Differential D86 Diff 301 swh/storage/archiver/director.py

Changeset View

Standalone View

swh/storage/archiver/director.py

	# Copyright (C) 2015-2016 The Software Heritage developers			# Copyright (C) 2015-2016 The Software Heritage developers
	# See the AUTHORS file at the top-level directory of this distribution			# See the AUTHORS file at the top-level directory of this distribution
	# License: GNU General Public License version 3, or any later version			# License: GNU General Public License version 3, or any later version
	# See top-level LICENSE file for more information			# See top-level LICENSE file for more information

	import logging
	import click			import click
	import time

	from swh.core import hashutil, config			from swh.core import config
	from swh.objstorage import PathSlicingObjStorage
	from swh.objstorage.api.client import RemoteObjStorage
	from swh.scheduler.celery_backend.config import app			from swh.scheduler.celery_backend.config import app

	from . import tasks # NOQA			from . import tasks # NOQA
	from .storage import ArchiverStorage			from .storage import ArchiverStorage


	DEFAULT_CONFIG = {
	'objstorage_type': ('str', 'local_storage'),
	'objstorage_path': ('str', '/tmp/swh-storage/objects'),
	'objstorage_slicing': ('str', '0:2/2:4/4:6'),
	'objstorage_url': ('str', 'http://localhost:5003/'),

	'batch_max_size': ('int', 50),
	'archival_max_age': ('int', 3600),
	'retention_policy': ('int', 2),
	'asynchronous': ('bool', True),

	'dbconn': ('str', 'dbname=softwareheritage-archiver-dev user=guest')
	}

	task_name = 'swh.storage.archiver.tasks.SWHArchiverTask'			task_name = 'swh.storage.archiver.tasks.SWHArchiverTask'
				olasdUnsubmitted Done Inline Actions The type name should probably be `dict` rather than json, as json will never be used for that purpose anyway olasd: The type name should probably be `dict` rather than json, as json will never be used for that…

	logger = logging.getLogger()

				class ArchiverDirector(config.SWHConfig):
	class ArchiverDirector():
	"""Process the files in order to know which one is needed as backup.			"""Process the files in order to know which one is needed as backup.

				olasdUnsubmitted Done Inline Actions The indentation of that dict looks broken :) olasd: The indentation of that dict looks broken :)
	The archiver director processes the files in the local storage in order			The archiver director processes the files in the local storage in order
	to know which one needs archival and it delegates this task to			to know which one needs archival and it delegates this task to
	archiver workers.			archiver workers.
	Attributes:			"""
	master_objstorage: the local storage of the master server.
	master_objstorage_args (dict): arguments of the master objstorage
	initialization.

	archiver_storage: a wrapper for archiver db operations.			DEFAULT_CONFIG = {
	db_conn_archiver: Either a libpq connection string,			'batch_max_size': ('int', 1500),
	or a psycopg2 connection for the archiver db.			'retention_policy': ('int', 2),
				'asynchronous': ('bool', True),

	slave_objstorages: Iterable of remote obj storages to the slaves			'dbconn': ('str', 'dbname=softwareheritage-archiver-dev user=guest')
	servers used for backup.			}
				olasdUnsubmitted Done Inline Actions Please make that `archiver/director` (therefore storing the configuration in `/etc/softwareheritage/archiver/director.{ini,yml}`) olasd: Please make that `archiver/director` (therefore storing the configuration in…
	config: Archiver_configuration. A dictionary that must contain			CONFIG_BASE_FILENAME = 'archiver/director'
	the following keys:

	objstorage_type (str): type of objstorage used (local_storage
	or remote_storage).
	If the storage is local, the arguments keys must be present
	objstorage_path (str): master's objstorage path
	objstorage_slicing (str): masters's objstorage slicing
	Otherwise, if it's a remote objstorage, the keys must be:
	objstorage_url (str): url of the remote objstorage

	batch_max_size (int): The number of content items that can be
	given to the same archiver worker.
	archival_max_age (int): Delay given to the worker to copy all
	the files in a given batch.
	retention_policy (int): Required number of copies for the
	content to be considered safe.
	asynchronous (boolean): Indicate whenever the archival should
	run in asynchronous mode or not.
	"""

	def __init__(self, db_conn_archiver, config):			def __init__(self, add_config):
	""" Constructor of the archiver director.			""" Constructor of the archiver director.

	Args:			Args:
	db_conn_archiver: Either a libpq connection string,			db_conn_archiver: Either a libpq connection string,
	or a psycopg2 connection for the archiver db.			or a psycopg2 connection for the archiver db.
	config: Archiver_configuration. A dictionary that must contain			config: optionnal additional configuration. Keys in the dict will
	the following keys:			override the one parsed from the configuration file.

	objstorage_type (str): type of objstorage used
	(local_objstorage or remote_objstorage).
	If the storage is local, the arguments keys must be present
	objstorage_path (str): master's objstorage path
	objstorage_slicing (str): masters's objstorage slicing
	Otherwise, if it's a remote objstorage, the keys must be:
	objstorage_url (str): url of the remote objstorage

	batch_max_size (int): The number of content items that can be
	given to the same archiver worker.
	archival_max_age (int): Delay given to the worker to copy all
	the files in a given batch.
	retention_policy (int): Required number of copies for the
	content to be considered safe.
	asynchronous (boolean): Indicate whenever the archival should
	run in asynchronous mode or not.
	"""			"""
	# Get the slave storages			self.config = self.parse_config_file(additional_configs=[add_config])
	self.db_conn_archiver = db_conn_archiver			self.archiver_storage = ArchiverStorage(self.config['dbconn'])
	self.archiver_storage = ArchiverStorage(db_conn_archiver)
	self.slave_objstorages = {
	id: url
	for id, url
	in self.archiver_storage.archive_ls()
	}
	# Check that there is enough backup servers for the retention policy
	if config['retention_policy'] > len(self.slave_objstorages) + 1:
	raise ValueError(
	"Can't have a retention policy of %d with %d backup servers"
	% (config['retention_policy'], len(self.slave_objstorages))
	)

	# Get the master storage that contains content to be archived
	if config['objstorage_type'] == 'local_objstorage':
	master_objstorage_args = {
	'root': config['objstorage_path'],
	'slicing': config['objstorage_slicing']
	}
	master_objstorage = PathSlicingObjStorage(
	**master_objstorage_args
	)
	elif config['objstorage_type'] == 'remote_objstorage':
	master_objstorage_args = {'base_url': config['objstorage_url']}
	master_objstorage = RemoteObjStorage(**master_objstorage_args)
	else:
	raise ValueError(
	'Unknow objstorage class `%s`' % config['objstorage_type']
	)
	self.master_objstorage = master_objstorage
	self.master_objstorage_args = master_objstorage_args

	# Keep the full configuration
	self.config = config

	def run(self):			def run(self):
	""" Run the archiver director.			""" Run the archiver director.

	The archiver director will check all the contents of the archiver			The archiver director will check all the contents of the archiver
	database and do the required backup jobs.			database and do the required backup jobs.
	"""			"""
	if self.config['asynchronous']:			if self.config['asynchronous']:
	run_fn = self.run_async_worker			run_fn = self.run_async_worker
	else:			else:
	run_fn = self.run_sync_worker			run_fn = self.run_sync_worker

	for batch in self.get_unarchived_content():			for batch in self.get_unarchived_content_batch():
	run_fn(batch)			run_fn(batch)

				olasdUnsubmitted Done Inline Actions This function can be removed now. olasd: This function can be removed now.
				def _worker_args(self, batch):
				""" Generates a dict that contains the arguments for a worker.
				"""
				return {
				'batch': batch
				olasdUnsubmitted Done Inline Actions We do not want to pass configuration to workers via Celery, ever. Considering that, I think the `add_config` instance attribute should be removed. The only actual use of that is overriding the file configuration when running the director on the command line, and we can just override specific configuration items in the director class's `__init__` method. This makes the "worker arguments" just a plain list of objects that need archiving. No need for a wrapping dict, or anything else. olasd: We do not want to pass configuration to workers via Celery, ever. Considering that, I think the…
				}

	def run_async_worker(self, batch):			def run_async_worker(self, batch):
	""" Produce a worker that will be added to the task queue.			""" Produce a worker that will be added to the task queue.
	"""			"""
				olasdUnsubmitted Done Inline Actions The worker should really read all of this from its configuration rather than having it passed around as a message. The access to the same objstorage can be different from one worker to another The archival policy can have changed since we sent the messages in the queue db access can also be different from one worker to another the dbconn string can contain passwords that we don't really want to have flying around All in all, all the director should enqueue is the list of objects that might need copies. olasd: The worker should really read all of this from its configuration rather than having it passed…
	task = app.tasks[task_name]			task = app.tasks[task_name]
	task.delay(batch,			task.delay(**self._worker_args(batch))
				olasdUnsubmitted Done Inline Actions `task.delay(batch)` olasd: `task.delay(batch)`
	archiver_args=self.db_conn_archiver,
	master_objstorage_args=self.master_objstorage_args,
	slave_objstorages=self.slave_objstorages,
	config=self.config)

	def run_sync_worker(self, batch):			def run_sync_worker(self, batch):
	""" Run synchronously a worker on the given batch.			""" Run synchronously a worker on the given batch.
	"""			"""
	task = app.tasks[task_name]			task = app.tasks[task_name]
	task(batch,			task(**self._worker_args(batch))
				olasdUnsubmitted Done Inline Actions `task(batch)` olasd: `task(batch)`
	archiver_args=self.db_conn_archiver,
	master_objstorage_args=self.master_objstorage_args,
	slave_objstorages=self.slave_objstorages,
	config=self.config)

	def get_unarchived_content(self):			def get_unarchived_content_batch(self):
	""" Get contents that need to be archived.			""" Create batch of contents that needs to be archived

	Yields:			Yields:
	A batch of contents. Batches are dictionaries which associates			batch of sha1 that corresponds to contents that needs more archive
	a content id to the data about servers that contains it or not.			copies.

	{'id1':
	{'present': [('slave1', 'slave1_url')],
	'missing': [('slave2', 'slave2_url'),
	('slave3', 'slave3_url')]
	},
	'id2':
	{'present': [],
	'missing': [
	('slave1', 'slave1_url'),
	('slave2', 'slave2_url'),
	('slave3', 'slave3_url')
	]}
	}

	Where keys (idX) are sha1 of the content and (slaveX, slaveX_url)
	are ids and urls of the storage slaves.

	At least all the content that don't have enough copies on the
	backups servers are distributed into these batches.
	"""			"""
	contents = {}			contents = []
	# Get the archives			for content in self._get_unarchived_content():
	archives = dict(self.archiver_storage.archive_ls())			contents.append(content)
	# Get all the contents referenced into the archiver tables			if len(contents) > self.config['batch_max_size']:
	last_object = b''
	while True:
	archived_contents = list(
	self.archiver_storage.content_archive_get_copies(last_object))

	if not archived_contents:
	break

	for content_id, present, ongoing in archived_contents:
	last_object = content_id
	data = {
	'present': set(present),
	'missing': set(archives) - set(present) - set(ongoing),
	}

	for archive_id, mtime in ongoing.items():
	status = self.get_virtual_status('ongoing', mtime)
	data[status].add(archive_id)

	if not data['missing']:
	continue

	contents[r'\x%s' % hashutil.hash_to_hex(content_id)] = {
	k: [(archive_id, archives[archive_id]) for archive_id in v]
	for k, v in data.items()
	}

	if len(contents) >= self.config['batch_max_size']:
	yield contents			yield contents
	contents = {}			contents = []

	if len(contents) > 0:			if len(contents) > 0:
	yield contents			yield contents

	def get_virtual_status(self, status, mtime):			def _get_unarchived_content(self):
	""" Compute the virtual presence of a content.			""" Get all the content ids in the db that needs more copies

	If the status is ongoing but the time is not elasped, the archiver			Yields:
	consider it will be present in the futur, and so consider it as			sha1 of contents that needs to be archived.
	present.
	However, if the time is elasped, the copy may have failed, so consider
	the content as missing.

	Arguments:
	status (string): One of ('present', 'missing', 'ongoing'). The
	status of the content.
	mtime (datetime): Time at which the content have been updated for
	the last time.

	Returns:
	The virtual status of the studied content, which is 'present' or
	'missing'.

	Raises:
	ValueError: if the status is not one 'present', 'missing'
	or 'ongoing'
	"""			"""
	if status in ('present', 'missing'):			for content_id, present, _ongoing in self._get_all_contents():
	return status			if len(present) < self.config['retention_policy']:
				olasdUnsubmitted Not Done Inline Actions This is okay as a first step but a tiny refactoring of the database should allow us to do that server-side rather than client-side. olasd: This is okay as a first step but a tiny refactoring of the database should allow us to do that…
				qcamposAuthorUnsubmitted Not Done Inline Actions Thats probably a better idea. I can't get the correct request for that however. What should I use in the WHERE section to count the number of `present` into the jsonb field ? qcampos: Thats probably a better idea. I can't get the correct request for that however. What should I…
				olasdUnsubmitted Not Done Inline Actions Let's look at that after this huge diff lands. olasd: Let's look at that after this huge diff lands.
				yield content_id
	# If the status is 'ongoing' but there is still time, another worker
	# may still be on the task.
	if status == 'ongoing':
	elapsed = int(time.time()) - mtime
	if elapsed <= self.config['archival_max_age']:
	return 'present'
	else:
	return 'missing'
	else:			else:
	raise ValueError("status must be either 'present', 'missing' "			continue
	"or 'ongoing'")
				def _get_all_contents(self):
				""" Get batchs from the archiver db and yield it as continous stream

				Yields:
				Datas about a content as a tuple
				(content_id, present_copies, ongoing_copies) where ongoing_copies
				is a dict mapping copy to mtime.
				"""
				last_object = b''
				while True:
				archiver_contents = list(
				self.archiver_storage.content_archive_get_copies(last_object)
				)
				if not archiver_contents:
				return
				for content in archiver_contents:
				last_object = content[0]
				yield content


	@click.command()			@click.command()
	@click.argument('config-path', required=1)			@click.option('--batch-size', help="Maximal number of objects in a batch")
				olasdUnsubmitted Done Inline Actions This argument is not needed any longer: the configuration file can be loaded from the default Software Heritage paths, that is in /etc or the user home directory. olasd: This argument is not needed any longer: the configuration file can be loaded from the default…
	@click.option('--dbconn', default=DEFAULT_CONFIG['dbconn'][1],			@click.option('--retention-policy',
				help="Minimal number of copies the archiver will create")
				@click.option('--dbconn',
	help="Connection string for the archiver database")			help="Connection string for the archiver database")
	@click.option('--async/--sync', default=DEFAULT_CONFIG['asynchronous'][1],			@click.option('--async/--sync',
	help="Indicates if the archiver should run asynchronously")			help="Indicates if the archiver should run asynchronously")
				olasdUnsubmitted Done Inline Actions The default for the command-line director should be to read the config file. When you're setting defaults for the command-line args, you make the command line _always_ override the contents of the configuration file with the default configuration. What you should do is: drop the defaults make the arguments non-mandatory only override the contents of the configuration file if the command line argument is passed. olasd: The default for the command-line director should be to read the config file. When you're…
	def launch(config_path, dbconn, async):			def launch(batch_size, retention_policy, dbconn, async):
	# The configuration have following priority :			# The configuration have following priority :
	# command line > file config > default config			# command line > file config > default config
	cl_config = {			# Values are None if not provided
				olasdUnsubmitted Done Inline Actions You should not need to do that any longer olasd: You should not need to do that any longer
	'dbconn': dbconn,			cl_config = create_conf(batch_size, retention_policy, dbconn, async)
	'asynchronous': async			# Rrun the archiver with the overriding conf.
	}			archiver = ArchiverDirector(cl_config)
				olasdUnsubmitted Done Inline Actions There, you can just pass the command-line config. olasd: There, you can just pass the command-line config.
	conf = config.read(config_path, DEFAULT_CONFIG)
	conf.update(cl_config)
	# Create connection data and run the archiver.
	archiver = ArchiverDirector(conf['dbconn'], conf)
	logger.info("Starting an archival at", time.time())
	archiver.run()			archiver.run()


				def create_conf(batch_size, retention_policy, dbconn, async):
				""" Create a dict that contains only the given arguments
				"""
				return dict(filter(lambda k, v: v is not None,
				(
				('batch_max_size', batch_size),
				('retention_policy', retention_policy),
				('dbconn', dbconn),
				('async', async)
				)))
				olasdUnsubmitted Done Inline Actions I'm sorry, but I'd rather we removed all the command line arguments than adding this complexity. This function has you duplicate the configuration keys at three places (command line argument names, function argument names, string constants), and realistically it will never be used as we will deploy a configuration file with puppet anyway... Down the line, we should make sure that our "config module" can generate a command-line parser, rather than doing stuff ad-hoc everywhere. Can we please just scrap the command line parsing and finally land this diff? olasd: I'm sorry, but I'd rather we removed all the command line arguments than adding this complexity.

	if __name__ == '__main__':			if __name__ == '__main__':
	launch()			launch()