s/%s: Successfully finished %s",
- idx + 1, count, op)
+ logging.info("Op %s/%s: Successfully finished opcode %s",
+ idx + 1, count, op_summary)
except CancelJob:
# Will be handled further up
raise
try:
try:
op.status = constants.OP_STATUS_ERROR
- op.result = str(err)
+ if isinstance(err, errors.GenericError):
+ op.result = errors.EncodeException(err)
+ else:
+ op.result = str(err)
op.end_timestamp = TimeStampNow()
- logging.debug("Op %s/%s: Error in %s", idx + 1, count, op)
+ logging.info("Op %s/%s: Error in opcode %s: %s",
+ idx + 1, count, op_summary, err)
finally:
queue.UpdateJobUnlocked(job)
finally:
queue.acquire()
try:
try:
- job.run_op_idx = -1
+ job.lock_status = None
job.end_timestamp = TimeStampNow()
queue.UpdateJobUnlocked(job)
finally:
status = job.CalcStatus()
finally:
queue.release()
- logging.debug("Worker %s finished job %s, status = %s",
- self.worker_id, job_id, status)
+
+ logging.info("Worker %s finished job %s, status = %s",
+ self.worker_id, job_id, status)
class _JobQueueWorkerPool(workerpool.WorkerPool):
class JobQueue(object):
- """Quue used to manaage the jobs.
+ """Queue used to manage the jobs.
@cvar _RE_JOB_FILE: regex matching the valid job file names
constants.JOB_STATUS_CANCELING):
logging.warning("Unfinished job %s found: %s", job.id, job)
try:
- for op in job.ops:
- op.status = constants.OP_STATUS_ERROR
- op.result = "Unclean master daemon shutdown"
+ job.MarkUnfinishedOps(constants.OP_STATUS_ERROR,
+ "Unclean master daemon shutdown")
finally:
self.UpdateJobUnlocked(job)
assert node_name != self._my_hostname
# Clean queue directory on added node
- rpc.RpcRunner.call_jobqueue_purge(node_name)
+ result = rpc.RpcRunner.call_jobqueue_purge(node_name)
+ msg = result.fail_msg
+ if msg:
+ logging.warning("Cannot cleanup queue directory on node %s: %s",
+ node_name, msg)
if not node.master_candidate:
# remove if existing, ignoring errors
for file_name in files:
# Read file content
- fd = open(file_name, "r")
- try:
- content = fd.read()
- finally:
- fd.close()
+ content = utils.ReadFile(file_name)
result = rpc.RpcRunner.call_jobqueue_update([node_name],
[node.primary_ip],
file_name, content)
- if not result[node_name]:
- logging.error("Failed to upload %s to %s", file_name, node_name)
+ msg = result[node_name].fail_msg
+ if msg:
+ logging.error("Failed to upload file %s to node %s: %s",
+ file_name, node_name, msg)
self._nodes[node_name] = node.primary_ip
Since we aim to keep consistency should this node (the current
master) fail, we will log errors if our rpc fail, and especially
- log the case when more than half of the nodes failes.
+ log the case when more than half of the nodes fails.
@param result: the data as returned from the rpc call
@type nodes: list
success = []
for node in nodes:
- if result[node]:
- success.append(node)
- else:
+ msg = result[node].fail_msg
+ if msg:
failed.append(node)
-
- if failed:
- logging.error("%s failed on %s", failmsg, ", ".join(failed))
+ logging.error("RPC call %s failed on node %s: %s",
+ result[node].call, node, msg)
+ else:
+ success.append(node)
# +1 for the master node
if (len(success) + 1) < len(failed):
self._CheckRpcResult(result, self._nodes,
"Updating %s" % file_name)
- def _RenameFileUnlocked(self, old, new):
+ def _RenameFilesUnlocked(self, rename):
"""Renames a file locally and then replicate the change.
This function will rename a file in the local queue directory
and then replicate this rename to all the other nodes we have.
- @type old: str
- @param old: the current name of the file
- @type new: str
- @param new: the new name of the file
+ @type rename: list of (old, new)
+ @param rename: List containing tuples mapping old to new names
"""
- utils.RenameFile(old, new, mkdir=True)
+ # Rename them locally
+ for old, new in rename:
+ utils.RenameFile(old, new, mkdir=True)
+ # ... and on all nodes
names, addrs = self._GetNodeIp()
- result = rpc.RpcRunner.call_jobqueue_rename(names, addrs, old, new)
- self._CheckRpcResult(result, self._nodes,
- "Moving %s to %s" % (old, new))
+ result = rpc.RpcRunner.call_jobqueue_rename(names, addrs, rename)
+ self._CheckRpcResult(result, self._nodes, "Renaming files (%r)" % rename)
def _FormatJobID(self, job_id):
"""Convert a job ID to string format.
"""
return str(int(job_id) / JOBS_PER_ARCHIVE_DIRECTORY)
- def _NewSerialUnlocked(self):
+ def _NewSerialsUnlocked(self, count):
"""Generates a new job identifier.
Job identifiers are unique during the lifetime of a cluster.
+ @type count: integer
+ @param count: how many serials to return
@rtype: str
@return: a string representing the job identifier.
"""
+ assert count > 0
# New number
- serial = self._last_serial + 1
+ serial = self._last_serial + count
# Write to file
self._WriteAndReplicateFileUnlocked(constants.JOB_QUEUE_SERIAL_FILE,
"%s\n" % serial)
+ result = [self._FormatJobID(v)
+ for v in range(self._last_serial, serial + 1)]
# Keep it only if we were able to write the file
self._last_serial = serial
- return self._FormatJobID(serial)
+ return result
@staticmethod
def _GetJobPath(job_id):
filepath = self._GetJobPath(job_id)
logging.debug("Loading job from %s", filepath)
try:
- fd = open(filepath, "r")
+ raw_data = utils.ReadFile(filepath)
except IOError, err:
if err.errno in (errno.ENOENT, ):
return None
raise
- try:
- data = serializer.LoadJson(fd.read())
- finally:
- fd.close()
+
+ data = serializer.LoadJson(raw_data)
try:
job = _QueuedJob.Restore(self, data)
else:
# non-archived case
logging.exception("Can't parse job %s, will archive.", job_id)
- self._RenameFileUnlocked(filepath, new_path)
+ self._RenameFilesUnlocked([(filepath, new_path)])
return None
self._memcache[job_id] = job
and in the future we might merge them.
@type drain_flag: boolean
- @param drain_flag: wheter to set or unset the drain flag
+ @param drain_flag: Whether to set or unset the drain flag
"""
if drain_flag:
utils.RemoveFile(constants.JOB_QUEUE_DRAIN_FILE)
return True
- @utils.LockedMethod
@_RequireOpenQueue
- def SubmitJob(self, ops):
+ def _SubmitJobUnlocked(self, job_id, ops):
"""Create and store a new job.
This enters the job into our job queue and also puts it on the new
queue, in order for it to be picked up by the queue processors.
+ @type job_id: job ID
+ @param job_id: the job ID for the new job
@type ops: list
@param ops: The list of OpCodes that will become the new job.
@rtype: job ID
"""
if self._IsQueueMarkedDrain():
- raise errors.JobQueueDrainError()
+ raise errors.JobQueueDrainError("Job queue is drained, refusing job")
# Check job queue size
size = len(self._ListJobFiles())
if size >= constants.JOB_QUEUE_SIZE_HARD_LIMIT:
raise errors.JobQueueFull()
- # Get job identifier
- job_id = self._NewSerialUnlocked()
job = _QueuedJob(self, job_id, ops)
# Write to disk
return job.id
+ @utils.LockedMethod
+ @_RequireOpenQueue
+ def SubmitJob(self, ops):
+ """Create and store a new job.
+
+ @see: L{_SubmitJobUnlocked}
+
+ """
+ job_id = self._NewSerialsUnlocked(1)[0]
+ return self._SubmitJobUnlocked(job_id, ops)
+
+ @utils.LockedMethod
+ @_RequireOpenQueue
+ def SubmitManyJobs(self, jobs):
+ """Create and store multiple jobs.
+
+ @see: L{_SubmitJobUnlocked}
+
+ """
+ results = []
+ all_job_ids = self._NewSerialsUnlocked(len(jobs))
+ for job_id, ops in zip(all_job_ids, jobs):
+ try:
+ data = self._SubmitJobUnlocked(job_id, ops)
+ status = True
+ except errors.GenericError, err:
+ data = str(err)
+ status = False
+ results.append((status, data))
+
+ return results
+
@_RequireOpenQueue
def UpdateJobUnlocked(self, job):
"""Update a job's on disk storage.
as such by the clients
"""
- logging.debug("Waiting for changes in job %s", job_id)
- end_time = time.time() + timeout
- while True:
- delta_time = end_time - time.time()
- if delta_time < 0:
- return constants.JOB_NOTCHANGED
+ job = self._LoadJobUnlocked(job_id)
+ if not job:
+ logging.debug("Job %s not found", job_id)
+ return None
- job = self._LoadJobUnlocked(job_id)
- if not job:
- logging.debug("Job %s not found", job_id)
- break
+ def _CheckForChanges():
+ logging.debug("Waiting for changes in job %s", job_id)
status = job.CalcStatus()
job_info = self._GetJobInfoUnlocked(job, fields)
job_info = serializer.LoadJson(serializer.DumpJson(job_info))
log_entries = serializer.LoadJson(serializer.DumpJson(log_entries))
- if status not in (constants.JOB_STATUS_QUEUED,
- constants.JOB_STATUS_RUNNING,
- constants.JOB_STATUS_WAITLOCK):
- # Don't even try to wait if the job is no longer running, there will be
- # no changes.
- break
-
- if (prev_job_info != job_info or
+ # Don't even try to wait if the job is no longer running, there will be
+ # no changes.
+ if (status not in (constants.JOB_STATUS_QUEUED,
+ constants.JOB_STATUS_RUNNING,
+ constants.JOB_STATUS_WAITLOCK) or
+ prev_job_info != job_info or
(log_entries and prev_log_serial != log_entries[0][0])):
- break
-
- logging.debug("Waiting again")
+ logging.debug("Job %s changed", job_id)
+ return (job_info, log_entries)
- # Release the queue lock while waiting
- job.change.wait(delta_time)
+ raise utils.RetryAgain()
- logging.debug("Job %s changed", job_id)
-
- return (job_info, log_entries)
+ try:
+ # Setting wait function to release the queue lock while waiting
+ return utils.Retry(_CheckForChanges, utils.RETRY_REMAINING_TIME, timeout,
+ wait_fn=job.change.wait)
+ except utils.RetryTimeout:
+ return constants.JOB_NOTCHANGED
@utils.LockedMethod
@_RequireOpenQueue
if job_status not in (constants.JOB_STATUS_QUEUED,
constants.JOB_STATUS_WAITLOCK):
- logging.debug("Job %s is no longer in the queue", job.id)
- return (False, "Job %s is no longer in the queue" % job.id)
+ logging.debug("Job %s is no longer waiting in the queue", job.id)
+ return (False, "Job %s is no longer waiting in the queue" % job.id)
if job_status == constants.JOB_STATUS_QUEUED:
self.CancelJobUnlocked(job)
elif job_status == constants.JOB_STATUS_WAITLOCK:
# The worker will notice the new status and cancel the job
try:
- for op in job.ops:
- op.status = constants.OP_STATUS_CANCELING
+ job.MarkUnfinishedOps(constants.OP_STATUS_CANCELING, None)
finally:
self.UpdateJobUnlocked(job)
return (True, "Job %s will be canceled" % job.id)
"""
try:
- for op in job.ops:
- op.status = constants.OP_STATUS_ERROR
- op.result = "Job canceled by request"
+ job.MarkUnfinishedOps(constants.OP_STATUS_CANCELED,
+ "Job canceled by request")
finally:
self.UpdateJobUnlocked(job)
@_RequireOpenQueue
- def _ArchiveJobUnlocked(self, job):
- """Archives a job.
+ def _ArchiveJobsUnlocked(self, jobs):
+ """Archives jobs.
- @type job: L{_QueuedJob}
- @param job: Job object
- @rtype bool
- @return Whether job was archived
+ @type jobs: list of L{_QueuedJob}
+ @param jobs: Job objects
+ @rtype: int
+ @return: Number of archived jobs
"""
- if job.CalcStatus() not in (constants.JOB_STATUS_CANCELED,
- constants.JOB_STATUS_SUCCESS,
- constants.JOB_STATUS_ERROR):
- logging.debug("Job %s is not yet done", job.id)
- return False
+ archive_jobs = []
+ rename_files = []
+ for job in jobs:
+ if job.CalcStatus() not in (constants.JOB_STATUS_CANCELED,
+ constants.JOB_STATUS_SUCCESS,
+ constants.JOB_STATUS_ERROR):
+ logging.debug("Job %s is not yet done", job.id)
+ continue
- old = self._GetJobPath(job.id)
- new = self._GetArchivedJobPath(job.id)
+ archive_jobs.append(job)
- self._RenameFileUnlocked(old, new)
+ old = self._GetJobPath(job.id)
+ new = self._GetArchivedJobPath(job.id)
+ rename_files.append((old, new))
- logging.debug("Successfully archived job %s", job.id)
+ # TODO: What if 1..n files fail to rename?
+ self._RenameFilesUnlocked(rename_files)
- return True
+ logging.debug("Successfully archived job(s) %s",
+ ", ".join(job.id for job in archive_jobs))
+
+ return len(archive_jobs)
@utils.LockedMethod
@_RequireOpenQueue
def ArchiveJob(self, job_id):
"""Archives a job.
- This is just a wrapper over L{_ArchiveJobUnlocked}.
+ This is just a wrapper over L{_ArchiveJobsUnlocked}.
@type job_id: string
@param job_id: Job ID of job to be archived.
logging.debug("Job %s not found", job_id)
return False
- return self._ArchiveJobUnlocked(job)
+ return self._ArchiveJobsUnlocked([job]) == 1
@utils.LockedMethod
@_RequireOpenQueue
last_touched = 0
all_job_ids = self._GetJobIDsUnlocked(archived=False)
+ pending = []
for idx, job_id in enumerate(all_job_ids):
last_touched = idx
+ # Not optimal because jobs could be pending
+ # TODO: Measure average duration for job archival and take number of
+ # pending jobs into account.
if time.time() > end_time:
break
job_age = job.end_timestamp
if age == -1 or now - job_age[0] > age:
- archived = self._ArchiveJobUnlocked(job)
- if archived:
- archived_count += 1
- continue
+ pending.append(job)
+
+ # Archive 10 jobs at a time
+ if len(pending) >= 10:
+ archived_count += self._ArchiveJobsUnlocked(pending)
+ pending = []
+ if pending:
+ archived_count += self._ArchiveJobsUnlocked(pending)
return (archived_count, len(all_job_ids) - last_touched - 1)
row.append(job.start_timestamp)
elif fname == "end_ts":
row.append(job.end_timestamp)
+ elif fname == "lock_status":
+ row.append(job.lock_status)
elif fname == "summary":
row.append([op.input.Summary() for op in job.ops])
else: