def _upload_pg_cluster_dir(self, start_backup_info, pg_cluster_dir,
version, pool_size, rate_limit=None):
"""
Upload to url_prefix from pg_cluster_dir
This function ignores the directory pg_xlog, which contains WAL
files and are not generally part of a base backup.
Note that this is also lzo compresses the files: thus, the number
of pooled processes involves doing a full sequential scan of the
uncompressed Postgres heap file that is pipelined into lzo. Once
lzo is completely finished (necessary to have access to the file
size) the file is sent to S3 or WABS.
TODO: Investigate an optimization to decouple the compression and
upload steps to make sure that the most efficient possible use of
pipelining of network and disk resources occurs. Right now it
possible to bounce back and forth between bottlenecking on reading
from the database block device and subsequently the S3/WABS sending
steps should the processes be at the same stage of the upload
pipeline: this can have a very negative impact on being able to
make full use of system resources.
Furthermore, it desirable to overflowing the page cache: having
separate tunables for number of simultanious compression jobs
(which occupy /tmp space and page cache) and number of uploads
(which affect upload throughput) would help.
"""
spec, parts = tar_partition.partition(pg_cluster_dir)
# TODO :: Move arbitray path construction to StorageLayout Object
backup_prefix = '{0}/basebackups_{1}/base_{file_name}_{file_offset}'\
.format(self.layout.prefix.rstrip('/'), FILE_STRUCTURE_VERSION,
**start_backup_info)
if rate_limit is None:
per_process_limit = None
else:
per_process_limit = int(rate_limit / pool_size)
# Reject tiny per-process rate limits. They should be
# rejected more nicely elsewhere.
assert per_process_limit > 0 or per_process_limit is None
total_size = 0
# Make an attempt to upload extended version metadata
extended_version_url = backup_prefix + '/extended_version.txt'
logger.info(
msg='start upload postgres version metadata',
detail=('Uploading to {extended_version_url}.'
.format(extended_version_url=extended_version_url)))
uri_put_file(self.creds,
extended_version_url, StringIO(version),
content_encoding='text/plain')
logger.info(msg='postgres version metadata upload complete')
uploader = PartitionUploader(self.creds, backup_prefix,
per_process_limit, self.gpg_key_id)
pool = TarUploadPool(uploader, pool_size)
# Enqueue uploads for parallel execution
for tpart in parts:
total_size += tpart.total_member_size
# 'put' can raise an exception for a just-failed upload,
# aborting the process.
pool.put(tpart)
# Wait for remaining parts to upload. An exception can be
# raised to signal failure of the upload.
pool.join()
return spec, backup_prefix, total_size
def _upload_pg_cluster_dir(self,
start_backup_info,
pg_cluster_dir,
version,
pool_size,
rate_limit=None):
"""
Upload to url_prefix from pg_cluster_dir
This function ignores the directory pg_xlog, which contains WAL
files and are not generally part of a base backup.
Note that this is also lzo compresses the files: thus, the number
of pooled processes involves doing a full sequential scan of the
uncompressed Postgres heap file that is pipelined into lzo. Once
lzo is completely finished (necessary to have access to the file
size) the file is sent to S3 or WABS.
TODO: Investigate an optimization to decouple the compression and
upload steps to make sure that the most efficient possible use of
pipelining of network and disk resources occurs. Right now it
possible to bounce back and forth between bottlenecking on reading
from the database block device and subsequently the S3/WABS sending
steps should the processes be at the same stage of the upload
pipeline: this can have a very negative impact on being able to
make full use of system resources.
Furthermore, it desirable to overflowing the page cache: having
separate tunables for number of simultanious compression jobs
(which occupy /tmp space and page cache) and number of uploads
(which affect upload throughput) would help.
"""
spec, parts = tar_partition.partition(pg_cluster_dir)
# TODO :: Move arbitray path construction to StorageLayout Object
backup_prefix = '{0}/basebackups_{1}/base_{file_name}_{file_offset}'\
.format(self.layout.prefix.rstrip('/'), FILE_STRUCTURE_VERSION,
**start_backup_info)
if rate_limit is None:
per_process_limit = None
else:
per_process_limit = int(rate_limit / pool_size)
# Reject tiny per-process rate limits. They should be
# rejected more nicely elsewhere.
assert per_process_limit is None or per_process_limit > 0
total_size = 0
# Make an attempt to upload extended version metadata
extended_version_url = backup_prefix + '/extended_version.txt'
logger.info(msg='start upload postgres version metadata',
detail=('Uploading to {extended_version_url}.'.format(
extended_version_url=extended_version_url)))
uri_put_file(self.creds,
extended_version_url,
BytesIO(version.encode("utf8")),
content_type='text/plain')
logger.info(msg='postgres version metadata upload complete')
uploader = PartitionUploader(self.creds, backup_prefix,
per_process_limit, self.gpg_key_id)
pool = TarUploadPool(uploader, pool_size)
# Enqueue uploads for parallel execution
for tpart in parts:
total_size += tpart.total_member_size
# 'put' can raise an exception for a just-failed upload,
# aborting the process.
pool.put(tpart)
# Wait for remaining parts to upload. An exception can be
# raised to signal failure of the upload.
pool.join()
return spec, backup_prefix, total_size
def database_backup(self, data_directory, *args, **kwargs):
"""Uploads a PostgreSQL file cluster to S3 or Windows Azure Blob
Service
Mechanism: just wraps _upload_pg_cluster_dir with
start/stop backup actions with exception handling.
In particular there is a 'finally' block to stop the backup in
most situations.
"""
upload_good = False
backup_stop_good = False
while_offline = False
start_backup_info = None
if 'while_offline' in kwargs:
while_offline = kwargs.pop('while_offline')
try:
if not while_offline:
start_backup_info = PgBackupStatements.run_start_backup()
version = PgBackupStatements.pg_version()['version']
else:
if os.path.exists(os.path.join(data_directory,
'postmaster.pid')):
hint = ('Shut down postgres. '
'If there is a stale lockfile, '
'then remove it after being very sure postgres '
'is not running.')
raise UserException(
msg='while_offline set, but pg looks to be running',
detail='Found a postmaster.pid lockfile, and aborting',
hint=hint)
ctrl_data = PgControlDataParser(data_directory)
start_backup_info = ctrl_data.last_xlog_file_name_and_offset()
version = ctrl_data.pg_version()
ret_tuple = self._upload_pg_cluster_dir(
start_backup_info, data_directory, version=version, *args,
**kwargs)
spec, uploaded_to, expanded_size_bytes = ret_tuple
upload_good = True
finally:
if not upload_good:
logger.warning(
'blocking on sending WAL segments',
detail=('The backup was not completed successfully, '
'but we have to wait anyway. '
'See README: TODO about pg_cancel_backup'))
if not while_offline:
stop_backup_info = PgBackupStatements.run_stop_backup()
else:
stop_backup_info = start_backup_info
backup_stop_good = True
# XXX: Ugly, this is more of a 'worker' task because it might
# involve retries and error messages, something that is not
# treated by the "operator" category of modules. So
# basically, if this small upload fails, the whole upload
# fails!
if upload_good and backup_stop_good:
# Try to write a sentinel file to the cluster backup
# directory that indicates that the base backup upload has
# definitely run its course and also communicates what WAL
# segments are needed to get to consistency.
sentinel_content = StringIO()
json.dump(
{'wal_segment_backup_stop':
stop_backup_info['file_name'],
'wal_segment_offset_backup_stop':
stop_backup_info['file_offset'],
'expanded_size_bytes': expanded_size_bytes,
'spec': spec},
sentinel_content)
# XXX: should use the storage operators.
#
# XXX: distinguish sentinels by *PREFIX* not suffix,
# which makes searching harder. (For the next version
# bump).
sentinel_content.seek(0)
uri_put_file(self.creds,
uploaded_to + '_backup_stop_sentinel.json',
sentinel_content, content_encoding='application/json')
else:
# NB: Other exceptions should be raised before this that
# have more informative results, it is intended that this
# exception never will get raised.
raise UserCritical('could not complete backup process')
def database_backup(self, data_directory, *args, **kwargs):
"""Uploads a PostgreSQL file cluster to S3 or Windows Azure Blob
Service
Mechanism: just wraps _upload_pg_cluster_dir with
start/stop backup actions with exception handling.
In particular there is a 'finally' block to stop the backup in
most situations.
"""
upload_good = False
backup_stop_good = False
while_offline = False
start_backup_info = None
if 'while_offline' in kwargs:
while_offline = kwargs.pop('while_offline')
try:
if not while_offline:
start_backup_info = PgBackupStatements.run_start_backup()
version = PgBackupStatements.pg_version()['version']
else:
if os.path.exists(
os.path.join(data_directory, 'postmaster.pid')):
hint = ('Shut down postgres. '
'If there is a stale lockfile, '
'then remove it after being very sure postgres '
'is not running.')
raise UserException(
msg='while_offline set, but pg looks to be running',
detail='Found a postmaster.pid lockfile, and aborting',
hint=hint)
ctrl_data = PgControlDataParser(data_directory)
start_backup_info = ctrl_data.last_xlog_file_name_and_offset()
version = ctrl_data.pg_version()
ret_tuple = self._upload_pg_cluster_dir(start_backup_info,
data_directory,
version=version,
*args,
**kwargs)
spec, uploaded_to, expanded_size_bytes = ret_tuple
upload_good = True
finally:
if not upload_good:
logger.warning(
'blocking on sending WAL segments',
detail=('The backup was not completed successfully, '
'but we have to wait anyway. '
'See README: TODO about pg_cancel_backup'))
if not while_offline:
stop_backup_info = PgBackupStatements.run_stop_backup()
else:
stop_backup_info = start_backup_info
backup_stop_good = True
# XXX: Ugly, this is more of a 'worker' task because it might
# involve retries and error messages, something that is not
# treated by the "operator" category of modules. So
# basically, if this small upload fails, the whole upload
# fails!
if upload_good and backup_stop_good:
# Try to write a sentinel file to the cluster backup
# directory that indicates that the base backup upload has
# definitely run its course and also communicates what WAL
# segments are needed to get to consistency.
sentinel_content = json.dumps({
'wal_segment_backup_stop':
stop_backup_info['file_name'],
'wal_segment_offset_backup_stop':
stop_backup_info['file_offset'],
'expanded_size_bytes':
expanded_size_bytes,
'spec':
spec
})
# XXX: should use the storage operators.
#
# XXX: distinguish sentinels by *PREFIX* not suffix,
# which makes searching harder. (For the next version
# bump).
uri_put_file(self.creds,
uploaded_to + '_backup_stop_sentinel.json',
BytesIO(sentinel_content.encode("utf8")),
content_type='application/json')
else:
# NB: Other exceptions should be raised before this that
# have more informative results, it is intended that this
# exception never will get raised.
raise UserCritical('could not complete backup process')