143
138
self._content = ''.join(self._content_chunks)
144
139
self._content_chunks = None
145
140
if self._content is None:
146
# We join self._z_content_chunks here, because if we are
147
# decompressing, then it is *very* likely that we have a single
149
if self._z_content_chunks is None:
141
if self._z_content is None:
150
142
raise AssertionError('No content to decompress')
151
z_content = ''.join(self._z_content_chunks)
143
if self._z_content == '':
153
144
self._content = ''
154
145
elif self._compressor_name == 'lzma':
155
146
# We don't do partial lzma decomp yet
156
self._content = pylzma.decompress(z_content)
147
self._content = pylzma.decompress(self._z_content)
157
148
elif self._compressor_name == 'zlib':
158
149
# Start a zlib decompressor
159
if num_bytes * 4 > self._content_length * 3:
160
# If we are requesting more that 3/4ths of the content,
161
# just extract the whole thing in a single pass
162
num_bytes = self._content_length
163
self._content = zlib.decompress(z_content)
150
if num_bytes is None:
151
self._content = zlib.decompress(self._z_content)
165
153
self._z_content_decompressor = zlib.decompressobj()
166
154
# Seed the decompressor with the uncompressed bytes, so
167
155
# that the rest of the code is simplified
168
156
self._content = self._z_content_decompressor.decompress(
169
z_content, num_bytes + _ZLIB_DECOMP_WINDOW)
170
if not self._z_content_decompressor.unconsumed_tail:
171
self._z_content_decompressor = None
157
self._z_content, num_bytes + _ZLIB_DECOMP_WINDOW)
173
159
raise AssertionError('Unknown compressor: %r'
174
160
% self._compressor_name)
176
162
# 'unconsumed_tail'
178
164
# Do we have enough bytes already?
179
if len(self._content) >= num_bytes:
165
if num_bytes is not None and len(self._content) >= num_bytes:
167
if num_bytes is None and self._z_content_decompressor is None:
168
# We must have already decompressed everything
181
170
# If we got this far, and don't have a decompressor, something is wrong
182
171
if self._z_content_decompressor is None:
183
172
raise AssertionError(
184
173
'No decompressor to decompress %d bytes' % num_bytes)
185
174
remaining_decomp = self._z_content_decompressor.unconsumed_tail
186
if not remaining_decomp:
187
raise AssertionError('Nothing left to decompress')
188
needed_bytes = num_bytes - len(self._content)
189
# We always set max_size to 32kB over the minimum needed, so that
190
# zlib will give us as much as we really want.
191
# TODO: If this isn't good enough, we could make a loop here,
192
# that keeps expanding the request until we get enough
193
self._content += self._z_content_decompressor.decompress(
194
remaining_decomp, needed_bytes + _ZLIB_DECOMP_WINDOW)
195
if len(self._content) < num_bytes:
196
raise AssertionError('%d bytes wanted, only %d available'
197
% (num_bytes, len(self._content)))
198
if not self._z_content_decompressor.unconsumed_tail:
199
# The stream is finished
200
self._z_content_decompressor = None
175
if num_bytes is None:
177
# We don't know how much is left, but we'll decompress it all
178
self._content += self._z_content_decompressor.decompress(
180
# Note: There's what I consider a bug in zlib.decompressobj
181
# If you pass back in the entire unconsumed_tail, only
182
# this time you don't pass a max-size, it doesn't
183
# change the unconsumed_tail back to None/''.
184
# However, we know we are done with the whole stream
185
self._z_content_decompressor = None
186
# XXX: Why is this the only place in this routine we set this?
187
self._content_length = len(self._content)
189
if not remaining_decomp:
190
raise AssertionError('Nothing left to decompress')
191
needed_bytes = num_bytes - len(self._content)
192
# We always set max_size to 32kB over the minimum needed, so that
193
# zlib will give us as much as we really want.
194
# TODO: If this isn't good enough, we could make a loop here,
195
# that keeps expanding the request until we get enough
196
self._content += self._z_content_decompressor.decompress(
197
remaining_decomp, needed_bytes + _ZLIB_DECOMP_WINDOW)
198
if len(self._content) < num_bytes:
199
raise AssertionError('%d bytes wanted, only %d available'
200
% (num_bytes, len(self._content)))
201
if not self._z_content_decompressor.unconsumed_tail:
202
# The stream is finished
203
self._z_content_decompressor = None
202
205
def _parse_bytes(self, bytes, pos):
203
206
"""Read the various lengths from the header.
305
298
self._content_chunks = None
306
299
if self._content is None:
307
300
raise AssertionError('Nothing to compress')
308
z_content = pylzma.compress(self._content)
309
self._z_content_chunks = (z_content,)
310
self._z_content_length = len(z_content)
301
self._z_content = pylzma.compress(self._content)
302
self._z_content_length = len(self._z_content)
312
def _create_z_content_from_chunks(self, chunks):
304
def _create_z_content_from_chunks(self):
313
305
compressor = zlib.compressobj(zlib.Z_DEFAULT_COMPRESSION)
314
# Peak in this point is 1 fulltext, 1 compressed text, + zlib overhead
315
# (measured peak is maybe 30MB over the above...)
316
compressed_chunks = map(compressor.compress, chunks)
306
compressed_chunks = map(compressor.compress, self._content_chunks)
317
307
compressed_chunks.append(compressor.flush())
318
# Ignore empty chunks
319
self._z_content_chunks = [c for c in compressed_chunks if c]
320
self._z_content_length = sum(map(len, self._z_content_chunks))
308
self._z_content = ''.join(compressed_chunks)
309
self._z_content_length = len(self._z_content)
322
311
def _create_z_content(self):
323
if self._z_content_chunks is not None:
312
if self._z_content is not None:
326
315
self._create_z_content_using_lzma()
328
317
if self._content_chunks is not None:
329
chunks = self._content_chunks
331
chunks = (self._content,)
332
self._create_z_content_from_chunks(chunks)
318
self._create_z_content_from_chunks()
320
self._z_content = zlib.compress(self._content)
321
self._z_content_length = len(self._z_content)
335
"""Create the byte stream as a series of 'chunks'"""
324
"""Encode the information into a byte stream."""
336
325
self._create_z_content()
338
327
header = self.GCB_LZ_HEADER
340
329
header = self.GCB_HEADER
341
chunks = ['%s%d\n%d\n'
342
% (header, self._z_content_length, self._content_length),
331
'%d\n%d\n' % (self._z_content_length, self._content_length),
344
chunks.extend(self._z_content_chunks)
345
total_len = sum(map(len, chunks))
346
return total_len, chunks
349
"""Encode the information into a byte stream."""
350
total_len, chunks = self.to_chunks()
351
334
return ''.join(chunks)
353
336
def _dump(self, include_text=False):
487
468
class _LazyGroupContentManager(object):
488
469
"""This manages a group of _LazyGroupCompressFactory objects."""
490
_max_cut_fraction = 0.75 # We allow a block to be trimmed to 75% of
491
# current size, and still be considered
493
_full_block_size = 4*1024*1024
494
_full_mixed_block_size = 2*1024*1024
495
_full_enough_block_size = 3*1024*1024 # size at which we won't repack
496
_full_enough_mixed_block_size = 2*768*1024 # 1.5MB
498
def __init__(self, block, get_compressor_settings=None):
471
def __init__(self, block):
499
472
self._block = block
500
473
# We need to preserve the ordering
501
474
self._factories = []
502
475
self._last_byte = 0
503
self._get_settings = get_compressor_settings
504
self._compressor_settings = None
506
def _get_compressor_settings(self):
507
if self._compressor_settings is not None:
508
return self._compressor_settings
510
if self._get_settings is not None:
511
settings = self._get_settings()
513
vf = GroupCompressVersionedFiles
514
settings = vf._DEFAULT_COMPRESSOR_SETTINGS
515
self._compressor_settings = settings
516
return self._compressor_settings
518
477
def add_factory(self, key, parents, start, end):
519
478
if not self._factories:
594
545
# time (self._block._content) is a little expensive.
595
546
self._block._ensure_content(self._last_byte)
597
def _check_rebuild_action(self):
548
def _check_rebuild_block(self):
598
549
"""Check to see if our block should be repacked."""
599
550
total_bytes_used = 0
600
551
last_byte_used = 0
601
552
for factory in self._factories:
602
553
total_bytes_used += factory._end - factory._start
603
if last_byte_used < factory._end:
604
last_byte_used = factory._end
605
# If we are using more than half of the bytes from the block, we have
606
# nothing else to check
554
last_byte_used = max(last_byte_used, factory._end)
555
# If we are using most of the bytes from the block, we have nothing
556
# else to check (currently more that 1/2)
607
557
if total_bytes_used * 2 >= self._block._content_length:
608
return None, last_byte_used, total_bytes_used
609
# We are using less than 50% of the content. Is the content we are
610
# using at the beginning of the block? If so, we can just trim the
611
# tail, rather than rebuilding from scratch.
559
# Can we just strip off the trailing bytes? If we are going to be
560
# transmitting more than 50% of the front of the content, go ahead
612
561
if total_bytes_used * 2 > last_byte_used:
613
return 'trim', last_byte_used, total_bytes_used
562
self._trim_block(last_byte_used)
615
565
# We are using a small amount of the data, and it isn't just packed
616
566
# nicely at the front, so rebuild the content.
623
573
# expanding many deltas into fulltexts, as well.
624
574
# If we build a cheap enough 'strip', then we could try a strip,
625
575
# if that expands the content, we then rebuild.
626
return 'rebuild', last_byte_used, total_bytes_used
628
def check_is_well_utilized(self):
629
"""Is the current block considered 'well utilized'?
631
This heuristic asks if the current block considers itself to be a fully
632
developed group, rather than just a loose collection of data.
634
if len(self._factories) == 1:
635
# A block of length 1 could be improved by combining with other
636
# groups - don't look deeper. Even larger than max size groups
637
# could compress well with adjacent versions of the same thing.
639
action, last_byte_used, total_bytes_used = self._check_rebuild_action()
640
block_size = self._block._content_length
641
if total_bytes_used < block_size * self._max_cut_fraction:
642
# This block wants to trim itself small enough that we want to
643
# consider it under-utilized.
645
# TODO: This code is meant to be the twin of _insert_record_stream's
646
# 'start_new_block' logic. It would probably be better to factor
647
# out that logic into a shared location, so that it stays
649
# We currently assume a block is properly utilized whenever it is >75%
650
# of the size of a 'full' block. In normal operation, a block is
651
# considered full when it hits 4MB of same-file content. So any block
652
# >3MB is 'full enough'.
653
# The only time this isn't true is when a given block has large-object
654
# content. (a single file >4MB, etc.)
655
# Under these circumstances, we allow a block to grow to
656
# 2 x largest_content. Which means that if a given block had a large
657
# object, it may actually be under-utilized. However, given that this
658
# is 'pack-on-the-fly' it is probably reasonable to not repack large
659
# content blobs on-the-fly. Note that because we return False for all
660
# 1-item blobs, we will repack them; we may wish to reevaluate our
661
# treatment of large object blobs in the future.
662
if block_size >= self._full_enough_block_size:
664
# If a block is <3MB, it still may be considered 'full' if it contains
665
# mixed content. The current rule is 2MB of mixed content is considered
666
# full. So check to see if this block contains mixed content, and
667
# set the threshold appropriately.
669
for factory in self._factories:
670
prefix = factory.key[:-1]
671
if common_prefix is None:
672
common_prefix = prefix
673
elif prefix != common_prefix:
674
# Mixed content, check the size appropriately
675
if block_size >= self._full_enough_mixed_block_size:
678
# The content failed both the mixed check and the single-content check
679
# so obviously it is not fully utilized
680
# TODO: there is one other constraint that isn't being checked
681
# namely, that the entries in the block are in the appropriate
682
# order. For example, you could insert the entries in exactly
683
# reverse groupcompress order, and we would think that is ok.
684
# (all the right objects are in one group, and it is fully
685
# utilized, etc.) For now, we assume that case is rare,
686
# especially since we should always fetch in 'groupcompress'
690
def _check_rebuild_block(self):
691
action, last_byte_used, total_bytes_used = self._check_rebuild_action()
695
self._trim_block(last_byte_used)
696
elif action == 'rebuild':
697
self._rebuild_block()
699
raise ValueError('unknown rebuild action: %r' % (action,))
576
self._rebuild_block()
701
578
def _wire_bytes(self):
702
579
"""Return a byte stream suitable for transmitting over the wire."""
736
613
z_header_bytes = zlib.compress(header_bytes)
738
615
z_header_bytes_len = len(z_header_bytes)
739
block_bytes_len, block_chunks = self._block.to_chunks()
616
block_bytes = self._block.to_bytes()
740
617
lines.append('%d\n%d\n%d\n' % (z_header_bytes_len, header_bytes_len,
742
619
lines.append(z_header_bytes)
743
lines.extend(block_chunks)
744
del z_header_bytes, block_chunks
745
# TODO: This is a point where we will double the memory consumption. To
746
# avoid this, we probably have to switch to a 'chunked' api
620
lines.append(block_bytes)
621
del z_header_bytes, block_bytes
747
622
return ''.join(lines)
750
625
def from_bytes(cls, bytes):
751
626
# TODO: This does extra string copying, probably better to do it a
752
# different way. At a minimum this creates 2 copies of the
754
628
(storage_kind, z_header_len, header_len,
755
629
block_len, rest) = bytes.split('\n', 4)
990
868
It contains code very similar to SequenceMatcher because of having a similar
991
869
task. However some key differences apply:
993
* there is no junk, we want a minimal edit not a human readable diff.
994
* we don't filter very common lines (because we don't know where a good
995
range will start, and after the first text we want to be emitting minmal
997
* we chain the left side, not the right side
998
* we incrementally update the adjacency matrix as new lines are provided.
999
* we look for matches in all of the left side, so the routine which does
1000
the analagous task of find_longest_match does not need to filter on the
870
- there is no junk, we want a minimal edit not a human readable diff.
871
- we don't filter very common lines (because we don't know where a good
872
range will start, and after the first text we want to be emitting minmal
874
- we chain the left side, not the right side
875
- we incrementally update the adjacency matrix as new lines are provided.
876
- we look for matches in all of the left side, so the routine which does
877
the analagous task of find_longest_match does not need to filter on the
1004
def __init__(self, settings=None):
1005
super(PyrexGroupCompressor, self).__init__(settings)
1006
max_bytes_to_index = self._settings.get('max_bytes_to_index', 0)
1007
self._delta_index = DeltaIndex(max_bytes_to_index=max_bytes_to_index)
882
super(PyrexGroupCompressor, self).__init__()
883
self._delta_index = DeltaIndex()
1009
885
def _compress(self, key, bytes, max_delta_size, soft=False):
1010
886
"""see _CommonGroupCompressor._compress"""
1099
975
versioned_files.stream.close()
1102
class _BatchingBlockFetcher(object):
1103
"""Fetch group compress blocks in batches.
1105
:ivar total_bytes: int of expected number of bytes needed to fetch the
1106
currently pending batch.
1109
def __init__(self, gcvf, locations, get_compressor_settings=None):
1111
self.locations = locations
1113
self.batch_memos = {}
1114
self.memos_to_get = []
1115
self.total_bytes = 0
1116
self.last_read_memo = None
1118
self._get_compressor_settings = get_compressor_settings
1120
def add_key(self, key):
1121
"""Add another to key to fetch.
1123
:return: The estimated number of bytes needed to fetch the batch so
1126
self.keys.append(key)
1127
index_memo, _, _, _ = self.locations[key]
1128
read_memo = index_memo[0:3]
1129
# Three possibilities for this read_memo:
1130
# - it's already part of this batch; or
1131
# - it's not yet part of this batch, but is already cached; or
1132
# - it's not yet part of this batch and will need to be fetched.
1133
if read_memo in self.batch_memos:
1134
# This read memo is already in this batch.
1135
return self.total_bytes
1137
cached_block = self.gcvf._group_cache[read_memo]
1139
# This read memo is new to this batch, and the data isn't cached
1141
self.batch_memos[read_memo] = None
1142
self.memos_to_get.append(read_memo)
1143
byte_length = read_memo[2]
1144
self.total_bytes += byte_length
1146
# This read memo is new to this batch, but cached.
1147
# Keep a reference to the cached block in batch_memos because it's
1148
# certain that we'll use it when this batch is processed, but
1149
# there's a risk that it would fall out of _group_cache between now
1151
self.batch_memos[read_memo] = cached_block
1152
return self.total_bytes
1154
def _flush_manager(self):
1155
if self.manager is not None:
1156
for factory in self.manager.get_record_stream():
1159
self.last_read_memo = None
1161
def yield_factories(self, full_flush=False):
1162
"""Yield factories for keys added since the last yield. They will be
1163
returned in the order they were added via add_key.
1165
:param full_flush: by default, some results may not be returned in case
1166
they can be part of the next batch. If full_flush is True, then
1167
all results are returned.
1169
if self.manager is None and not self.keys:
1171
# Fetch all memos in this batch.
1172
blocks = self.gcvf._get_blocks(self.memos_to_get)
1173
# Turn blocks into factories and yield them.
1174
memos_to_get_stack = list(self.memos_to_get)
1175
memos_to_get_stack.reverse()
1176
for key in self.keys:
1177
index_memo, _, parents, _ = self.locations[key]
1178
read_memo = index_memo[:3]
1179
if self.last_read_memo != read_memo:
1180
# We are starting a new block. If we have a
1181
# manager, we have found everything that fits for
1182
# now, so yield records
1183
for factory in self._flush_manager():
1185
# Now start a new manager.
1186
if memos_to_get_stack and memos_to_get_stack[-1] == read_memo:
1187
# The next block from _get_blocks will be the block we
1189
block_read_memo, block = blocks.next()
1190
if block_read_memo != read_memo:
1191
raise AssertionError(
1192
"block_read_memo out of sync with read_memo"
1193
"(%r != %r)" % (block_read_memo, read_memo))
1194
self.batch_memos[read_memo] = block
1195
memos_to_get_stack.pop()
1197
block = self.batch_memos[read_memo]
1198
self.manager = _LazyGroupContentManager(block,
1199
get_compressor_settings=self._get_compressor_settings)
1200
self.last_read_memo = read_memo
1201
start, end = index_memo[3:5]
1202
self.manager.add_factory(key, parents, start, end)
1204
for factory in self._flush_manager():
1207
self.batch_memos.clear()
1208
del self.memos_to_get[:]
1209
self.total_bytes = 0
1212
class GroupCompressVersionedFiles(VersionedFilesWithFallbacks):
978
class GroupCompressVersionedFiles(VersionedFiles):
1213
979
"""A group-compress based VersionedFiles implementation."""
1215
# This controls how the GroupCompress DeltaIndex works. Basically, we
1216
# compute hash pointers into the source blocks (so hash(text) => text).
1217
# However each of these references costs some memory in trade against a
1218
# more accurate match result. For very large files, they either are
1219
# pre-compressed and change in bulk whenever they change, or change in just
1220
# local blocks. Either way, 'improved resolution' is not very helpful,
1221
# versus running out of memory trying to track everything. The default max
1222
# gives 100% sampling of a 1MB file.
1223
_DEFAULT_MAX_BYTES_TO_INDEX = 1024 * 1024
1224
_DEFAULT_COMPRESSOR_SETTINGS = {'max_bytes_to_index':
1225
_DEFAULT_MAX_BYTES_TO_INDEX}
1227
def __init__(self, index, access, delta=True, _unadded_refs=None,
981
def __init__(self, index, access, delta=True):
1229
982
"""Create a GroupCompressVersionedFiles object.
1231
984
:param index: The index object storing access and graph data.
1232
985
:param access: The access object storing raw data.
1233
986
:param delta: Whether to delta compress or just entropy compress.
1234
:param _unadded_refs: private parameter, don't use.
1235
:param _group_cache: private parameter, don't use.
1237
988
self._index = index
1238
989
self._access = access
1239
990
self._delta = delta
1240
if _unadded_refs is None:
1242
self._unadded_refs = _unadded_refs
1243
if _group_cache is None:
1244
_group_cache = LRUSizeCache(max_size=50*1024*1024)
1245
self._group_cache = _group_cache
1246
self._immediate_fallback_vfs = []
1247
self._max_bytes_to_index = None
1249
def without_fallbacks(self):
1250
"""Return a clone of this object without any fallbacks configured."""
1251
return GroupCompressVersionedFiles(self._index, self._access,
1252
self._delta, _unadded_refs=dict(self._unadded_refs),
1253
_group_cache=self._group_cache)
991
self._unadded_refs = {}
992
self._group_cache = LRUSizeCache(max_size=50*1024*1024)
993
self._fallback_vfs = []
1255
995
def add_lines(self, key, parents, lines, parent_texts=None,
1256
996
left_matching_blocks=None, nostore_sha=None, random_id=False,
1397
1128
missing.difference_update(set(new_result))
1398
1129
return result, source_results
1400
def _get_blocks(self, read_memos):
1401
"""Get GroupCompressBlocks for the given read_memos.
1403
:returns: a series of (read_memo, block) pairs, in the order they were
1407
for read_memo in read_memos:
1409
block = self._group_cache[read_memo]
1413
cached[read_memo] = block
1415
not_cached_seen = set()
1416
for read_memo in read_memos:
1417
if read_memo in cached:
1418
# Don't fetch what we already have
1420
if read_memo in not_cached_seen:
1421
# Don't try to fetch the same data twice
1423
not_cached.append(read_memo)
1424
not_cached_seen.add(read_memo)
1425
raw_records = self._access.get_raw_records(not_cached)
1426
for read_memo in read_memos:
1428
yield read_memo, cached[read_memo]
1430
# Read the block, and cache it.
1431
zdata = raw_records.next()
1432
block = GroupCompressBlock.from_bytes(zdata)
1433
self._group_cache[read_memo] = block
1434
cached[read_memo] = block
1435
yield read_memo, block
1131
def _get_block(self, index_memo):
1132
read_memo = index_memo[0:3]
1135
block = self._group_cache[read_memo]
1138
zdata = self._access.get_raw_records([read_memo]).next()
1139
# decompress - whole thing - this is not a bug, as it
1140
# permits caching. We might want to store the partially
1141
# decompresed group and decompress object, so that recent
1142
# texts are not penalised by big groups.
1143
block = GroupCompressBlock.from_bytes(zdata)
1144
self._group_cache[read_memo] = block
1146
# print len(zdata), len(plain)
1147
# parse - requires split_lines, better to have byte offsets
1148
# here (but not by much - we only split the region for the
1149
# recipe, and we often want to end up with lines anyway.
1437
1152
def get_missing_compression_parent_keys(self):
1438
1153
"""Return the keys of missing compression parents.
1605
1319
unadded_keys, source_result)
1606
1320
for key in missing:
1607
1321
yield AbsentContentFactory(key)
1608
# Batch up as many keys as we can until either:
1609
# - we encounter an unadded ref, or
1610
# - we run out of keys, or
1611
# - the total bytes to retrieve for this batch > BATCH_SIZE
1612
batcher = _BatchingBlockFetcher(self, locations,
1613
get_compressor_settings=self._get_compressor_settings)
1323
last_read_memo = None
1324
# TODO: This works fairly well at batching up existing groups into a
1325
# streamable format, and possibly allowing for taking one big
1326
# group and splitting it when it isn't fully utilized.
1327
# However, it doesn't allow us to find under-utilized groups and
1328
# combine them into a bigger group on the fly.
1329
# (Consider the issue with how chk_map inserts texts
1330
# one-at-a-time.) This could be done at insert_record_stream()
1331
# time, but it probably would decrease the number of
1332
# bytes-on-the-wire for fetch.
1614
1333
for source, keys in source_keys:
1615
1334
if source is self:
1616
1335
for key in keys:
1617
1336
if key in self._unadded_refs:
1618
# Flush batch, then yield unadded ref from
1620
for factory in batcher.yield_factories(full_flush=True):
1337
if manager is not None:
1338
for factory in manager.get_record_stream():
1340
last_read_memo = manager = None
1622
1341
bytes, sha1 = self._compressor.extract(key)
1623
1342
parents = self._unadded_refs[key]
1624
1343
yield FulltextContentFactory(key, parents, sha1, bytes)
1626
if batcher.add_key(key) > BATCH_SIZE:
1627
# Ok, this batch is big enough. Yield some results.
1628
for factory in batcher.yield_factories():
1345
index_memo, _, parents, (method, _) = locations[key]
1346
read_memo = index_memo[0:3]
1347
if last_read_memo != read_memo:
1348
# We are starting a new block. If we have a
1349
# manager, we have found everything that fits for
1350
# now, so yield records
1351
if manager is not None:
1352
for factory in manager.get_record_stream():
1354
# Now start a new manager
1355
block = self._get_block(index_memo)
1356
manager = _LazyGroupContentManager(block)
1357
last_read_memo = read_memo
1358
start, end = index_memo[3:5]
1359
manager.add_factory(key, parents, start, end)
1631
for factory in batcher.yield_factories(full_flush=True):
1361
if manager is not None:
1362
for factory in manager.get_record_stream():
1364
last_read_memo = manager = None
1633
1365
for record in source.get_record_stream(keys, ordering,
1634
1366
include_delta_closure):
1636
for factory in batcher.yield_factories(full_flush=True):
1368
if manager is not None:
1369
for factory in manager.get_record_stream():
1639
1372
def get_sha1s(self, keys):
1640
1373
"""See VersionedFiles.get_sha1s()."""
1715
1424
# This will go up to fulltexts for gc to gc fetching, which isn't
1717
self._compressor = self._make_group_compressor()
1426
self._compressor = GroupCompressor()
1718
1427
self._unadded_refs = {}
1719
1428
keys_to_add = []
1721
bytes_len, chunks = self._compressor.flush().to_chunks()
1722
self._compressor = self._make_group_compressor()
1723
# Note: At this point we still have 1 copy of the fulltext (in
1724
# record and the var 'bytes'), and this generates 2 copies of
1725
# the compressed text (one for bytes, one in chunks)
1726
# TODO: Push 'chunks' down into the _access api, so that we don't
1727
# have to double compressed memory here
1728
# TODO: Figure out how to indicate that we would be happy to free
1729
# the fulltext content at this point. Note that sometimes we
1730
# will want it later (streaming CHK pages), but most of the
1731
# time we won't (everything else)
1732
bytes = ''.join(chunks)
1430
bytes = self._compressor.flush().to_bytes()
1734
1431
index, start, length = self._access.add_raw_records(
1735
1432
[(None, len(bytes))], bytes)[0]
1748
1446
block_length = None
1749
1447
# XXX: TODO: remove this, it is just for safety checking for now
1750
1448
inserted_keys = set()
1751
reuse_this_block = reuse_blocks
1752
1449
for record in stream:
1753
1450
# Raise an error when a record is missing.
1754
1451
if record.storage_kind == 'absent':
1755
1452
raise errors.RevisionNotPresent(record.key, self)
1757
1454
if record.key in inserted_keys:
1758
trace.note(gettext('Insert claimed random_id=True,'
1759
' but then inserted %r two times'), record.key)
1455
trace.note('Insert claimed random_id=True,'
1456
' but then inserted %r two times', record.key)
1761
1458
inserted_keys.add(record.key)
1762
1459
if reuse_blocks:
1763
1460
# If the reuse_blocks flag is set, check to see if we can just
1764
1461
# copy a groupcompress block as-is.
1765
# We only check on the first record (groupcompress-block) not
1766
# on all of the (groupcompress-block-ref) entries.
1767
# The reuse_this_block flag is then kept for as long as
1768
if record.storage_kind == 'groupcompress-block':
1769
# Check to see if we really want to re-use this block
1770
insert_manager = record._manager
1771
reuse_this_block = insert_manager.check_is_well_utilized()
1773
reuse_this_block = False
1774
if reuse_this_block:
1775
# We still want to reuse this block
1776
1462
if record.storage_kind == 'groupcompress-block':
1777
1463
# Insert the raw block into the target repo
1778
1464
insert_manager = record._manager
1465
insert_manager._check_rebuild_block()
1779
1466
bytes = record._manager._block.to_bytes()
1780
1467
_, start, length = self._access.add_raw_records(
1781
1468
[(None, len(bytes))], bytes)[0]
1901
1578
"""See VersionedFiles.keys."""
1902
1579
if 'evil' in debug.debug_flags:
1903
1580
trace.mutter_callsite(2, "keys scales with size of history")
1904
sources = [self._index] + self._immediate_fallback_vfs
1581
sources = [self._index] + self._fallback_vfs
1906
1583
for source in sources:
1907
1584
result.update(source.keys())
1911
class _GCBuildDetails(object):
1912
"""A blob of data about the build details.
1914
This stores the minimal data, which then allows compatibility with the old
1915
api, without taking as much memory.
1918
__slots__ = ('_index', '_group_start', '_group_end', '_basis_end',
1919
'_delta_end', '_parents')
1922
compression_parent = None
1924
def __init__(self, parents, position_info):
1925
self._parents = parents
1926
(self._index, self._group_start, self._group_end, self._basis_end,
1927
self._delta_end) = position_info
1930
return '%s(%s, %s)' % (self.__class__.__name__,
1931
self.index_memo, self._parents)
1934
def index_memo(self):
1935
return (self._index, self._group_start, self._group_end,
1936
self._basis_end, self._delta_end)
1939
def record_details(self):
1940
return static_tuple.StaticTuple(self.method, None)
1942
def __getitem__(self, offset):
1943
"""Compatibility thunk to act like a tuple."""
1945
return self.index_memo
1947
return self.compression_parent # Always None
1949
return self._parents
1951
return self.record_details
1953
raise IndexError('offset out of range')
1959
1588
class _GCGraphIndex(object):
1960
1589
"""Mapper from GroupCompressVersionedFiles needs into GraphIndex storage."""
1962
1591
def __init__(self, graph_index, is_locked, parents=True,
1963
1592
add_callback=None, track_external_parent_refs=False,
1964
inconsistency_fatal=True, track_new_keys=False):
1593
inconsistency_fatal=True):
1965
1594
"""Construct a _GCGraphIndex on a graph_index.
1967
1596
:param graph_index: An implementation of bzrlib.index.GraphIndex.
2170
1787
"""Convert an index value to position details."""
2171
1788
bits = node[2].split(' ')
2172
1789
# It would be nice not to read the entire gzip.
2173
# start and stop are put into _int_cache because they are very common.
2174
# They define the 'group' that an entry is in, and many groups can have
2175
# thousands of objects.
2176
# Branching Launchpad, for example, saves ~600k integers, at 12 bytes
2177
# each, or about 7MB. Note that it might be even more when you consider
2178
# how PyInt is allocated in separate slabs. And you can't return a slab
2179
# to the OS if even 1 int on it is in use. Note though that Python uses
2180
# a LIFO when re-using PyInt slots, which might cause more
2182
1790
start = int(bits[0])
2183
start = self._int_cache.setdefault(start, start)
2184
1791
stop = int(bits[1])
2185
stop = self._int_cache.setdefault(stop, stop)
2186
1792
basis_end = int(bits[2])
2187
1793
delta_end = int(bits[3])
2188
# We can't use StaticTuple here, because node[0] is a BTreeGraphIndex
2190
return (node[0], start, stop, basis_end, delta_end)
1794
return node[0], start, stop, basis_end, delta_end
2192
1796
def scan_unvalidated_index(self, graph_index):
2193
1797
"""Inform this _GCGraphIndex that there is an unvalidated index.
2195
1799
This allows this _GCGraphIndex to keep track of any missing
2196
1800
compression parents we may want to have filled in to make those
2197
indices valid. It also allows _GCGraphIndex to track any new keys.
2199
1803
:param graph_index: A GraphIndex
2201
key_dependencies = self._key_dependencies
2202
if key_dependencies is None:
2204
for node in graph_index.iter_all_entries():
2205
# Add parent refs from graph_index (and discard parent refs
2206
# that the graph_index has).
2207
key_dependencies.add_references(node[1], node[3][0])
1805
if self._key_dependencies is not None:
1806
# Add parent refs from graph_index (and discard parent refs that
1807
# the graph_index has).
1808
add_refs = self._key_dependencies.add_references
1809
for node in graph_index.iter_all_entries():
1810
add_refs(node[1], node[3][0])
2210
1814
from bzrlib._groupcompress_py import (