Skip to content

Commit

Permalink
[Batch] Patch TaskOperations.add_collection with convenience function…
Browse files Browse the repository at this point in the history
…ality (#3217)

* patch file for bulk add task

* Update bulk task add to use local sdk references and added tests

* Update error handling

* update doc text

* Dynamically inject custom error into models
  • Loading branch information
bgklein authored and lmazuel committed Sep 17, 2018
1 parent 5b87ef6 commit ceeb803
Show file tree
Hide file tree
Showing 8 changed files with 95,907 additions and 190 deletions.
9 changes: 9 additions & 0 deletions azure-batch/HISTORY.rst
Original file line number Diff line number Diff line change
Expand Up @@ -3,6 +3,15 @@
Release History
===============

5.1.0 (2018-08-28)
++++++++++++++++++

- Update operation TaskOperations.add_collection with the following added functionality:
- Retry server side errors.
- Automatically chunk lists of more than 100 tasks to multiple requests.
- If tasks are too large to be submitted in chunks of 100, reduces number of tasks per request.
- Add a parameter to specify number of threads to use when submitting tasks.

5.0.0 (2018-08-24)
++++++++++++++++++

Expand Down
3 changes: 3 additions & 0 deletions azure-batch/azure/batch/batch_service_client.py
Original file line number Diff line number Diff line change
Expand Up @@ -23,6 +23,7 @@
from .operations.task_operations import TaskOperations
from .operations.compute_node_operations import ComputeNodeOperations
from . import models
from .custom.patch import patch_client


class BatchServiceClientConfiguration(AzureConfiguration):
Expand Down Expand Up @@ -112,3 +113,5 @@ def __init__(
self._client, self.config, self._serialize, self._deserialize)
self.compute_node = ComputeNodeOperations(
self._client, self.config, self._serialize, self._deserialize)

patch_client(self)
Empty file.
19 changes: 19 additions & 0 deletions azure-batch/azure/batch/custom/custom_errors.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,19 @@
# --------------------------------------------------------------------------------------------
# Copyright (c) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License. See License.txt in the project root for license information.
# --------------------------------------------------------------------------------------------


class CreateTasksErrorException(Exception):
""" Aggregate Exception containing details for any failures from a task add operation.
:param str message: Error message describing exit reason
:param [~TaskAddParameter] pending_task_list: List of tasks remaining to be submitted.
:param [~TaskAddResult] failure_tasks: List of tasks which failed to add
:param [~Exception] errors: List of unknown errors forcing early termination
"""
def __init__(self, message, pending_task_list=None, failure_tasks=None, errors=None):
self.message = message
self.pending_tasks = list(pending_task_list)
self.failure_tasks = list(failure_tasks)
self.errors = list(errors)
296 changes: 296 additions & 0 deletions azure-batch/azure/batch/custom/patch.py
Original file line number Diff line number Diff line change
@@ -0,0 +1,296 @@
import collections
import importlib
import logging
import threading
import types
import sys

from ..models import BatchErrorException, TaskAddCollectionResult, TaskAddStatus
from ..custom.custom_errors import CreateTasksErrorException
from ..operations.task_operations import TaskOperations

MAX_TASKS_PER_REQUEST = 100
_LOGGER = logging.getLogger(__name__)

class _TaskWorkflowManager(object):
"""Worker class for one add_collection request
:param ~TaskOperations task_operations: Parent object which instantiated this
:param str job_id: The ID of the job to which the task collection is to be
added.
:param tasks_to_add: The collection of tasks to add.
:type tasks_to_add: list of :class:`TaskAddParameter
<azure.batch.models.TaskAddParameter>`
:param task_add_collection_options: Additional parameters for the
operation
:type task_add_collection_options: :class:`TaskAddCollectionOptions
<azure.batch.models.TaskAddCollectionOptions>`
:param dict custom_headers: headers that will be added to the request
:param bool raw: returns the direct response alongside the
deserialized response
"""

def __init__(
self,
client,
job_id,
tasks_to_add,
task_add_collection_options=None,
custom_headers=None,
raw=False,
**kwargs):
# Append operations thread safe - Only read once all threads have completed
# List of tasks which failed to add due to a returned client error
self._failure_tasks = collections.deque()
# List of unknown exceptions which occurred during requests.
self._errors = collections.deque()

# synchronized through lock variables
self.error = None # Only written once all threads have completed
self._max_tasks_per_request = MAX_TASKS_PER_REQUEST
self._tasks_to_add = collections.deque(tasks_to_add)

self._error_lock = threading.Lock()
self._max_tasks_lock = threading.Lock()
self._pending_queue_lock = threading.Lock()

# Variables to be used for task add_collection requests
self._client = TaskOperations(
client._client, client.config, client._serialize, client._deserialize)
self._job_id = job_id
self._task_add_collection_options = task_add_collection_options
self._custom_headers = custom_headers
self._raw = raw
self._kwargs = dict(**kwargs)

def _bulk_add_tasks(self, results_queue, chunk_tasks_to_add):
"""Adds a chunk of tasks to the job
Retry chunk if body exceeds the maximum request size and retry tasks
if failed due to server errors.
:param results_queue: Queue to place the return value of the request
:type results_queue: collections.deque
:param chunk_tasks_to_add: Chunk of at most 100 tasks with retry details
:type chunk_tasks_to_add: list[~TrackedCloudTask]
"""

try:
add_collection_response = self._client.add_collection(
self._job_id,
chunk_tasks_to_add,
self._task_add_collection_options,
self._custom_headers,
self._raw)
except BatchErrorException as e:
# In case of a chunk exceeding the MaxMessageSize split chunk in half
# and resubmit smaller chunk requests
# TODO: Replace string with constant variable once available in SDK
if e.error.code == "RequestBodyTooLarge": # pylint: disable=no-member
# In this case the task is misbehaved and will not be able to be added due to:
# 1) The task exceeding the max message size
# 2) A single cell of the task exceeds the per-cell limit, or
# 3) Sum of all cells exceeds max row limit
if len(chunk_tasks_to_add) == 1:
failed_task = chunk_tasks_to_add.pop()
self._errors.appendleft(e)
_LOGGER.error("Failed to add task with ID %s due to the body"
" exceeding the maximum request size", failed_task.id)
else:
# Assumption: Tasks are relatively close in size therefore if one batch exceeds size limit
# we should decrease the initial task collection size to avoid repeating the error
# Midpoint is lower bounded by 1 due to above base case
midpoint = int(len(chunk_tasks_to_add) / 2)
# Restrict one thread at a time to do this compare and set,
# therefore forcing max_tasks_per_request to be strictly decreasing
with self._max_tasks_lock:
if midpoint < self._max_tasks_per_request:
self._max_tasks_per_request = midpoint
_LOGGER.info("Amount of tasks per request reduced from %s to %s due to the"
" request body being too large", str(self._max_tasks_per_request),
str(midpoint))

# Not the most efficient solution for all cases, but the goal of this is to handle this
# exception and have it work in all cases where tasks are well behaved
# Behavior retries as a smaller chunk and
# appends extra tasks to queue to be picked up by another thread .
self._tasks_to_add.extendleft(chunk_tasks_to_add[midpoint:])
self._bulk_add_tasks(results_queue, chunk_tasks_to_add[:midpoint])
# Retry server side errors
elif 500 <= e.response.status_code <= 599:
self._tasks_to_add.extendleft(chunk_tasks_to_add)
else:
# Re-add to pending queue as unknown status / don't have result
self._tasks_to_add.extendleft(chunk_tasks_to_add)
# Unknown State - don't know if tasks failed to add or were successful
self._errors.appendleft(e)
except Exception as e: # pylint: disable=broad-except
# Re-add to pending queue as unknown status / don't have result
self._tasks_to_add.extendleft(chunk_tasks_to_add)
# Unknown State - don't know if tasks failed to add or were successful
self._errors.appendleft(e)
else:
try:
add_collection_response = add_collection_response.output
except AttributeError:
pass

for task_result in add_collection_response.value: # pylint: disable=no-member
if task_result.status == TaskAddStatus.server_error:
# Server error will be retried
with self._pending_queue_lock:
for task in chunk_tasks_to_add:
if task.id == task_result.task_id:
self._tasks_to_add.appendleft(task)
elif (task_result.status == TaskAddStatus.client_error
and not task_result.error.code == "TaskExists"):
# Client error will be recorded unless Task already exists
self._failure_tasks.appendleft(task_result)
else:
results_queue.appendleft(task_result)

def task_collection_thread_handler(self, results_queue):
"""Main method for worker to run
Pops a chunk of tasks off the collection of pending tasks to be added and submits them to be added.
:param collections.deque results_queue: Queue for worker to output results to
"""
# Add tasks until either we run out or we run into an unexpected error
while self._tasks_to_add and not self._errors:
max_tasks = self._max_tasks_per_request # local copy
chunk_tasks_to_add = []
with self._pending_queue_lock:
while len(chunk_tasks_to_add) < max_tasks and self._tasks_to_add:
chunk_tasks_to_add.append(self._tasks_to_add.pop())

if chunk_tasks_to_add:
self._bulk_add_tasks(results_queue, chunk_tasks_to_add)

# Only define error if all threads have finished and there were failures
with self._error_lock:
if threading.active_count() == 1 and (self._failure_tasks or self._errors):
self.error = CreateTasksErrorException(
"One or more tasks failed to be added",
self._failure_tasks,
self._tasks_to_add,
self._errors)


def _handle_output(results_queue):
"""Scan output for exceptions
If there is an output from an add task collection call add it to the results.
:param results_queue: Queue containing results of attempted add_collection's
:type results_queue: collections.deque
:return: list of TaskAddResults
:rtype: list[~TaskAddResult]
"""
results = []
while results_queue:
queue_item = results_queue.pop()
results.append(queue_item)
return results

def patch_client(client):
try:
models = sys.modules['azure.batch.models']
except KeyError:
models = importlib.import_module('azure.batch.models')
setattr(models, 'CreateTasksErrorException', CreateTasksErrorException)
sys.modules['azure.batch.models'] = models
client.task.add_collection = types.MethodType(bulk_add_collection, client.task)

def bulk_add_collection(
client,
job_id,
value,
task_add_collection_options=None,
custom_headers=None,
raw=False,
threads=0,
**operation_config):
"""Adds a collection of tasks to the specified job.
Note that each task must have a unique ID. The Batch service may not
return the results for each task in the same order the tasks were
submitted in this request. If the server times out or the connection is
closed during the request, the request may have been partially or fully
processed, or not at all. In such cases, the user should re-issue the
request. Note that it is up to the user to correctly handle failures
when re-issuing a request. For example, you should use the same task
IDs during a retry so that if the prior operation succeeded, the retry
will not create extra tasks unexpectedly. If the response contains any
tasks which failed to add, a client can retry the request. In a retry,
it is most efficient to resubmit only tasks that failed to add, and to
omit tasks that were successfully added on the first attempt. The
maximum lifetime of a task from addition to completion is 7 days. If a
task has not completed within 7 days of being added it will be
terminated by the Batch service and left in whatever state it was in at
that time.
:param job_id: The ID of the job to which the task collection is to be
added.
:type job_id: str
:param value: The collection of tasks to add. The total serialized
size of this collection must be less than 4MB. If it is greater than
4MB (for example if each task has 100's of resource files or
environment variables), the request will fail with code
'RequestBodyTooLarge' and should be retried again with fewer tasks.
:type value: list of :class:`TaskAddParameter
<azure.batch.models.TaskAddParameter>`
:param task_add_collection_options: Additional parameters for the
operation
:type task_add_collection_options: :class:`TaskAddCollectionOptions
<azure.batch.models.TaskAddCollectionOptions>`
:param dict custom_headers: headers that will be added to the request
:param bool raw: returns the direct response alongside the
deserialized response
:param int threads: number of threads to use in parallel when adding tasks. If specified
and greater than 0, will start additional threads to submit requests and wait for them to finish.
Otherwise will submit add_collection requests sequentially on main thread
:return: :class:`TaskAddCollectionResult
<azure.batch.models.TaskAddCollectionResult>` or
:class:`ClientRawResponse<msrest.pipeline.ClientRawResponse>` if
raw=true
:rtype: :class:`TaskAddCollectionResult
<azure.batch.models.TaskAddCollectionResult>` or
:class:`ClientRawResponse<msrest.pipeline.ClientRawResponse>`
:raises:
:class:`BatchErrorException<azure.batch.models.BatchErrorException>`
"""

results_queue = collections.deque() # deque operations(append/pop) are thread-safe
task_workflow_manager = _TaskWorkflowManager(
client,
job_id,
value,
task_add_collection_options,
custom_headers,
raw,
**operation_config)

# multi-threaded behavior
if threads:
if threads < 0:
raise ValueError("Threads must be positive or 0")

active_threads = []
for i in range(threads):
active_threads.append(threading.Thread(
target=task_workflow_manager.task_collection_thread_handler,
args=(results_queue,)))
active_threads[-1].start()
for thread in active_threads:
thread.join()
# single-threaded behavior
else:
task_workflow_manager.task_collection_thread_handler(results_queue)

if task_workflow_manager.error:
raise task_workflow_manager.error # pylint: disable=raising-bad-type
else:
submitted_tasks = _handle_output(results_queue)
return TaskAddCollectionResult(value=submitted_tasks)
bulk_add_collection.metadata = {'url': '/jobs/{jobId}/addtaskcollection'}
2 changes: 1 addition & 1 deletion azure-batch/azure/batch/version.py
Original file line number Diff line number Diff line change
Expand Up @@ -9,5 +9,5 @@
# regenerated.
# --------------------------------------------------------------------------

VERSION = "5.0.0"
VERSION = "5.1.0"

Loading

0 comments on commit ceeb803

Please sign in to comment.