1
# Copyright (C) 2008-2011 Canonical Ltd
3
# This program is free software; you can redistribute it and/or modify
4
# it under the terms of the GNU General Public License as published by
5
# the Free Software Foundation; either version 2 of the License, or
6
# (at your option) any later version.
8
# This program is distributed in the hope that it will be useful,
9
# but WITHOUT ANY WARRANTY; without even the implied warranty of
10
# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
11
# GNU General Public License for more details.
13
# You should have received a copy of the GNU General Public License
14
# along with this program; if not, write to the Free Software
15
# Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20
from __future__ import absolute_import
22
from .lazy_import import lazy_import
23
lazy_import(globals(), """
42
from .index import _OPTION_NODE_REFS, _OPTION_KEY_ELEMENTS, _OPTION_LEN
53
_BTSIGNATURE = b"B+Tree Graph Index 2\n"
54
_OPTION_ROW_LENGTHS = b"row_lengths="
55
_LEAF_FLAG = b"type=leaf\n"
56
_INTERNAL_FLAG = b"type=internal\n"
57
_INTERNAL_OFFSET = b"offset="
59
_RESERVED_HEADER_BYTES = 120
62
# 4K per page: 4MB - 1000 entries
63
_NODE_CACHE_SIZE = 1000
66
class _BuilderRow(object):
67
"""The stored state accumulated while writing out a row in the index.
69
:ivar spool: A temporary file used to accumulate nodes for this row
71
:ivar nodes: The count of nodes emitted so far.
75
"""Create a _BuilderRow."""
77
self.spool = None# tempfile.TemporaryFile(prefix='bzr-index-row-')
80
def finish_node(self, pad=True):
81
byte_lines, _, padding = self.writer.finish()
83
self.spool = BytesIO()
85
self.spool.write("\x00" * _RESERVED_HEADER_BYTES)
87
# We got bigger than 1 node, switch to a temp file
88
spool = tempfile.TemporaryFile(prefix='bzr-index-row-')
89
spool.write(self.spool.getvalue())
92
if not pad and padding:
94
skipped_bytes = padding
95
self.spool.writelines(byte_lines)
96
remainder = (self.spool.tell() + skipped_bytes) % _PAGE_SIZE
98
raise AssertionError("incorrect node length: %d, %d"
99
% (self.spool.tell(), remainder))
104
class _InternalBuilderRow(_BuilderRow):
105
"""The stored state accumulated while writing out internal rows."""
107
def finish_node(self, pad=True):
109
raise AssertionError("Must pad internal nodes only.")
110
_BuilderRow.finish_node(self)
113
class _LeafBuilderRow(_BuilderRow):
114
"""The stored state accumulated while writing out a leaf rows."""
117
class BTreeBuilder(index.GraphIndexBuilder):
118
"""A Builder for B+Tree based Graph indices.
120
The resulting graph has the structure:
122
_SIGNATURE OPTIONS NODES
123
_SIGNATURE := 'B+Tree Graph Index 1' NEWLINE
124
OPTIONS := REF_LISTS KEY_ELEMENTS LENGTH
125
REF_LISTS := 'node_ref_lists=' DIGITS NEWLINE
126
KEY_ELEMENTS := 'key_elements=' DIGITS NEWLINE
127
LENGTH := 'len=' DIGITS NEWLINE
128
ROW_LENGTHS := 'row_lengths' DIGITS (COMMA DIGITS)*
129
NODES := NODE_COMPRESSED*
130
NODE_COMPRESSED:= COMPRESSED_BYTES{4096}
131
NODE_RAW := INTERNAL | LEAF
132
INTERNAL := INTERNAL_FLAG POINTERS
133
LEAF := LEAF_FLAG ROWS
134
KEY_ELEMENT := Not-whitespace-utf8
135
KEY := KEY_ELEMENT (NULL KEY_ELEMENT)*
137
ROW := KEY NULL ABSENT? NULL REFERENCES NULL VALUE NEWLINE
139
REFERENCES := REFERENCE_LIST (TAB REFERENCE_LIST){node_ref_lists - 1}
140
REFERENCE_LIST := (REFERENCE (CR REFERENCE)*)?
142
VALUE := no-newline-no-null-bytes
145
def __init__(self, reference_lists=0, key_elements=1, spill_at=100000):
146
"""See GraphIndexBuilder.__init__.
148
:param spill_at: Optional parameter controlling the maximum number
149
of nodes that BTreeBuilder will hold in memory.
151
index.GraphIndexBuilder.__init__(self, reference_lists=reference_lists,
152
key_elements=key_elements)
153
self._spill_at = spill_at
154
self._backing_indices = []
155
# A map of {key: (node_refs, value)}
157
# Indicate it hasn't been built yet
158
self._nodes_by_key = None
159
self._optimize_for_size = False
161
def add_node(self, key, value, references=()):
162
"""Add a node to the index.
164
If adding the node causes the builder to reach its spill_at threshold,
165
disk spilling will be triggered.
167
:param key: The key. keys are non-empty tuples containing
168
as many whitespace-free utf8 bytestrings as the key length
169
defined for this index.
170
:param references: An iterable of iterables of keys. Each is a
171
reference to another key.
172
:param value: The value to associate with the key. It may be any
173
bytes as long as it does not contain \\0 or \\n.
175
# Ensure that 'key' is a StaticTuple
176
key = static_tuple.StaticTuple.from_sequence(key).intern()
177
# we don't care about absent_references
178
node_refs, _ = self._check_key_ref_value(key, references, value)
179
if key in self._nodes:
180
raise errors.BadIndexDuplicateKey(key, self)
181
self._nodes[key] = static_tuple.StaticTuple(node_refs, value)
182
if self._nodes_by_key is not None and self._key_length > 1:
183
self._update_nodes_by_key(key, value, node_refs)
184
if len(self._nodes) < self._spill_at:
186
self._spill_mem_keys_to_disk()
188
def _spill_mem_keys_to_disk(self):
189
"""Write the in memory keys down to disk to cap memory consumption.
191
If we already have some keys written to disk, we will combine them so
192
as to preserve the sorted order. The algorithm for combining uses
193
powers of two. So on the first spill, write all mem nodes into a
194
single index. On the second spill, combine the mem nodes with the nodes
195
on disk to create a 2x sized disk index and get rid of the first index.
196
On the third spill, create a single new disk index, which will contain
197
the mem nodes, and preserve the existing 2x sized index. On the fourth,
198
combine mem with the first and second indexes, creating a new one of
199
size 4x. On the fifth create a single new one, etc.
201
if self._combine_backing_indices:
202
(new_backing_file, size,
203
backing_pos) = self._spill_mem_keys_and_combine()
205
new_backing_file, size = self._spill_mem_keys_without_combining()
206
# Note: The transport here isn't strictly needed, because we will use
207
# direct access to the new_backing._file object
208
new_backing = BTreeGraphIndex(transport.get_transport_from_path('.'),
210
# GC will clean up the file
211
new_backing._file = new_backing_file
212
if self._combine_backing_indices:
213
if len(self._backing_indices) == backing_pos:
214
self._backing_indices.append(None)
215
self._backing_indices[backing_pos] = new_backing
216
for backing_pos in range(backing_pos):
217
self._backing_indices[backing_pos] = None
219
self._backing_indices.append(new_backing)
221
self._nodes_by_key = None
223
def _spill_mem_keys_without_combining(self):
224
return self._write_nodes(self._iter_mem_nodes(), allow_optimize=False)
226
def _spill_mem_keys_and_combine(self):
227
iterators_to_combine = [self._iter_mem_nodes()]
229
for pos, backing in enumerate(self._backing_indices):
233
iterators_to_combine.append(backing.iter_all_entries())
234
backing_pos = pos + 1
235
new_backing_file, size = \
236
self._write_nodes(self._iter_smallest(iterators_to_combine),
237
allow_optimize=False)
238
return new_backing_file, size, backing_pos
240
def add_nodes(self, nodes):
241
"""Add nodes to the index.
243
:param nodes: An iterable of (key, node_refs, value) entries to add.
245
if self.reference_lists:
246
for (key, value, node_refs) in nodes:
247
self.add_node(key, value, node_refs)
249
for (key, value) in nodes:
250
self.add_node(key, value)
252
def _iter_mem_nodes(self):
253
"""Iterate over the nodes held in memory."""
255
if self.reference_lists:
256
for key in sorted(nodes):
257
references, value = nodes[key]
258
yield self, key, value, references
260
for key in sorted(nodes):
261
references, value = nodes[key]
262
yield self, key, value
264
def _iter_smallest(self, iterators_to_combine):
265
if len(iterators_to_combine) == 1:
266
for value in iterators_to_combine[0]:
270
for iterator in iterators_to_combine:
272
current_values.append(next(iterator))
273
except StopIteration:
274
current_values.append(None)
277
# Decorate candidates with the value to allow 2.4's min to be used.
278
candidates = [(item[1][1], item) for item
279
in enumerate(current_values) if item[1] is not None]
280
if not len(candidates):
282
selected = min(candidates)
283
# undecorate back to (pos, node)
284
selected = selected[1]
285
if last == selected[1][1]:
286
raise errors.BadIndexDuplicateKey(last, self)
287
last = selected[1][1]
288
# Yield, with self as the index
289
yield (self,) + selected[1][1:]
292
current_values[pos] = next(iterators_to_combine[pos])
293
except StopIteration:
294
current_values[pos] = None
296
def _add_key(self, string_key, line, rows, allow_optimize=True):
297
"""Add a key to the current chunk.
299
:param string_key: The key to add.
300
:param line: The fully serialised key and value.
301
:param allow_optimize: If set to False, prevent setting the optimize
302
flag when writing out. This is used by the _spill_mem_keys_to_disk
306
if rows[-1].writer is None:
307
# opening a new leaf chunk;
309
for pos, internal_row in enumerate(rows[:-1]):
310
# flesh out any internal nodes that are needed to
311
# preserve the height of the tree
312
if internal_row.writer is None:
314
if internal_row.nodes == 0:
315
length -= _RESERVED_HEADER_BYTES # padded
317
optimize_for_size = self._optimize_for_size
319
optimize_for_size = False
320
internal_row.writer = chunk_writer.ChunkWriter(length, 0,
321
optimize_for_size=optimize_for_size)
322
internal_row.writer.write(_INTERNAL_FLAG)
323
internal_row.writer.write(_INTERNAL_OFFSET +
324
str(rows[pos + 1].nodes) + "\n")
327
if rows[-1].nodes == 0:
328
length -= _RESERVED_HEADER_BYTES # padded
329
rows[-1].writer = chunk_writer.ChunkWriter(length,
330
optimize_for_size=self._optimize_for_size)
331
rows[-1].writer.write(_LEAF_FLAG)
332
if rows[-1].writer.write(line):
333
# if we failed to write, despite having an empty page to write to,
334
# then line is too big. raising the error avoids infinite recursion
335
# searching for a suitably large page that will not be found.
337
raise errors.BadIndexKey(string_key)
338
# this key did not fit in the node:
339
rows[-1].finish_node()
340
key_line = string_key + "\n"
342
for row in reversed(rows[:-1]):
343
# Mark the start of the next node in the node above. If it
344
# doesn't fit then propagate upwards until we find one that
346
if row.writer.write(key_line):
349
# We've found a node that can handle the pointer.
352
# If we reached the current root without being able to mark the
353
# division point, then we need a new root:
356
if 'index' in debug.debug_flags:
357
trace.mutter('Inserting new global row.')
358
new_row = _InternalBuilderRow()
360
rows.insert(0, new_row)
361
# This will be padded, hence the -100
362
new_row.writer = chunk_writer.ChunkWriter(
363
_PAGE_SIZE - _RESERVED_HEADER_BYTES,
365
optimize_for_size=self._optimize_for_size)
366
new_row.writer.write(_INTERNAL_FLAG)
367
new_row.writer.write(_INTERNAL_OFFSET +
368
str(rows[1].nodes - 1) + "\n")
369
new_row.writer.write(key_line)
370
self._add_key(string_key, line, rows, allow_optimize=allow_optimize)
372
def _write_nodes(self, node_iterator, allow_optimize=True):
373
"""Write node_iterator out as a B+Tree.
375
:param node_iterator: An iterator of sorted nodes. Each node should
376
match the output given by iter_all_entries.
377
:param allow_optimize: If set to False, prevent setting the optimize
378
flag when writing out. This is used by the _spill_mem_keys_to_disk
380
:return: A file handle for a temporary file containing a B+Tree for
383
# The index rows - rows[0] is the root, rows[1] is the layer under it
386
# forward sorted by key. In future we may consider topological sorting,
387
# at the cost of table scans for direct lookup, or a second index for
390
# A stack with the number of nodes of each size. 0 is the root node
391
# and must always be 1 (if there are any nodes in the tree).
392
self.row_lengths = []
393
# Loop over all nodes adding them to the bottom row
394
# (rows[-1]). When we finish a chunk in a row,
395
# propagate the key that didn't fit (comes after the chunk) to the
396
# row above, transitively.
397
for node in node_iterator:
399
# First key triggers the first row
400
rows.append(_LeafBuilderRow())
402
string_key, line = _btree_serializer._flatten_node(node,
403
self.reference_lists)
404
self._add_key(string_key, line, rows, allow_optimize=allow_optimize)
405
for row in reversed(rows):
406
pad = (not isinstance(row, _LeafBuilderRow))
407
row.finish_node(pad=pad)
408
lines = [_BTSIGNATURE]
409
lines.append(b'%s%d\n' % (_OPTION_NODE_REFS, self.reference_lists))
410
lines.append(b'%s%d\n' % (_OPTION_KEY_ELEMENTS, self._key_length))
411
lines.append(b'%s%d\n' % (_OPTION_LEN, key_count))
412
row_lengths = [row.nodes for row in rows]
413
lines.append(_OPTION_ROW_LENGTHS + b','.join(
414
map(bytes, row_lengths)) + b'\n')
415
if row_lengths and row_lengths[-1] > 1:
416
result = tempfile.NamedTemporaryFile(prefix='bzr-index-')
419
result.writelines(lines)
420
position = sum(map(len, lines))
422
if position > _RESERVED_HEADER_BYTES:
423
raise AssertionError("Could not fit the header in the"
424
" reserved space: %d > %d"
425
% (position, _RESERVED_HEADER_BYTES))
426
# write the rows out:
428
reserved = _RESERVED_HEADER_BYTES # reserved space for first node
431
# copy nodes to the finalised file.
432
# Special case the first node as it may be prefixed
433
node = row.spool.read(_PAGE_SIZE)
434
result.write(node[reserved:])
435
if len(node) == _PAGE_SIZE:
436
result.write("\x00" * (reserved - position))
437
position = 0 # Only the root row actually has an offset
438
copied_len = osutils.pumpfile(row.spool, result)
439
if copied_len != (row.nodes - 1) * _PAGE_SIZE:
440
if not isinstance(row, _LeafBuilderRow):
441
raise AssertionError("Incorrect amount of data copied"
442
" expected: %d, got: %d"
443
% ((row.nodes - 1) * _PAGE_SIZE,
451
"""Finalise the index.
453
:return: A file handle for a temporary file containing the nodes added
456
return self._write_nodes(self.iter_all_entries())[0]
458
def iter_all_entries(self):
459
"""Iterate over all keys within the index
461
:return: An iterable of (index, key, value, reference_lists). There is
462
no defined order for the result iteration - it will be in the most
463
efficient order for the index (in this case dictionary hash order).
465
if 'evil' in debug.debug_flags:
466
trace.mutter_callsite(3,
467
"iter_all_entries scales with size of history.")
468
# Doing serial rather than ordered would be faster; but this shouldn't
469
# be getting called routinely anyway.
470
iterators = [self._iter_mem_nodes()]
471
for backing in self._backing_indices:
472
if backing is not None:
473
iterators.append(backing.iter_all_entries())
474
if len(iterators) == 1:
476
return self._iter_smallest(iterators)
478
def iter_entries(self, keys):
479
"""Iterate over keys within the index.
481
:param keys: An iterable providing the keys to be retrieved.
482
:return: An iterable of (index, key, value, reference_lists). There is no
483
defined order for the result iteration - it will be in the most
484
efficient order for the index (keys iteration order in this case).
487
# Note: We don't use keys.intersection() here. If you read the C api,
488
# set.intersection(other) special cases when other is a set and
489
# will iterate the smaller of the two and lookup in the other.
490
# It does *not* do this for any other type (even dict, unlike
491
# some other set functions.) Since we expect keys is generally <<
492
# self._nodes, it is faster to iterate over it in a list
495
local_keys = [key for key in keys if key in nodes]
496
if self.reference_lists:
497
for key in local_keys:
499
yield self, key, node[1], node[0]
501
for key in local_keys:
503
yield self, key, node[1]
504
# Find things that are in backing indices that have not been handled
506
if not self._backing_indices:
507
return # We won't find anything there either
508
# Remove all of the keys that we found locally
509
keys.difference_update(local_keys)
510
for backing in self._backing_indices:
515
for node in backing.iter_entries(keys):
517
yield (self,) + node[1:]
519
def iter_entries_prefix(self, keys):
520
"""Iterate over keys within the index using prefix matching.
522
Prefix matching is applied within the tuple of a key, not to within
523
the bytestring of each key element. e.g. if you have the keys ('foo',
524
'bar'), ('foobar', 'gam') and do a prefix search for ('foo', None) then
525
only the former key is returned.
527
:param keys: An iterable providing the key prefixes to be retrieved.
528
Each key prefix takes the form of a tuple the length of a key, but
529
with the last N elements 'None' rather than a regular bytestring.
530
The first element cannot be 'None'.
531
:return: An iterable as per iter_all_entries, but restricted to the
532
keys with a matching prefix to those supplied. No additional keys
533
will be returned, and every match that is in the index will be
539
for backing in self._backing_indices:
542
for node in backing.iter_entries_prefix(keys):
543
yield (self,) + node[1:]
544
if self._key_length == 1:
546
index._sanity_check_key(self, key)
548
node = self._nodes[key]
551
if self.reference_lists:
552
yield self, key, node[1], node[0]
554
yield self, key, node[1]
556
nodes_by_key = self._get_nodes_by_key()
557
for entry in index._iter_entries_prefix(self, nodes_by_key, keys):
560
def _get_nodes_by_key(self):
561
if self._nodes_by_key is None:
563
if self.reference_lists:
564
for key, (references, value) in viewitems(self._nodes):
565
key_dict = nodes_by_key
566
for subkey in key[:-1]:
567
key_dict = key_dict.setdefault(subkey, {})
568
key_dict[key[-1]] = key, value, references
570
for key, (references, value) in viewitems(self._nodes):
571
key_dict = nodes_by_key
572
for subkey in key[:-1]:
573
key_dict = key_dict.setdefault(subkey, {})
574
key_dict[key[-1]] = key, value
575
self._nodes_by_key = nodes_by_key
576
return self._nodes_by_key
579
"""Return an estimate of the number of keys in this index.
581
For InMemoryGraphIndex the estimate is exact.
583
return len(self._nodes) + sum(backing.key_count() for backing in
584
self._backing_indices if backing is not None)
587
"""In memory index's have no known corruption at the moment."""
590
class _LeafNode(dict):
591
"""A leaf node for a serialised B+Tree index."""
593
__slots__ = ('min_key', 'max_key', '_keys')
595
def __init__(self, bytes, key_length, ref_list_length):
596
"""Parse bytes to create a leaf node object."""
597
# splitlines mangles the \r delimiters.. don't use it.
598
key_list = _btree_serializer._parse_leaf_lines(bytes,
599
key_length, ref_list_length)
601
self.min_key = key_list[0][0]
602
self.max_key = key_list[-1][0]
604
self.min_key = self.max_key = None
605
super(_LeafNode, self).__init__(key_list)
606
self._keys = dict(self)
609
"""Return a sorted list of (key, (value, refs)) items"""
610
items = sorted(self.items())
614
"""Return a sorted list of all keys."""
615
keys = sorted(self.keys())
619
class _InternalNode(object):
620
"""An internal node for a serialised B+Tree index."""
622
__slots__ = ('keys', 'offset')
624
def __init__(self, bytes):
625
"""Parse bytes to create an internal node object."""
626
# splitlines mangles the \r delimiters.. don't use it.
627
self.keys = self._parse_lines(bytes.split('\n'))
629
def _parse_lines(self, lines):
631
self.offset = int(lines[1][7:])
632
as_st = static_tuple.StaticTuple.from_sequence
633
for line in lines[2:]:
636
# GZ 2017-05-24: Used to intern() each chunk of line as well, need
637
# to recheck performance and perhaps adapt StaticTuple to adjust.
638
nodes.append(as_st(line.split(b'\0')).intern())
642
class BTreeGraphIndex(object):
643
"""Access to nodes via the standard GraphIndex interface for B+Tree's.
645
Individual nodes are held in a LRU cache. This holds the root node in
646
memory except when very large walks are done.
649
def __init__(self, transport, name, size, unlimited_cache=False,
651
"""Create a B+Tree index object on the index name.
653
:param transport: The transport to read data for the index from.
654
:param name: The file name of the index on transport.
655
:param size: Optional size of the index in bytes. This allows
656
compatibility with the GraphIndex API, as well as ensuring that
657
the initial read (to read the root node header) can be done
658
without over-reading even on empty indices, and on small indices
659
allows single-IO to read the entire index.
660
:param unlimited_cache: If set to True, then instead of using an
661
LRUCache with size _NODE_CACHE_SIZE, we will use a dict and always
662
cache all leaf nodes.
663
:param offset: The start of the btree index data isn't byte 0 of the
664
file. Instead it starts at some point later.
666
self._transport = transport
670
self._recommended_pages = self._compute_recommended_pages()
671
self._root_node = None
672
self._base_offset = offset
673
self._leaf_factory = _LeafNode
674
# Default max size is 100,000 leave values
675
self._leaf_value_cache = None # lru_cache.LRUCache(100*1000)
677
self._leaf_node_cache = {}
678
self._internal_node_cache = {}
680
self._leaf_node_cache = lru_cache.LRUCache(_NODE_CACHE_SIZE)
681
# We use a FIFO here just to prevent possible blowout. However, a
682
# 300k record btree has only 3k leaf nodes, and only 20 internal
683
# nodes. A value of 100 scales to ~100*100*100 = 1M records.
684
self._internal_node_cache = fifo_cache.FIFOCache(100)
685
self._key_count = None
686
self._row_lengths = None
687
self._row_offsets = None # Start of each row, [-1] is the end
689
def __eq__(self, other):
690
"""Equal when self and other were created with the same parameters."""
692
isinstance(self, type(other)) and
693
self._transport == other._transport and
694
self._name == other._name and
695
self._size == other._size)
697
def __ne__(self, other):
698
return not self.__eq__(other)
700
def _get_and_cache_nodes(self, nodes):
701
"""Read nodes and cache them in the lru.
703
The nodes list supplied is sorted and then read from disk, each node
704
being inserted it into the _node_cache.
706
Note: Asking for more nodes than the _node_cache can contain will
707
result in some of the results being immediately discarded, to prevent
708
this an assertion is raised if more nodes are asked for than are
711
:return: A dict of {node_pos: node}
714
start_of_leaves = None
715
for node_pos, node in self._read_nodes(sorted(nodes)):
716
if node_pos == 0: # Special case
717
self._root_node = node
719
if start_of_leaves is None:
720
start_of_leaves = self._row_offsets[-2]
721
if node_pos < start_of_leaves:
722
self._internal_node_cache[node_pos] = node
724
self._leaf_node_cache[node_pos] = node
725
found[node_pos] = node
728
def _compute_recommended_pages(self):
729
"""Convert transport's recommended_page_size into btree pages.
731
recommended_page_size is in bytes, we want to know how many _PAGE_SIZE
732
pages fit in that length.
734
recommended_read = self._transport.recommended_page_size()
735
recommended_pages = int(math.ceil(recommended_read /
737
return recommended_pages
739
def _compute_total_pages_in_index(self):
740
"""How many pages are in the index.
742
If we have read the header we will use the value stored there.
743
Otherwise it will be computed based on the length of the index.
745
if self._size is None:
746
raise AssertionError('_compute_total_pages_in_index should not be'
747
' called when self._size is None')
748
if self._root_node is not None:
749
# This is the number of pages as defined by the header
750
return self._row_offsets[-1]
751
# This is the number of pages as defined by the size of the index. They
752
# should be indentical.
753
total_pages = int(math.ceil(self._size / float(_PAGE_SIZE)))
756
def _expand_offsets(self, offsets):
757
"""Find extra pages to download.
759
The idea is that we always want to make big-enough requests (like 64kB
760
for http), so that we don't waste round trips. So given the entries
761
that we already have cached and the new pages being downloaded figure
762
out what other pages we might want to read.
764
See also doc/developers/btree_index_prefetch.txt for more details.
766
:param offsets: The offsets to be read
767
:return: A list of offsets to download
769
if 'index' in debug.debug_flags:
770
trace.mutter('expanding: %s\toffsets: %s', self._name, offsets)
772
if len(offsets) >= self._recommended_pages:
773
# Don't add more, we are already requesting more than enough
774
if 'index' in debug.debug_flags:
775
trace.mutter(' not expanding large request (%s >= %s)',
776
len(offsets), self._recommended_pages)
778
if self._size is None:
779
# Don't try anything, because we don't know where the file ends
780
if 'index' in debug.debug_flags:
781
trace.mutter(' not expanding without knowing index size')
783
total_pages = self._compute_total_pages_in_index()
784
cached_offsets = self._get_offsets_to_cached_pages()
785
# If reading recommended_pages would read the rest of the index, just
787
if total_pages - len(cached_offsets) <= self._recommended_pages:
788
# Read whatever is left
790
expanded = [x for x in range(total_pages)
791
if x not in cached_offsets]
793
expanded = list(range(total_pages))
794
if 'index' in debug.debug_flags:
795
trace.mutter(' reading all unread pages: %s', expanded)
798
if self._root_node is None:
799
# ATM on the first read of the root node of a large index, we don't
800
# bother pre-reading any other pages. This is because the
801
# likelyhood of actually reading interesting pages is very low.
802
# See doc/developers/btree_index_prefetch.txt for a discussion, and
803
# a possible implementation when we are guessing that the second
804
# layer index is small
805
final_offsets = offsets
807
tree_depth = len(self._row_lengths)
808
if len(cached_offsets) < tree_depth and len(offsets) == 1:
809
# We haven't read enough to justify expansion
810
# If we are only going to read the root node, and 1 leaf node,
811
# then it isn't worth expanding our request. Once we've read at
812
# least 2 nodes, then we are probably doing a search, and we
813
# start expanding our requests.
814
if 'index' in debug.debug_flags:
815
trace.mutter(' not expanding on first reads')
817
final_offsets = self._expand_to_neighbors(offsets, cached_offsets,
820
final_offsets = sorted(final_offsets)
821
if 'index' in debug.debug_flags:
822
trace.mutter('expanded: %s', final_offsets)
825
def _expand_to_neighbors(self, offsets, cached_offsets, total_pages):
826
"""Expand requests to neighbors until we have enough pages.
828
This is called from _expand_offsets after policy has determined that we
830
We only want to expand requests within a given layer. We cheat a little
831
bit and assume all requests will be in the same layer. This is true
832
given the current design, but if it changes this algorithm may perform
835
:param offsets: requested offsets
836
:param cached_offsets: offsets for pages we currently have cached
837
:return: A set() of offsets after expansion
839
final_offsets = set(offsets)
841
new_tips = set(final_offsets)
842
while len(final_offsets) < self._recommended_pages and new_tips:
846
first, end = self._find_layer_first_and_end(pos)
849
and previous not in cached_offsets
850
and previous not in final_offsets
851
and previous >= first):
852
next_tips.add(previous)
854
if (after < total_pages
855
and after not in cached_offsets
856
and after not in final_offsets
859
# This would keep us from going bigger than
860
# recommended_pages by only expanding the first offsets.
861
# However, if we are making a 'wide' request, it is
862
# reasonable to expand all points equally.
863
# if len(final_offsets) > recommended_pages:
865
final_offsets.update(next_tips)
869
def clear_cache(self):
870
"""Clear out any cached/memoized values.
872
This can be called at any time, but generally it is used when we have
873
extracted some information, but don't expect to be requesting any more
876
# Note that we don't touch self._root_node or self._internal_node_cache
877
# We don't expect either of those to be big, and it can save
878
# round-trips in the future. We may re-evaluate this if InternalNode
879
# memory starts to be an issue.
880
self._leaf_node_cache.clear()
882
def external_references(self, ref_list_num):
883
if self._root_node is None:
884
self._get_root_node()
885
if ref_list_num + 1 > self.node_ref_lists:
886
raise ValueError('No ref list %d, index has %d ref lists'
887
% (ref_list_num, self.node_ref_lists))
890
for node in self.iter_all_entries():
892
refs.update(node[3][ref_list_num])
895
def _find_layer_first_and_end(self, offset):
896
"""Find the start/stop nodes for the layer corresponding to offset.
898
:return: (first, end)
899
first is the first node in this layer
900
end is the first node of the next layer
903
for roffset in self._row_offsets:
910
def _get_offsets_to_cached_pages(self):
911
"""Determine what nodes we already have cached."""
912
cached_offsets = set(self._internal_node_cache)
913
# cache may be dict or LRUCache, keys() is the common method
914
cached_offsets.update(self._leaf_node_cache.keys())
915
if self._root_node is not None:
916
cached_offsets.add(0)
917
return cached_offsets
919
def _get_root_node(self):
920
if self._root_node is None:
921
# We may not have a root node yet
922
self._get_internal_nodes([0])
923
return self._root_node
925
def _get_nodes(self, cache, node_indexes):
928
for idx in node_indexes:
929
if idx == 0 and self._root_node is not None:
930
found[0] = self._root_node
933
found[idx] = cache[idx]
938
needed = self._expand_offsets(needed)
939
found.update(self._get_and_cache_nodes(needed))
942
def _get_internal_nodes(self, node_indexes):
943
"""Get a node, from cache or disk.
945
After getting it, the node will be cached.
947
return self._get_nodes(self._internal_node_cache, node_indexes)
949
def _cache_leaf_values(self, nodes):
950
"""Cache directly from key => value, skipping the btree."""
951
if self._leaf_value_cache is not None:
952
for node in viewvalues(nodes):
953
for key, value in node.all_items():
954
if key in self._leaf_value_cache:
955
# Don't add the rest of the keys, we've seen this node
958
self._leaf_value_cache[key] = value
960
def _get_leaf_nodes(self, node_indexes):
961
"""Get a bunch of nodes, from cache or disk."""
962
found = self._get_nodes(self._leaf_node_cache, node_indexes)
963
self._cache_leaf_values(found)
966
def iter_all_entries(self):
967
"""Iterate over all keys within the index.
969
:return: An iterable of (index, key, value) or (index, key, value, reference_lists).
970
The former tuple is used when there are no reference lists in the
971
index, making the API compatible with simple key:value index types.
972
There is no defined order for the result iteration - it will be in
973
the most efficient order for the index.
975
if 'evil' in debug.debug_flags:
976
trace.mutter_callsite(3,
977
"iter_all_entries scales with size of history.")
978
if not self.key_count():
980
if self._row_offsets[-1] == 1:
981
# There is only the root node, and we read that via key_count()
982
if self.node_ref_lists:
983
for key, (value, refs) in self._root_node.all_items():
984
yield (self, key, value, refs)
986
for key, (value, refs) in self._root_node.all_items():
987
yield (self, key, value)
989
start_of_leaves = self._row_offsets[-2]
990
end_of_leaves = self._row_offsets[-1]
991
needed_offsets = list(range(start_of_leaves, end_of_leaves))
992
if needed_offsets == [0]:
993
# Special case when we only have a root node, as we have already
995
nodes = [(0, self._root_node)]
997
nodes = self._read_nodes(needed_offsets)
998
# We iterate strictly in-order so that we can use this function
999
# for spilling index builds to disk.
1000
if self.node_ref_lists:
1001
for _, node in nodes:
1002
for key, (value, refs) in node.all_items():
1003
yield (self, key, value, refs)
1005
for _, node in nodes:
1006
for key, (value, refs) in node.all_items():
1007
yield (self, key, value)
1010
def _multi_bisect_right(in_keys, fixed_keys):
1011
"""Find the positions where each 'in_key' would fit in fixed_keys.
1013
This is equivalent to doing "bisect_right" on each in_key into
1016
:param in_keys: A sorted list of keys to match with fixed_keys
1017
:param fixed_keys: A sorted list of keys to match against
1018
:return: A list of (integer position, [key list]) tuples.
1023
# no pointers in the fixed_keys list, which means everything must
1025
return [(0, in_keys)]
1027
# TODO: Iterating both lists will generally take M + N steps
1028
# Bisecting each key will generally take M * log2 N steps.
1029
# If we had an efficient way to compare, we could pick the method
1030
# based on which has the fewer number of steps.
1031
# There is also the argument that bisect_right is a compiled
1032
# function, so there is even more to be gained.
1033
# iter_steps = len(in_keys) + len(fixed_keys)
1034
# bisect_steps = len(in_keys) * math.log(len(fixed_keys), 2)
1035
if len(in_keys) == 1: # Bisect will always be faster for M = 1
1036
return [(bisect.bisect_right(fixed_keys, in_keys[0]), in_keys)]
1037
# elif bisect_steps < iter_steps:
1039
# for key in in_keys:
1040
# offsets.setdefault(bisect_right(fixed_keys, key),
1042
# return [(o, offsets[o]) for o in sorted(offsets)]
1043
in_keys_iter = iter(in_keys)
1044
fixed_keys_iter = enumerate(fixed_keys)
1045
cur_in_key = next(in_keys_iter)
1046
cur_fixed_offset, cur_fixed_key = next(fixed_keys_iter)
1048
class InputDone(Exception): pass
1049
class FixedDone(Exception): pass
1054
# TODO: Another possibility is that rather than iterating on each side,
1055
# we could use a combination of bisecting and iterating. For
1056
# example, while cur_in_key < fixed_key, bisect to find its
1057
# point, then iterate all matching keys, then bisect (restricted
1058
# to only the remainder) for the next one, etc.
1061
if cur_in_key < cur_fixed_key:
1063
cur_out = (cur_fixed_offset, cur_keys)
1064
output.append(cur_out)
1065
while cur_in_key < cur_fixed_key:
1066
cur_keys.append(cur_in_key)
1068
cur_in_key = next(in_keys_iter)
1069
except StopIteration:
1071
# At this point cur_in_key must be >= cur_fixed_key
1072
# step the cur_fixed_key until we pass the cur key, or walk off
1074
while cur_in_key >= cur_fixed_key:
1076
cur_fixed_offset, cur_fixed_key = next(fixed_keys_iter)
1077
except StopIteration:
1080
# We consumed all of the input, nothing more to do
1083
# There was some input left, but we consumed all of fixed, so we
1084
# have to add one more for the tail
1085
cur_keys = [cur_in_key]
1086
cur_keys.extend(in_keys_iter)
1087
cur_out = (len(fixed_keys), cur_keys)
1088
output.append(cur_out)
1091
def _walk_through_internal_nodes(self, keys):
1092
"""Take the given set of keys, and find the corresponding LeafNodes.
1094
:param keys: An unsorted iterable of keys to search for
1095
:return: (nodes, index_and_keys)
1096
nodes is a dict mapping {index: LeafNode}
1097
keys_at_index is a list of tuples of [(index, [keys for Leaf])]
1099
# 6 seconds spent in miss_torture using the sorted() line.
1100
# Even with out of order disk IO it seems faster not to sort it when
1101
# large queries are being made.
1102
keys_at_index = [(0, sorted(keys))]
1104
for row_pos, next_row_start in enumerate(self._row_offsets[1:-1]):
1105
node_indexes = [idx for idx, s_keys in keys_at_index]
1106
nodes = self._get_internal_nodes(node_indexes)
1108
next_nodes_and_keys = []
1109
for node_index, sub_keys in keys_at_index:
1110
node = nodes[node_index]
1111
positions = self._multi_bisect_right(sub_keys, node.keys)
1112
node_offset = next_row_start + node.offset
1113
next_nodes_and_keys.extend([(node_offset + pos, s_keys)
1114
for pos, s_keys in positions])
1115
keys_at_index = next_nodes_and_keys
1116
# We should now be at the _LeafNodes
1117
node_indexes = [idx for idx, s_keys in keys_at_index]
1119
# TODO: We may *not* want to always read all the nodes in one
1120
# big go. Consider setting a max size on this.
1121
nodes = self._get_leaf_nodes(node_indexes)
1122
return nodes, keys_at_index
1124
def iter_entries(self, keys):
1125
"""Iterate over keys within the index.
1127
:param keys: An iterable providing the keys to be retrieved.
1128
:return: An iterable as per iter_all_entries, but restricted to the
1129
keys supplied. No additional keys will be returned, and every
1130
key supplied that is in the index will be returned.
1132
# 6 seconds spent in miss_torture using the sorted() line.
1133
# Even with out of order disk IO it seems faster not to sort it when
1134
# large queries are being made.
1135
# However, now that we are doing multi-way bisecting, we need the keys
1136
# in sorted order anyway. We could change the multi-way code to not
1137
# require sorted order. (For example, it bisects for the first node,
1138
# does an in-order search until a key comes before the current point,
1139
# which it then bisects for, etc.)
1140
keys = frozenset(keys)
1144
if not self.key_count():
1148
if self._leaf_value_cache is None:
1152
value = self._leaf_value_cache.get(key, None)
1153
if value is not None:
1154
# This key is known not to be here, skip it
1156
if self.node_ref_lists:
1157
yield (self, key, value, refs)
1159
yield (self, key, value)
1161
needed_keys.append(key)
1167
nodes, nodes_and_keys = self._walk_through_internal_nodes(needed_keys)
1168
for node_index, sub_keys in nodes_and_keys:
1171
node = nodes[node_index]
1172
for next_sub_key in sub_keys:
1173
if next_sub_key in node:
1174
value, refs = node[next_sub_key]
1175
if self.node_ref_lists:
1176
yield (self, next_sub_key, value, refs)
1178
yield (self, next_sub_key, value)
1180
def _find_ancestors(self, keys, ref_list_num, parent_map, missing_keys):
1181
"""Find the parent_map information for the set of keys.
1183
This populates the parent_map dict and missing_keys set based on the
1184
queried keys. It also can fill out an arbitrary number of parents that
1185
it finds while searching for the supplied keys.
1187
It is unlikely that you want to call this directly. See
1188
"CombinedGraphIndex.find_ancestry()" for a more appropriate API.
1190
:param keys: A keys whose ancestry we want to return
1191
Every key will either end up in 'parent_map' or 'missing_keys'.
1192
:param ref_list_num: This index in the ref_lists is the parents we
1194
:param parent_map: {key: parent_keys} for keys that are present in this
1195
index. This may contain more entries than were in 'keys', that are
1196
reachable ancestors of the keys requested.
1197
:param missing_keys: keys which are known to be missing in this index.
1198
This may include parents that were not directly requested, but we
1199
were able to determine that they are not present in this index.
1200
:return: search_keys parents that were found but not queried to know
1201
if they are missing or present. Callers can re-query this index for
1202
those keys, and they will be placed into parent_map or missing_keys
1204
if not self.key_count():
1205
# We use key_count() to trigger reading the root node and
1206
# determining info about this BTreeGraphIndex
1207
# If we don't have any keys, then everything is missing
1208
missing_keys.update(keys)
1210
if ref_list_num >= self.node_ref_lists:
1211
raise ValueError('No ref list %d, index has %d ref lists'
1212
% (ref_list_num, self.node_ref_lists))
1214
# The main trick we are trying to accomplish is that when we find a
1215
# key listing its parents, we expect that the parent key is also likely
1216
# to sit on the same page. Allowing us to expand parents quickly
1217
# without suffering the full stack of bisecting, etc.
1218
nodes, nodes_and_keys = self._walk_through_internal_nodes(keys)
1220
# These are parent keys which could not be immediately resolved on the
1221
# page where the child was present. Note that we may already be
1222
# searching for that key, and it may actually be present [or known
1223
# missing] on one of the other pages we are reading.
1225
# We could try searching for them in the immediate previous or next
1226
# page. If they occur "later" we could put them in a pending lookup
1227
# set, and then for each node we read thereafter we could check to
1228
# see if they are present.
1229
# However, we don't know the impact of keeping this list of things
1230
# that I'm going to search for every node I come across from here on
1232
# It doesn't handle the case when the parent key is missing on a
1233
# page that we *don't* read. So we already have to handle being
1234
# re-entrant for that.
1235
# Since most keys contain a date string, they are more likely to be
1236
# found earlier in the file than later, but we would know that right
1237
# away (key < min_key), and wouldn't keep searching it on every other
1238
# page that we read.
1239
# Mostly, it is an idea, one which should be benchmarked.
1240
parents_not_on_page = set()
1242
for node_index, sub_keys in nodes_and_keys:
1245
# sub_keys is all of the keys we are looking for that should exist
1246
# on this page, if they aren't here, then they won't be found
1247
node = nodes[node_index]
1248
parents_to_check = set()
1249
for next_sub_key in sub_keys:
1250
if next_sub_key not in node:
1251
# This one is just not present in the index at all
1252
missing_keys.add(next_sub_key)
1254
value, refs = node[next_sub_key]
1255
parent_keys = refs[ref_list_num]
1256
parent_map[next_sub_key] = parent_keys
1257
parents_to_check.update(parent_keys)
1258
# Don't look for things we've already found
1259
parents_to_check = parents_to_check.difference(parent_map)
1260
# this can be used to test the benefit of having the check loop
1262
# parents_not_on_page.update(parents_to_check)
1264
while parents_to_check:
1265
next_parents_to_check = set()
1266
for key in parents_to_check:
1268
value, refs = node[key]
1269
parent_keys = refs[ref_list_num]
1270
parent_map[key] = parent_keys
1271
next_parents_to_check.update(parent_keys)
1273
# This parent either is genuinely missing, or should be
1274
# found on another page. Perf test whether it is better
1275
# to check if this node should fit on this page or not.
1276
# in the 'everything-in-one-pack' scenario, this *not*
1277
# doing the check is 237ms vs 243ms.
1278
# So slightly better, but I assume the standard 'lots
1279
# of packs' is going to show a reasonable improvement
1280
# from the check, because it avoids 'going around
1281
# again' for everything that is in another index
1282
# parents_not_on_page.add(key)
1283
# Missing for some reason
1284
if key < node.min_key:
1285
# in the case of bzr.dev, 3.4k/5.3k misses are
1286
# 'earlier' misses (65%)
1287
parents_not_on_page.add(key)
1288
elif key > node.max_key:
1289
# This parent key would be present on a different
1291
parents_not_on_page.add(key)
1293
# assert key != node.min_key and key != node.max_key
1294
# If it was going to be present, it would be on
1295
# *this* page, so mark it missing.
1296
missing_keys.add(key)
1297
parents_to_check = next_parents_to_check.difference(parent_map)
1298
# Might want to do another .difference() from missing_keys
1299
# parents_not_on_page could have been found on a different page, or be
1300
# known to be missing. So cull out everything that has already been
1302
search_keys = parents_not_on_page.difference(
1303
parent_map).difference(missing_keys)
1306
def iter_entries_prefix(self, keys):
1307
"""Iterate over keys within the index using prefix matching.
1309
Prefix matching is applied within the tuple of a key, not to within
1310
the bytestring of each key element. e.g. if you have the keys ('foo',
1311
'bar'), ('foobar', 'gam') and do a prefix search for ('foo', None) then
1312
only the former key is returned.
1314
WARNING: Note that this method currently causes a full index parse
1315
unconditionally (which is reasonably appropriate as it is a means for
1316
thunking many small indices into one larger one and still supplies
1317
iter_all_entries at the thunk layer).
1319
:param keys: An iterable providing the key prefixes to be retrieved.
1320
Each key prefix takes the form of a tuple the length of a key, but
1321
with the last N elements 'None' rather than a regular bytestring.
1322
The first element cannot be 'None'.
1323
:return: An iterable as per iter_all_entries, but restricted to the
1324
keys with a matching prefix to those supplied. No additional keys
1325
will be returned, and every match that is in the index will be
1328
keys = sorted(set(keys))
1331
# Load if needed to check key lengths
1332
if self._key_count is None:
1333
self._get_root_node()
1334
# TODO: only access nodes that can satisfy the prefixes we are looking
1335
# for. For now, to meet API usage (as this function is not used by
1336
# current breezy) just suck the entire index and iterate in memory.
1338
if self.node_ref_lists:
1339
if self._key_length == 1:
1340
for _1, key, value, refs in self.iter_all_entries():
1341
nodes[key] = value, refs
1344
for _1, key, value, refs in self.iter_all_entries():
1345
key_value = key, value, refs
1346
# For a key of (foo, bar, baz) create
1347
# _nodes_by_key[foo][bar][baz] = key_value
1348
key_dict = nodes_by_key
1349
for subkey in key[:-1]:
1350
key_dict = key_dict.setdefault(subkey, {})
1351
key_dict[key[-1]] = key_value
1353
if self._key_length == 1:
1354
for _1, key, value in self.iter_all_entries():
1358
for _1, key, value in self.iter_all_entries():
1359
key_value = key, value
1360
# For a key of (foo, bar, baz) create
1361
# _nodes_by_key[foo][bar][baz] = key_value
1362
key_dict = nodes_by_key
1363
for subkey in key[:-1]:
1364
key_dict = key_dict.setdefault(subkey, {})
1365
key_dict[key[-1]] = key_value
1366
if self._key_length == 1:
1368
index._sanity_check_key(self, key)
1370
if self.node_ref_lists:
1371
value, node_refs = nodes[key]
1372
yield self, key, value, node_refs
1374
yield self, key, nodes[key]
1378
for entry in index._iter_entries_prefix(self, nodes_by_key, keys):
1381
def key_count(self):
1382
"""Return an estimate of the number of keys in this index.
1384
For BTreeGraphIndex the estimate is exact as it is contained in the
1387
if self._key_count is None:
1388
self._get_root_node()
1389
return self._key_count
1391
def _compute_row_offsets(self):
1392
"""Fill out the _row_offsets attribute based on _row_lengths."""
1395
for row in self._row_lengths:
1396
offsets.append(row_offset)
1398
offsets.append(row_offset)
1399
self._row_offsets = offsets
1401
def _parse_header_from_bytes(self, bytes):
1402
"""Parse the header from a region of bytes.
1404
:param bytes: The data to parse.
1405
:return: An offset, data tuple such as readv yields, for the unparsed
1406
data. (which may be of length 0).
1408
signature = bytes[0:len(self._signature())]
1409
if not signature == self._signature():
1410
raise errors.BadIndexFormatSignature(self._name, BTreeGraphIndex)
1411
lines = bytes[len(self._signature()):].splitlines()
1412
options_line = lines[0]
1413
if not options_line.startswith(_OPTION_NODE_REFS):
1414
raise errors.BadIndexOptions(self)
1416
self.node_ref_lists = int(options_line[len(_OPTION_NODE_REFS):])
1418
raise errors.BadIndexOptions(self)
1419
options_line = lines[1]
1420
if not options_line.startswith(_OPTION_KEY_ELEMENTS):
1421
raise errors.BadIndexOptions(self)
1423
self._key_length = int(options_line[len(_OPTION_KEY_ELEMENTS):])
1425
raise errors.BadIndexOptions(self)
1426
options_line = lines[2]
1427
if not options_line.startswith(_OPTION_LEN):
1428
raise errors.BadIndexOptions(self)
1430
self._key_count = int(options_line[len(_OPTION_LEN):])
1432
raise errors.BadIndexOptions(self)
1433
options_line = lines[3]
1434
if not options_line.startswith(_OPTION_ROW_LENGTHS):
1435
raise errors.BadIndexOptions(self)
1437
self._row_lengths = [int(length) for length in
1438
options_line[len(_OPTION_ROW_LENGTHS):].split(b',')
1441
raise errors.BadIndexOptions(self)
1442
self._compute_row_offsets()
1444
# calculate the bytes we have processed
1445
header_end = (len(signature) + sum(map(len, lines[0:4])) + 4)
1446
return header_end, bytes[header_end:]
1448
def _read_nodes(self, nodes):
1449
"""Read some nodes from disk into the LRU cache.
1451
This performs a readv to get the node data into memory, and parses each
1452
node, then yields it to the caller. The nodes are requested in the
1453
supplied order. If possible doing sort() on the list before requesting
1454
a read may improve performance.
1456
:param nodes: The nodes to read. 0 - first node, 1 - second node etc.
1459
# may be the byte string of the whole file
1461
# list of (offset, length) regions of the file that should, evenually
1462
# be read in to data_ranges, either from 'bytes' or from the transport
1464
base_offset = self._base_offset
1466
offset = (index * _PAGE_SIZE)
1469
# Root node - special case
1471
size = min(_PAGE_SIZE, self._size)
1473
# The only case where we don't know the size, is for very
1474
# small indexes. So we read the whole thing
1475
bytes = self._transport.get_bytes(self._name)
1476
num_bytes = len(bytes)
1477
self._size = num_bytes - base_offset
1478
# the whole thing should be parsed out of 'bytes'
1479
ranges = [(start, min(_PAGE_SIZE, num_bytes - start))
1480
for start in range(base_offset, num_bytes, _PAGE_SIZE)]
1483
if offset > self._size:
1484
raise AssertionError('tried to read past the end'
1485
' of the file %s > %s'
1486
% (offset, self._size))
1487
size = min(size, self._size - offset)
1488
ranges.append((base_offset + offset, size))
1491
elif bytes is not None:
1492
# already have the whole file
1493
data_ranges = [(start, bytes[start:start+size])
1494
for start, size in ranges]
1495
elif self._file is None:
1496
data_ranges = self._transport.readv(self._name, ranges)
1499
for offset, size in ranges:
1500
self._file.seek(offset)
1501
data_ranges.append((offset, self._file.read(size)))
1502
for offset, data in data_ranges:
1503
offset -= base_offset
1505
# extract the header
1506
offset, data = self._parse_header_from_bytes(data)
1509
bytes = zlib.decompress(data)
1510
if bytes.startswith(_LEAF_FLAG):
1511
node = self._leaf_factory(bytes, self._key_length,
1512
self.node_ref_lists)
1513
elif bytes.startswith(_INTERNAL_FLAG):
1514
node = _InternalNode(bytes)
1516
raise AssertionError("Unknown node type for %r" % bytes)
1517
yield offset / _PAGE_SIZE, node
1519
def _signature(self):
1520
"""The file signature for this index type."""
1524
"""Validate that everything in the index can be accessed."""
1525
# just read and parse every node.
1526
self._get_root_node()
1527
if len(self._row_lengths) > 1:
1528
start_node = self._row_offsets[1]
1530
# We shouldn't be reading anything anyway
1532
node_end = self._row_offsets[-1]
1533
for node in self._read_nodes(list(range(start_node, node_end))):
1537
_gcchk_factory = _LeafNode
1540
from breezy import _btree_serializer_pyx as _btree_serializer
1541
_gcchk_factory = _btree_serializer._parse_into_chk
1542
except ImportError as e:
1543
osutils.failed_to_load_extension(e)
1544
from breezy import _btree_serializer_py as _btree_serializer
added
removed
breezy/btree_index.py
