Commit 1213f8a6 authored by Christian Theune's avatar Christian Theune

Implemented new bushy layout for blob directories including backwards

compatibility for the old `lawn` layout.
parent de31996b
......@@ -8,6 +8,11 @@ Change History
New Features
------------
- Changed layout strategy for the blob directory to a bushy approach (8 levels
deep, at most ~256 entries per directory level, one directory for each
blob). Old directories are automatically detected and will be handled with
the old strategy.
- Versions are no-longer supported.
- ZEO cache files can be larger than 4G. Note that older ZEO cache
......
......@@ -855,9 +855,7 @@ class ClientStorage(object):
def _storeBlob_shared(self, oid, serial, data, filename, txn):
# First, move the blob into the blob directory
dir = self.fshelper.getPathForOID(oid)
if not os.path.exists(dir):
os.mkdir(dir)
self.fshelper.getPathForOID(oid, create=True)
fd, target = self.fshelper.blob_mkstemp(oid, serial)
os.close(fd)
......@@ -924,14 +922,7 @@ class ClientStorage(object):
raise POSException.POSKeyError("No blob file", oid, serial)
# First, we'll create the directory for this oid, if it doesn't exist.
targetpath = self.fshelper.getPathForOID(oid)
if not os.path.exists(targetpath):
try:
os.makedirs(targetpath, 0700)
except OSError:
# We might have lost a race. If so, the directory
# must exist now
assert os.path.exists(targetpath)
targetpath = self.fshelper.getPathForOID(oid, create=True)
# OK, it's not here and we (or someone) needs to get it. We
# want to avoid getting it multiple times. We want to avoid
......@@ -1118,19 +1109,15 @@ class ClientStorage(object):
assert s == tid, (s, tid)
self._cache.store(oid, s, None, data)
if self.fshelper is not None:
blobs = self._tbuf.blobs
while blobs:
oid, blobfilename = blobs.pop()
targetpath = self.fshelper.getPathForOID(oid)
if not os.path.exists(targetpath):
os.makedirs(targetpath, 0700)
targetpath = self.fshelper.getPathForOID(oid, create=True)
rename_or_copy_blob(blobfilename,
self.fshelper.getBlobFilename(oid, tid),
)
self._tbuf.clear()
def undo(self, trans_id, txn):
......
......@@ -515,8 +515,7 @@ class CommonBlobTests:
self._storage.tpc_abort(t)
raise
self.assert_(not os.path.exists(tfname))
filename = os.path.join(self.blobdir, oid_repr(oid),
tid_repr(revid) + BLOB_SUFFIX)
filename = self._storage.fshelper.getBlobFilename(oid, revid)
self.assert_(os.path.exists(filename))
self.assertEqual(somedata, open(filename).read())
......@@ -630,18 +629,16 @@ class BlobAdaptedFileStorageTests(GenericTests, CommonBlobTests):
d1 = f.read(8096)
d2 = somedata.read(8096)
self.assertEqual(d1, d2)
# The file should have been copied to the server:
filename = os.path.join(self.blobdir, oid_repr(oid),
tid_repr(revid) + BLOB_SUFFIX)
check_data(filename)
# It should also be in the cache:
filename = os.path.join(self.blob_cache_dir, oid_repr(oid),
tid_repr(revid) + BLOB_SUFFIX)
# The file should be in the cache ...
filename = self._storage.fshelper.getBlobFilename(oid, revid)
check_data(filename)
# ... and on the server
server_filename = filename.replace(self.blob_cache_dir, self.blobdir)
self.assert_(server_filename.startswith(self.blobdir))
check_data(server_filename)
# If we remove it from the cache and call loadBlob, it should
# come back. We can do this in many threads. We'll instrument
# the method that is used to request data from teh server to
......
......@@ -15,8 +15,10 @@
"""
import base64
import binascii
import logging
import os
import re
import shutil
import stat
import sys
......@@ -43,6 +45,9 @@ logger = logging.getLogger('ZODB.blob')
BLOB_SUFFIX = ".blob"
SAVEPOINT_SUFFIX = ".spb"
LAYOUT_MARKER = '.layout'
LAYOUTS = {}
valid_modes = 'r', 'w', 'r+', 'a'
# Threading issues:
......@@ -292,21 +297,42 @@ class FilesystemHelper:
# with blobs and storages needn't indirect through this if they
# want to perform blob storage differently.
def __init__(self, base_dir):
self.base_dir = base_dir
def __init__(self, base_dir, layout_name='automatic'):
self.base_dir = os.path.normpath(base_dir) + '/'
self.temp_dir = os.path.join(base_dir, 'tmp')
if layout_name == 'automatic':
layout_name = auto_layout_select(base_dir)
if layout_name == 'lawn':
log('The `lawn` blob directory layout is deprecated due to '
'scalability issues on some file systems, please consider '
'migrating to the `bushy` layout.', level=logging.WARN)
self.layout_name = layout_name
self.layout = LAYOUTS[layout_name]
def create(self):
if not os.path.exists(self.base_dir):
os.makedirs(self.base_dir, 0700)
log("Blob cache directory '%s' does not exist. "
"Created new directory." % self.base_dir,
level=logging.INFO)
log("Blob directory '%s' does not exist. "
"Created new directory." % self.base_dir)
if not os.path.exists(self.temp_dir):
os.makedirs(self.temp_dir, 0700)
log("Blob temporary directory '%s' does not exist. "
"Created new directory." % self.temp_dir,
level=logging.INFO)
"Created new directory." % self.temp_dir)
if not os.path.exists(os.path.join(self.base_dir, LAYOUT_MARKER)):
layout_marker = open(
os.path.join(self.base_dir, LAYOUT_MARKER), 'wb')
layout_marker.write(self.layout_name)
else:
layout_marker = open(
os.path.join(self.base_dir, LAYOUT_MARKER), 'rb')
layout = layout_marker.read().strip()
if layout != self.layout_name:
raise ValueError(
"Directory layout `%s` selected for blob directory %s, but "
"marker found for layout `%s`" %
(self.layout_name, self.base_dir, layout))
def isSecure(self, path):
"""Ensure that (POSIX) path mode bits are 0700."""
......@@ -317,12 +343,40 @@ class FilesystemHelper:
log('Blob dir %s has insecure mode setting' % self.base_dir,
level=logging.WARNING)
def getPathForOID(self, oid):
def getPathForOID(self, oid, create=False):
"""Given an OID, return the path on the filesystem where
the blob data relating to that OID is stored.
If the create flag is given, the path is also created if it didn't
exist already.
"""
# OIDs are numbers and sometimes passed around as integers. For our
# computations we rely on the 64-bit packed string representation.
if isinstance(oid, int):
oid = utils.p64(oid)
path = self.layout.oid_to_path(oid)
path = os.path.join(self.base_dir, path)
if create and not os.path.exists(path):
try:
os.makedirs(path, 0700)
except OSError:
# We might have lost a race. If so, the directory
# must exist now
assert os.path.exists(targetpath)
return path
def getOIDForPath(self, path):
"""Given a path, return an OID, if the path is a valid path for an
OID. The inverse function to `getPathForOID`.
Raises ValueError if the path is not valid for an OID.
"""
return os.path.join(self.base_dir, utils.oid_repr(oid))
path = path[len(self.base_dir):]
return self.layout.path_to_oid(path)
def getBlobFilename(self, oid, tid):
"""Given an oid and a tid, return the full filename of the
......@@ -330,6 +384,10 @@ class FilesystemHelper:
"""
oid_path = self.getPathForOID(oid)
# TIDs are numbers and sometimes passed around as integers. For our
# computations we rely on the 64-bit packed string representation
if isinstance(tid, int):
tid = utils.p64(tid)
filename = "%s%s" % (utils.tid_repr(tid), BLOB_SUFFIX)
return os.path.join(oid_path, filename)
......@@ -359,10 +417,9 @@ class FilesystemHelper:
if not filename.endswith(BLOB_SUFFIX):
return None, None
path, filename = os.path.split(filename)
oid = os.path.split(path)[1]
oid = self.getOIDForPath(path)
serial = filename[:-len(BLOB_SUFFIX)]
oid = utils.repr_to_oid(oid)
serial = utils.repr_to_oid(serial)
return oid, serial
......@@ -372,24 +429,104 @@ class FilesystemHelper:
"""
oids = []
base_dir = self.base_dir
for oidpath in os.listdir(base_dir):
for filename in os.listdir(os.path.join(base_dir, oidpath)):
blob_path = os.path.join(base_dir, oidpath, filename)
for oid, oidpath in self.listOIDs():
for filename in os.listdir(oidpath):
blob_path = os.path.join(oidpath, filename)
oid, serial = self.splitBlobFilename(blob_path)
if search_serial == serial:
oids.append(oid)
return oids
def listOIDs(self):
"""Lists all OIDs and their paths.
"""Iterates over all paths under the base directory that contain blob
files.
"""
for candidate in os.listdir(self.base_dir):
if candidate == 'tmp':
for path, dirs, files in os.walk(self.base_dir):
try:
oid = self.getOIDForPath(path)
except ValueError:
continue
oid = utils.repr_to_oid(candidate)
yield oid, self.getPathForOID(oid)
yield oid, path
def auto_layout_select(path):
# A heuristic to look at a path and determine which directory layout to
# use. Basically we try to figure out if the directory is either already
# used and contains an explicit marker, is unused or used without a
# marker.
layout_marker = os.path.join(path, LAYOUT_MARKER)
if not os.path.exists(path):
log('Blob directory %s does not exist. '
'Selected `bushy` layout. ' % path)
layout = 'bushy'
elif len(os.listdir(path)) == 0:
log('Blob directory %s is unused and has no layout marker set.'
'Selected `bushy` layout. ' % path)
layout = 'bushy'
elif LAYOUT_MARKER not in os.listdir(path):
log('Blob directory %s is used but has no layout marker set.'
'Selected `lawn` layout. ' % path)
layout = 'lawn'
else:
layout = open(layout_marker, 'rb').read()
layout = layout.strip()
log('Blob directory %s has layout marker set.'
'Selected `%s` layout. ' % (path, layout))
return layout
class BushyLayout(object):
"""A bushy directory layout for blob directories.
Creates an 8-level directory structure (one level per byte) in
little-endian order from the OID of an object.
"""
blob_path_pattern = r'^' + (r'0x[0-9]{1,2}/*'*8) + r'$'
blob_path_pattern = re.compile(blob_path_pattern)
def oid_to_path(self, oid):
directories = []
# Create the bushy directory structure with the least significant byte
# first
for byte in reversed(str(oid)):
directories.append('0x%s' % binascii.hexlify(byte))
return '/'.join(directories)
def path_to_oid(self, path):
if self.blob_path_pattern.match(path) is None:
raise ValueError("Not a valid OID path: %s" % path)
# The path always has a leading slash that we need to ignore.
path = path.split('/')
# The path contains the OID in little endian form but the OID itself
# is big endian.
path.reverse()
# Each path segment stores a byte in hex representation. Turn it into
# an int and then get the character for our byte string.
oid = ''.join(binascii.unhexlify(byte[2:]) for byte in path)
return oid
LAYOUTS['bushy'] = BushyLayout()
class LawnLayout(object):
"""A shallow directory layout for blob directories.
Creates a single level of directories (one for each oid).
"""
def oid_to_path(self, oid):
return utils.oid_repr(oid)
def path_to_oid(self, path):
try:
return utils.repr_to_oid(path)
except TypeError:
raise ValueError('Not a valid OID path: %s' % path)
LAYOUTS['lawn'] = LawnLayout()
class BlobStorage(SpecificationDecoratorBase):
......@@ -401,13 +538,13 @@ class BlobStorage(SpecificationDecoratorBase):
# us to have instance attributes explicitly on the proxy.
__slots__ = ('fshelper', 'dirty_oids', '_BlobStorage__supportsUndo')
def __new__(self, base_directory, storage):
def __new__(self, base_directory, storage, layout='automatic'):
return SpecificationDecoratorBase.__new__(self, storage)
def __init__(self, base_directory, storage):
def __init__(self, base_directory, storage, layout='automatic'):
# XXX Log warning if storage is ClientStorage
SpecificationDecoratorBase.__init__(self, storage)
self.fshelper = FilesystemHelper(base_directory)
self.fshelper = FilesystemHelper(base_directory, layout)
self.fshelper.create()
self.fshelper.checkSecure()
self.dirty_oids = []
......@@ -438,10 +575,7 @@ class BlobStorage(SpecificationDecoratorBase):
self._lock_acquire()
try:
targetpath = self.fshelper.getPathForOID(oid)
if not os.path.exists(targetpath):
os.makedirs(targetpath, 0700)
self.fshelper.getPathForOID(oid, create=True)
targetname = self.fshelper.getBlobFilename(oid, serial)
rename_or_copy_blob(blobfilename, targetname)
......@@ -487,14 +621,12 @@ class BlobStorage(SpecificationDecoratorBase):
# if they are still needed by attempting to load the revision
# of that object from the database. This is maybe the slowest
# possible way to do this, but it's safe.
base_dir = self.fshelper.base_dir
for oid, oid_path in self.fshelper.listOIDs():
files = os.listdir(oid_path)
for filename in files:
filepath = os.path.join(oid_path, filename)
whatever, serial = self.fshelper.splitBlobFilename(filepath)
try:
fn = self.fshelper.getBlobFilename(oid, serial)
self.loadSerial(oid, serial)
except POSKeyError:
remove_committed(filepath)
......@@ -504,7 +636,6 @@ class BlobStorage(SpecificationDecoratorBase):
@non_overridable
def _packNonUndoing(self, packtime, referencesf):
base_dir = self.fshelper.base_dir
for oid, oid_path in self.fshelper.listOIDs():
exists = True
try:
......@@ -552,15 +683,11 @@ class BlobStorage(SpecificationDecoratorBase):
"""Return the size of the database in bytes."""
orig_size = getProxiedObject(self).getSize()
blob_size = 0
base_dir = self.fshelper.base_dir
for oid in os.listdir(base_dir):
sub_dir = os.path.join(base_dir, oid)
if not os.path.isdir(sub_dir):
continue
for serial in os.listdir(sub_dir):
for oid, path in self.fshelper.listOIDs():
for serial in os.listdir(path):
if not serial.endswith(BLOB_SUFFIX):
continue
file_path = os.path.join(base_dir, oid, serial)
file_path = os.path.join(path, serial)
blob_size += os.stat(file_path).st_size
return orig_size + blob_size
......@@ -583,7 +710,6 @@ class BlobStorage(SpecificationDecoratorBase):
# we get all the blob oids on the filesystem related to the
# transaction we want to undo.
for oid in self.fshelper.getOIDsForSerial(serial_id):
# we want to find the serial id of the previous revision
# of this blob object.
load_result = self.loadBefore(oid, serial_id)
......
======================
Blob directory layouts
======================
The internal structure of the blob directories is governed by so called
`layouts`. The current default layout is called `bushy`.
The original blob implementation used a layout that we now call `lawn` and
which is still available for backwards compatibility.
Layouts implement two methods: one for computing a relative path for an
OID and one for turning a relative path back into an OID.
Our terminology is roughly the same as used in `DirectoryStorage`.
The `bushy` layout
==================
The bushy layout splits the OID into the 8 byte parts, reverses them and
creates one directory level for each part, named by the hexlified
representation of the byte value. This results in 8 levels of directories, the
leaf directories being used for the revisions of the blobs and at most 256
entries per directory level:
>>> from ZODB.blob import BushyLayout
>>> bushy = BushyLayout()
>>> bushy.oid_to_path('\x00\x00\x00\x00\x00\x00\x00\x00')
'0x00/0x00/0x00/0x00/0x00/0x00/0x00/0x00'
>>> bushy.oid_to_path('\x00\x00\x00\x00\x00\x00\x00\x01')
'0x01/0x00/0x00/0x00/0x00/0x00/0x00/0x00'
>>> bushy.path_to_oid('0x01/0x00/0x00/0x00/0x00/0x00/0x00/0x00')
'\x00\x00\x00\x00\x00\x00\x00\x01'
Paths that do not represent an OID will cause a ValueError:
>>> bushy.path_to_oid('tmp')
Traceback (most recent call last):
ValueError: Not a valid OID path: tmp
The `lawn` layout
=================
The lawn layout creates on directory for each blob named by the blob's hex
representation of its OID. This has some limitations on various file systems
like performance penalties or the inability to store more than a given number
of blobs at the same time (e.g. 32k on ext3).
>>> from ZODB.blob import LawnLayout
>>> lawn = LawnLayout()
>>> lawn.oid_to_path('\x00\x00\x00\x00\x00\x00\x00\x00')
'0x00'
>>> lawn.oid_to_path('\x00\x00\x00\x00\x00\x00\x00\x01')
'0x01'
>>> lawn.path_to_oid('0x01')
'\x00\x00\x00\x00\x00\x00\x00\x01'
Paths that do not represent an OID will cause a ValueError:
>>> lawn.path_to_oid('tmp')
Traceback (most recent call last):
ValueError: Not a valid OID path: tmp
Auto-detecting the layout of a directory
========================================
To allow easier migration, we provide an auto-detection feature that analyses a
blob directory and decides for a strategy to use. In general it prefers to
choose the `bushy` layout, except if it determines that the directory has
already been used to create a lawn structure.
>>> from ZODB.blob import auto_layout_select
1. Non-existing directories will trigger a bushy layout:
>>> import tempfile
>>> import shutil
>>> d = tempfile.mkdtemp()
>>> shutil.rmtree(d)
>>> auto_layout_select(d)
'bushy'
2. Empty directories will trigger a bushy layout too:
>>> d = tempfile.mkdtemp()
>>> auto_layout_select(d)
'bushy'
3. If the directory contains a marker for the strategy it will be used:
>>> from ZODB.blob import LAYOUT_MARKER
>>> import os.path
>>> open(os.path.join(d, LAYOUT_MARKER), 'wb').write('bushy')
>>> auto_layout_select(d)
'bushy'
>>> open(os.path.join(d, LAYOUT_MARKER), 'wb').write('lawn')
>>> auto_layout_select(d)
'lawn'
>>> shutil.rmtree(d)
4. If the directory does not contain a marker but other files, we assume that
it was created with an earlier version of the blob implementation and uses our
`lawn` layout:
>>> d = tempfile.mkdtemp()
>>> open(os.path.join(d, '0x0101'), 'wb').write('foo')
>>> auto_layout_select(d)
'lawn'
>>> shutil.rmtree(d)
Directory layout markers
========================
When the file system helper (FSH) is asked to create the directory structure,
it will leave a marker with the choosen layout if no marker exists yet:
>>> from ZODB.blob import FilesystemHelper
>>> d = tempfile.mkdtemp()
>>> blobs = os.path.join(d, 'blobs')
>>> fsh = FilesystemHelper(blobs)
>>> fsh.layout_name
'bushy'
>>> fsh.create()
>>> open(os.path.join(blobs, LAYOUT_MARKER), 'rb').read()
'bushy'
If the FSH finds a marker, then it verifies whether its content matches the
strategy that was chosen. It will raise an exception if we try to work with a
directory that has a different marker than the chosen strategy:
>>> fsh = FilesystemHelper(blobs, 'lawn')
>>> fsh.layout_name
'lawn'
>>> fsh.create() # doctest: +ELLIPSIS
Traceback (most recent call last):
ValueError: Directory layout `lawn` selected for blob directory /.../blobs/, but marker found for layout `bushy`
>>> shutil.rmtree(blobs)
This function interacts with the automatic detection in the way, that an
unmarked directory will be marked the first time when it is auto-guessed and
the marker will be used in the future:
>>> import ZODB.FileStorage
>>> from ZODB.blob import BlobStorage
>>> datafs = os.path.join(d, 'data.fs')
>>> base_storage = ZODB.FileStorage.FileStorage(datafs)
>>> os.mkdir(blobs)
>>> open(os.path.join(blobs, 'foo'), 'wb').write('foo')
>>> blob_storage = BlobStorage(blobs, base_storage)
>>> blob_storage.fshelper.layout_name
'lawn'
>>> open(os.path.join(blobs, LAYOUT_MARKER), 'rb').read()
'lawn'
>>> blob_storage = BlobStorage(blobs, base_storage, layout='bushy') # doctest: +ELLIPSIS
Traceback (most recent call last):
ValueError: Directory layout `bushy` selected for blob directory /.../blobs/, but marker found for layout `lawn`
>>> shutil.rmtree(d)
......@@ -32,7 +32,7 @@ First, we need a datatabase with blob support::
>>> from ZODB.DB import DB
>>> from tempfile import mkdtemp
>>> import os.path
>>> base_storage = MappingStorage("test")
>>> base_storage = MappingStorage('test')
>>> blob_dir = mkdtemp()
>>> blob_storage = BlobStorage(blob_dir, base_storage)
>>> database = DB(blob_storage)
......
......@@ -322,9 +322,9 @@ clean up dirty files:
>>> base_storage = DummyBaseStorage()
>>> blob_dir2 = mkdtemp()
>>> blob_storage2 = BlobStorage(blob_dir2, base_storage)
>>> committed_blob_dir = os.path.join(blob_dir2, '0')
>>> committed_blob_file = os.path.join(committed_blob_dir, '0.blob')
>>> os.mkdir(committed_blob_dir)
>>> committed_blob_dir = blob_storage2.fshelper.getPathForOID(0)
>>> os.makedirs(committed_blob_dir)
>>> committed_blob_file = blob_storage2.fshelper.getBlobFilename(0, 0)
>>> open(os.path.join(committed_blob_file), 'w').write('foo')
>>> os.path.exists(committed_blob_file)
True
......
......@@ -105,7 +105,6 @@ class BlobUndoTests(unittest.TestCase):
self.here = os.getcwd()
os.chdir(self.test_dir)
self.storagefile = 'Data.fs'
os.mkdir('blobs')
self.blob_dir = 'blobs'
def tearDown(self):
......@@ -483,7 +482,7 @@ def loadblob_tmpstore():
We can access the blob correctly:
>>> tmpstore.loadBlob(blob_oid, tid) # doctest: +ELLIPSIS
'.../0x01/0x...blob'
'.../0x01/0x00/0x00/0x00/0x00/0x00/0x00/0x00/0x...blob'
Clean up:
......@@ -503,7 +502,7 @@ def test_suite():
suite.addTest(doctest.DocFileSuite(
"blob_basic.txt", "blob_connection.txt", "blob_transaction.txt",
"blob_packing.txt", "blob_importexport.txt", "blob_consume.txt",
"blob_tempdir.txt",
"blob_tempdir.txt", "blob_layout.txt",
setUp=ZODB.tests.util.setUp,
tearDown=ZODB.tests.util.tearDown,
))
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment