17
17
"""Core compression logic for compressing streams of related files."""
19
from __future__ import absolute_import
26
24
from bzrlib.lazy_import import lazy_import
27
25
lazy_import(globals(), """
28
26
from bzrlib import (
32
31
graph as _mod_graph,
48
48
AbsentContentFactory,
49
49
ChunkedContentFactory,
50
50
FulltextContentFactory,
51
VersionedFilesWithFallbacks,
54
54
# Minimum number of uncompressed bytes to try fetch at once when retrieving
55
55
# groupcompress blocks.
58
_USE_LZMA = False and (pylzma is not None)
60
58
# osutils.sha_string('')
61
59
_null_sha1 = 'da39a3ee5e6b4b0d3255bfef95601890afd80709'
151
149
self._content = ''
152
150
elif self._compressor_name == 'lzma':
153
151
# We don't do partial lzma decomp yet
154
153
self._content = pylzma.decompress(z_content)
155
154
elif self._compressor_name == 'zlib':
156
155
# Start a zlib decompressor
297
296
self._content = content
298
297
self._z_content_chunks = None
300
def _create_z_content_using_lzma(self):
301
if self._content_chunks is not None:
302
self._content = ''.join(self._content_chunks)
303
self._content_chunks = None
304
if self._content is None:
305
raise AssertionError('Nothing to compress')
306
z_content = pylzma.compress(self._content)
307
self._z_content_chunks = (z_content,)
308
self._z_content_length = len(z_content)
310
299
def _create_z_content_from_chunks(self, chunks):
311
300
compressor = zlib.compressobj(zlib.Z_DEFAULT_COMPRESSION)
312
301
# Peak in this point is 1 fulltext, 1 compressed text, + zlib overhead
320
309
def _create_z_content(self):
321
310
if self._z_content_chunks is not None:
324
self._create_z_content_using_lzma()
326
312
if self._content_chunks is not None:
327
313
chunks = self._content_chunks
332
318
def to_chunks(self):
333
319
"""Create the byte stream as a series of 'chunks'"""
334
320
self._create_z_content()
336
header = self.GCB_LZ_HEADER
338
header = self.GCB_HEADER
321
header = self.GCB_HEADER
339
322
chunks = ['%s%d\n%d\n'
340
323
% (header, self._z_content_length, self._content_length),
465
448
# Grab and cache the raw bytes for this entry
466
449
# and break the ref-cycle with _manager since we don't need it
468
self._manager._prepare_for_extract()
452
self._manager._prepare_for_extract()
453
except zlib.error as value:
454
raise errors.DecompressCorruption("zlib: " + str(value))
469
455
block = self._manager._block
470
456
self._bytes = block.extract(self.key, self._start, self._end)
471
457
# There are code paths that first extract as fulltext, and then
490
476
_full_enough_block_size = 3*1024*1024 # size at which we won't repack
491
477
_full_enough_mixed_block_size = 2*768*1024 # 1.5MB
493
def __init__(self, block):
479
def __init__(self, block, get_compressor_settings=None):
494
480
self._block = block
495
481
# We need to preserve the ordering
496
482
self._factories = []
497
483
self._last_byte = 0
484
self._get_settings = get_compressor_settings
485
self._compressor_settings = None
487
def _get_compressor_settings(self):
488
if self._compressor_settings is not None:
489
return self._compressor_settings
491
if self._get_settings is not None:
492
settings = self._get_settings()
494
vf = GroupCompressVersionedFiles
495
settings = vf._DEFAULT_COMPRESSOR_SETTINGS
496
self._compressor_settings = settings
497
return self._compressor_settings
499
499
def add_factory(self, key, parents, start, end):
500
500
if not self._factories:
533
533
new_block.set_content(self._block._content[:last_byte])
534
534
self._block = new_block
536
def _make_group_compressor(self):
537
return GroupCompressor(self._get_compressor_settings())
536
539
def _rebuild_block(self):
537
540
"""Create a new GroupCompressBlock with only the referenced texts."""
538
compressor = GroupCompressor()
541
compressor = self._make_group_compressor()
539
542
tstart = time.time()
540
543
old_length = self._block._content_length
553
556
# block? It seems hard to come up with a method that it would
554
557
# expand, since we do full compression again. Perhaps based on a
555
558
# request that ends up poorly ordered?
559
# TODO: If the content would have expanded, then we would want to
560
# handle a case where we need to split the block.
561
# Now that we have a user-tweakable option
562
# (max_bytes_to_index), it is possible that one person set it
563
# to a very low value, causing poor compression.
556
564
delta = time.time() - tstart
557
565
self._block = new_block
558
566
trace.mutter('creating new compressed block on-the-fly in %.3fs'
790
798
self.labels_deltas = {}
791
799
self._delta_index = None # Set by the children
792
800
self._block = GroupCompressBlock()
804
self._settings = settings
794
806
def compress(self, key, bytes, expected_sha, nostore_sha=None, soft=False):
795
807
"""Compress lines with label key.
911
923
class PythonGroupCompressor(_CommonGroupCompressor):
925
def __init__(self, settings=None):
914
926
"""Create a GroupCompressor.
916
928
Used only if the pyrex version is not available.
918
super(PythonGroupCompressor, self).__init__()
930
super(PythonGroupCompressor, self).__init__(settings)
919
931
self._delta_index = LinesDeltaIndex([])
920
932
# The actual content is managed by LinesDeltaIndex
921
933
self.chunks = self._delta_index.lines
959
971
It contains code very similar to SequenceMatcher because of having a similar
960
972
task. However some key differences apply:
961
- there is no junk, we want a minimal edit not a human readable diff.
962
- we don't filter very common lines (because we don't know where a good
963
range will start, and after the first text we want to be emitting minmal
965
- we chain the left side, not the right side
966
- we incrementally update the adjacency matrix as new lines are provided.
967
- we look for matches in all of the left side, so the routine which does
968
the analagous task of find_longest_match does not need to filter on the
974
* there is no junk, we want a minimal edit not a human readable diff.
975
* we don't filter very common lines (because we don't know where a good
976
range will start, and after the first text we want to be emitting minmal
978
* we chain the left side, not the right side
979
* we incrementally update the adjacency matrix as new lines are provided.
980
* we look for matches in all of the left side, so the routine which does
981
the analagous task of find_longest_match does not need to filter on the
973
super(PyrexGroupCompressor, self).__init__()
974
self._delta_index = DeltaIndex()
985
def __init__(self, settings=None):
986
super(PyrexGroupCompressor, self).__init__(settings)
987
max_bytes_to_index = self._settings.get('max_bytes_to_index', 0)
988
self._delta_index = DeltaIndex(max_bytes_to_index=max_bytes_to_index)
976
990
def _compress(self, key, bytes, max_delta_size, soft=False):
977
991
"""see _CommonGroupCompressor._compress"""
1069
1083
class _BatchingBlockFetcher(object):
1070
1084
"""Fetch group compress blocks in batches.
1072
1086
:ivar total_bytes: int of expected number of bytes needed to fetch the
1073
1087
currently pending batch.
1076
def __init__(self, gcvf, locations):
1090
def __init__(self, gcvf, locations, get_compressor_settings=None):
1077
1091
self.gcvf = gcvf
1078
1092
self.locations = locations
1127
1142
def yield_factories(self, full_flush=False):
1128
1143
"""Yield factories for keys added since the last yield. They will be
1129
1144
returned in the order they were added via add_key.
1131
1146
:param full_flush: by default, some results may not be returned in case
1132
1147
they can be part of the next batch. If full_flush is True, then
1133
1148
all results are returned.
1161
1176
memos_to_get_stack.pop()
1163
1178
block = self.batch_memos[read_memo]
1164
self.manager = _LazyGroupContentManager(block)
1179
self.manager = _LazyGroupContentManager(block,
1180
get_compressor_settings=self._get_compressor_settings)
1165
1181
self.last_read_memo = read_memo
1166
1182
start, end = index_memo[3:5]
1167
1183
self.manager.add_factory(key, parents, start, end)
1174
1190
self.total_bytes = 0
1177
class GroupCompressVersionedFiles(VersionedFiles):
1193
class GroupCompressVersionedFiles(VersionedFilesWithFallbacks):
1178
1194
"""A group-compress based VersionedFiles implementation."""
1180
def __init__(self, index, access, delta=True, _unadded_refs=None):
1196
# This controls how the GroupCompress DeltaIndex works. Basically, we
1197
# compute hash pointers into the source blocks (so hash(text) => text).
1198
# However each of these references costs some memory in trade against a
1199
# more accurate match result. For very large files, they either are
1200
# pre-compressed and change in bulk whenever they change, or change in just
1201
# local blocks. Either way, 'improved resolution' is not very helpful,
1202
# versus running out of memory trying to track everything. The default max
1203
# gives 100% sampling of a 1MB file.
1204
_DEFAULT_MAX_BYTES_TO_INDEX = 1024 * 1024
1205
_DEFAULT_COMPRESSOR_SETTINGS = {'max_bytes_to_index':
1206
_DEFAULT_MAX_BYTES_TO_INDEX}
1208
def __init__(self, index, access, delta=True, _unadded_refs=None,
1181
1210
"""Create a GroupCompressVersionedFiles object.
1183
1212
:param index: The index object storing access and graph data.
1184
1213
:param access: The access object storing raw data.
1185
1214
:param delta: Whether to delta compress or just entropy compress.
1186
1215
:param _unadded_refs: private parameter, don't use.
1216
:param _group_cache: private parameter, don't use.
1188
1218
self._index = index
1189
1219
self._access = access
1191
1221
if _unadded_refs is None:
1192
1222
_unadded_refs = {}
1193
1223
self._unadded_refs = _unadded_refs
1194
self._group_cache = LRUSizeCache(max_size=50*1024*1024)
1224
if _group_cache is None:
1225
_group_cache = LRUSizeCache(max_size=50*1024*1024)
1226
self._group_cache = _group_cache
1195
1227
self._immediate_fallback_vfs = []
1228
self._max_bytes_to_index = None
1197
1230
def without_fallbacks(self):
1198
1231
"""Return a clone of this object without any fallbacks configured."""
1199
1232
return GroupCompressVersionedFiles(self._index, self._access,
1200
self._delta, _unadded_refs=dict(self._unadded_refs))
1233
self._delta, _unadded_refs=dict(self._unadded_refs),
1234
_group_cache=self._group_cache)
1202
1236
def add_lines(self, key, parents, lines, parent_texts=None,
1203
1237
left_matching_blocks=None, nostore_sha=None, random_id=False,
1207
1241
:param key: The key tuple of the text to add.
1208
1242
:param parents: The parents key tuples of the text to add.
1209
1243
:param lines: A list of lines. Each line must be a bytestring. And all
1210
of them except the last must be terminated with \n and contain no
1211
other \n's. The last line may either contain no \n's or a single
1212
terminating \n. If the lines list does meet this constraint the add
1213
routine may error or may succeed - but you will be unable to read
1214
the data back accurately. (Checking the lines have been split
1244
of them except the last must be terminated with \\n and contain no
1245
other \\n's. The last line may either contain no \\n's or a single
1246
terminating \\n. If the lines list does meet this constraint the
1247
add routine may error or may succeed - but you will be unable to
1248
read the data back accurately. (Checking the lines have been split
1215
1249
correctly is expensive and extremely unlikely to catch bugs so it
1216
1250
is not done at runtime unless check_content is True.)
1217
1251
:param parent_texts: An optional dictionary containing the opaque
1312
1346
self._check_lines_not_unicode(lines)
1313
1347
self._check_lines_are_lines(lines)
1315
def get_known_graph_ancestry(self, keys):
1316
"""Get a KnownGraph instance with the ancestry of keys."""
1317
# Note that this is identical to
1318
# KnitVersionedFiles.get_known_graph_ancestry, but they don't share
1320
parent_map, missing_keys = self._index.find_ancestry(keys)
1321
for fallback in self._transitive_fallbacks():
1322
if not missing_keys:
1324
(f_parent_map, f_missing_keys) = fallback._index.find_ancestry(
1326
parent_map.update(f_parent_map)
1327
missing_keys = f_missing_keys
1328
kg = _mod_graph.KnownGraph(parent_map)
1331
1349
def get_parent_map(self, keys):
1332
1350
"""Get a map of the graph parents of keys.
1472
1490
The returned objects should be in the order defined by 'ordering',
1473
1491
which can weave between different sources.
1474
1493
:param ordering: Must be one of 'topological' or 'groupcompress'
1475
1494
:return: List of [(source, [keys])] tuples, such that all keys are in
1476
1495
the defined order, regardless of source.
1571
1590
# - we encounter an unadded ref, or
1572
1591
# - we run out of keys, or
1573
1592
# - the total bytes to retrieve for this batch > BATCH_SIZE
1574
batcher = _BatchingBlockFetcher(self, locations)
1593
batcher = _BatchingBlockFetcher(self, locations,
1594
get_compressor_settings=self._get_compressor_settings)
1575
1595
for source, keys in source_keys:
1576
1596
if source is self:
1577
1597
for key in keys:
1623
1643
for _ in self._insert_record_stream(stream, random_id=False):
1646
def _get_compressor_settings(self):
1647
if self._max_bytes_to_index is None:
1648
# TODO: VersionedFiles don't know about their containing
1649
# repository, so they don't have much of an idea about their
1650
# location. So for now, this is only a global option.
1651
c = config.GlobalConfig()
1652
val = c.get_user_option('bzr.groupcompress.max_bytes_to_index')
1656
except ValueError, e:
1657
trace.warning('Value for '
1658
'"bzr.groupcompress.max_bytes_to_index"'
1659
' %r is not an integer'
1663
val = self._DEFAULT_MAX_BYTES_TO_INDEX
1664
self._max_bytes_to_index = val
1665
return {'max_bytes_to_index': self._max_bytes_to_index}
1667
def _make_group_compressor(self):
1668
return GroupCompressor(self._get_compressor_settings())
1626
1670
def _insert_record_stream(self, stream, random_id=False, nostore_sha=None,
1627
1671
reuse_blocks=True):
1628
1672
"""Internal core to insert a record stream into this container.
1652
1696
# This will go up to fulltexts for gc to gc fetching, which isn't
1654
self._compressor = GroupCompressor()
1698
self._compressor = self._make_group_compressor()
1655
1699
self._unadded_refs = {}
1656
1700
keys_to_add = []
1658
1702
bytes_len, chunks = self._compressor.flush().to_chunks()
1659
self._compressor = GroupCompressor()
1703
self._compressor = self._make_group_compressor()
1660
1704
# Note: At this point we still have 1 copy of the fulltext (in
1661
1705
# record and the var 'bytes'), and this generates 2 copies of
1662
1706
# the compressed text (one for bytes, one in chunks)
1692
1736
raise errors.RevisionNotPresent(record.key, self)
1694
1738
if record.key in inserted_keys:
1695
trace.note('Insert claimed random_id=True,'
1696
' but then inserted %r two times', record.key)
1739
trace.note(gettext('Insert claimed random_id=True,'
1740
' but then inserted %r two times'), record.key)
1698
1742
inserted_keys.add(record.key)
1699
1743
if reuse_blocks:
2073
2117
:param keys: An iterable of keys.
2074
2118
:return: A dict of key:
2075
2119
(index_memo, compression_parent, parents, record_details).
2077
opaque structure to pass to read_records to extract the raw
2080
Content that this record is built upon, may be None
2082
Logical parents of this node
2084
extra information about the content which needs to be passed to
2085
Factory.parse_record
2121
* index_memo: opaque structure to pass to read_records to extract
2123
* compression_parent: Content that this record is built upon, may
2125
* parents: Logical parents of this node
2126
* record_details: extra information about the content which needs
2127
to be passed to Factory.parse_record
2087
2129
self._check_read()
added
removed
bzrlib/groupcompress.py
