1
# Copyright (C) 2008, 2009, 2010 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
21
from bisect import bisect_right
37
from bzrlib.index import _OPTION_NODE_REFS, _OPTION_KEY_ELEMENTS, _OPTION_LEN
38
from bzrlib.transport import get_transport
41
_BTSIGNATURE = "B+Tree Graph Index 2\n"
42
_OPTION_ROW_LENGTHS = "row_lengths="
43
_LEAF_FLAG = "type=leaf\n"
44
_INTERNAL_FLAG = "type=internal\n"
45
_INTERNAL_OFFSET = "offset="
47
_RESERVED_HEADER_BYTES = 120
50
# 4K per page: 4MB - 1000 entries
51
_NODE_CACHE_SIZE = 1000
54
class _BuilderRow(object):
55
"""The stored state accumulated while writing out a row in the index.
57
:ivar spool: A temporary file used to accumulate nodes for this row
59
:ivar nodes: The count of nodes emitted so far.
63
"""Create a _BuilderRow."""
65
self.spool = None# tempfile.TemporaryFile(prefix='bzr-index-row-')
68
def finish_node(self, pad=True):
69
byte_lines, _, padding = self.writer.finish()
71
self.spool = cStringIO.StringIO()
73
self.spool.write("\x00" * _RESERVED_HEADER_BYTES)
75
# We got bigger than 1 node, switch to a temp file
76
spool = tempfile.TemporaryFile(prefix='bzr-index-row-')
77
spool.write(self.spool.getvalue())
80
if not pad and padding:
82
skipped_bytes = padding
83
self.spool.writelines(byte_lines)
84
remainder = (self.spool.tell() + skipped_bytes) % _PAGE_SIZE
86
raise AssertionError("incorrect node length: %d, %d"
87
% (self.spool.tell(), remainder))
92
class _InternalBuilderRow(_BuilderRow):
93
"""The stored state accumulated while writing out internal rows."""
95
def finish_node(self, pad=True):
97
raise AssertionError("Must pad internal nodes only.")
98
_BuilderRow.finish_node(self)
101
class _LeafBuilderRow(_BuilderRow):
102
"""The stored state accumulated while writing out a leaf rows."""
105
class BTreeBuilder(index.GraphIndexBuilder):
106
"""A Builder for B+Tree based Graph indices.
108
The resulting graph has the structure:
110
_SIGNATURE OPTIONS NODES
111
_SIGNATURE := 'B+Tree Graph Index 1' NEWLINE
112
OPTIONS := REF_LISTS KEY_ELEMENTS LENGTH
113
REF_LISTS := 'node_ref_lists=' DIGITS NEWLINE
114
KEY_ELEMENTS := 'key_elements=' DIGITS NEWLINE
115
LENGTH := 'len=' DIGITS NEWLINE
116
ROW_LENGTHS := 'row_lengths' DIGITS (COMMA DIGITS)*
117
NODES := NODE_COMPRESSED*
118
NODE_COMPRESSED:= COMPRESSED_BYTES{4096}
119
NODE_RAW := INTERNAL | LEAF
120
INTERNAL := INTERNAL_FLAG POINTERS
121
LEAF := LEAF_FLAG ROWS
122
KEY_ELEMENT := Not-whitespace-utf8
123
KEY := KEY_ELEMENT (NULL KEY_ELEMENT)*
125
ROW := KEY NULL ABSENT? NULL REFERENCES NULL VALUE NEWLINE
127
REFERENCES := REFERENCE_LIST (TAB REFERENCE_LIST){node_ref_lists - 1}
128
REFERENCE_LIST := (REFERENCE (CR REFERENCE)*)?
130
VALUE := no-newline-no-null-bytes
133
def __init__(self, reference_lists=0, key_elements=1, spill_at=100000):
134
"""See GraphIndexBuilder.__init__.
136
:param spill_at: Optional parameter controlling the maximum number
137
of nodes that BTreeBuilder will hold in memory.
139
index.GraphIndexBuilder.__init__(self, reference_lists=reference_lists,
140
key_elements=key_elements)
141
self._spill_at = spill_at
142
self._backing_indices = []
143
# A map of {key: (node_refs, value)}
145
# Indicate it hasn't been built yet
146
self._nodes_by_key = None
147
self._optimize_for_size = False
149
def add_node(self, key, value, references=()):
150
"""Add a node to the index.
152
If adding the node causes the builder to reach its spill_at threshold,
153
disk spilling will be triggered.
155
:param key: The key. keys are non-empty tuples containing
156
as many whitespace-free utf8 bytestrings as the key length
157
defined for this index.
158
:param references: An iterable of iterables of keys. Each is a
159
reference to another key.
160
:param value: The value to associate with the key. It may be any
161
bytes as long as it does not contain \0 or \n.
163
# Ensure that 'key' is a StaticTuple
164
key = static_tuple.StaticTuple.from_sequence(key).intern()
165
# we don't care about absent_references
166
node_refs, _ = self._check_key_ref_value(key, references, value)
167
if key in self._nodes:
168
raise errors.BadIndexDuplicateKey(key, self)
169
self._nodes[key] = static_tuple.StaticTuple(node_refs, value)
170
if self._nodes_by_key is not None and self._key_length > 1:
171
self._update_nodes_by_key(key, value, node_refs)
172
if len(self._nodes) < self._spill_at:
174
self._spill_mem_keys_to_disk()
176
def _spill_mem_keys_to_disk(self):
177
"""Write the in memory keys down to disk to cap memory consumption.
179
If we already have some keys written to disk, we will combine them so
180
as to preserve the sorted order. The algorithm for combining uses
181
powers of two. So on the first spill, write all mem nodes into a
182
single index. On the second spill, combine the mem nodes with the nodes
183
on disk to create a 2x sized disk index and get rid of the first index.
184
On the third spill, create a single new disk index, which will contain
185
the mem nodes, and preserve the existing 2x sized index. On the fourth,
186
combine mem with the first and second indexes, creating a new one of
187
size 4x. On the fifth create a single new one, etc.
189
if self._combine_backing_indices:
190
(new_backing_file, size,
191
backing_pos) = self._spill_mem_keys_and_combine()
193
new_backing_file, size = self._spill_mem_keys_without_combining()
194
# Note: The transport here isn't strictly needed, because we will use
195
# direct access to the new_backing._file object
196
new_backing = BTreeGraphIndex(get_transport('.'), '<temp>', size)
197
# GC will clean up the file
198
new_backing._file = new_backing_file
199
if self._combine_backing_indices:
200
if len(self._backing_indices) == backing_pos:
201
self._backing_indices.append(None)
202
self._backing_indices[backing_pos] = new_backing
203
for backing_pos in range(backing_pos):
204
self._backing_indices[backing_pos] = None
206
self._backing_indices.append(new_backing)
208
self._nodes_by_key = None
210
def _spill_mem_keys_without_combining(self):
211
return self._write_nodes(self._iter_mem_nodes(), allow_optimize=False)
213
def _spill_mem_keys_and_combine(self):
214
iterators_to_combine = [self._iter_mem_nodes()]
216
for pos, backing in enumerate(self._backing_indices):
220
iterators_to_combine.append(backing.iter_all_entries())
221
backing_pos = pos + 1
222
new_backing_file, size = \
223
self._write_nodes(self._iter_smallest(iterators_to_combine),
224
allow_optimize=False)
225
return new_backing_file, size, backing_pos
227
def add_nodes(self, nodes):
228
"""Add nodes to the index.
230
:param nodes: An iterable of (key, node_refs, value) entries to add.
232
if self.reference_lists:
233
for (key, value, node_refs) in nodes:
234
self.add_node(key, value, node_refs)
236
for (key, value) in nodes:
237
self.add_node(key, value)
239
def _iter_mem_nodes(self):
240
"""Iterate over the nodes held in memory."""
242
if self.reference_lists:
243
for key in sorted(nodes):
244
references, value = nodes[key]
245
yield self, key, value, references
247
for key in sorted(nodes):
248
references, value = nodes[key]
249
yield self, key, value
251
def _iter_smallest(self, iterators_to_combine):
252
if len(iterators_to_combine) == 1:
253
for value in iterators_to_combine[0]:
257
for iterator in iterators_to_combine:
259
current_values.append(iterator.next())
260
except StopIteration:
261
current_values.append(None)
264
# Decorate candidates with the value to allow 2.4's min to be used.
265
candidates = [(item[1][1], item) for item
266
in enumerate(current_values) if item[1] is not None]
267
if not len(candidates):
269
selected = min(candidates)
270
# undecorate back to (pos, node)
271
selected = selected[1]
272
if last == selected[1][1]:
273
raise errors.BadIndexDuplicateKey(last, self)
274
last = selected[1][1]
275
# Yield, with self as the index
276
yield (self,) + selected[1][1:]
279
current_values[pos] = iterators_to_combine[pos].next()
280
except StopIteration:
281
current_values[pos] = None
283
def _add_key(self, string_key, line, rows, allow_optimize=True):
284
"""Add a key to the current chunk.
286
:param string_key: The key to add.
287
:param line: The fully serialised key and value.
288
:param allow_optimize: If set to False, prevent setting the optimize
289
flag when writing out. This is used by the _spill_mem_keys_to_disk
292
if rows[-1].writer is None:
293
# opening a new leaf chunk;
294
for pos, internal_row in enumerate(rows[:-1]):
295
# flesh out any internal nodes that are needed to
296
# preserve the height of the tree
297
if internal_row.writer is None:
299
if internal_row.nodes == 0:
300
length -= _RESERVED_HEADER_BYTES # padded
302
optimize_for_size = self._optimize_for_size
304
optimize_for_size = False
305
internal_row.writer = chunk_writer.ChunkWriter(length, 0,
306
optimize_for_size=optimize_for_size)
307
internal_row.writer.write(_INTERNAL_FLAG)
308
internal_row.writer.write(_INTERNAL_OFFSET +
309
str(rows[pos + 1].nodes) + "\n")
312
if rows[-1].nodes == 0:
313
length -= _RESERVED_HEADER_BYTES # padded
314
rows[-1].writer = chunk_writer.ChunkWriter(length,
315
optimize_for_size=self._optimize_for_size)
316
rows[-1].writer.write(_LEAF_FLAG)
317
if rows[-1].writer.write(line):
318
# this key did not fit in the node:
319
rows[-1].finish_node()
320
key_line = string_key + "\n"
322
for row in reversed(rows[:-1]):
323
# Mark the start of the next node in the node above. If it
324
# doesn't fit then propagate upwards until we find one that
326
if row.writer.write(key_line):
329
# We've found a node that can handle the pointer.
332
# If we reached the current root without being able to mark the
333
# division point, then we need a new root:
336
if 'index' in debug.debug_flags:
337
trace.mutter('Inserting new global row.')
338
new_row = _InternalBuilderRow()
340
rows.insert(0, new_row)
341
# This will be padded, hence the -100
342
new_row.writer = chunk_writer.ChunkWriter(
343
_PAGE_SIZE - _RESERVED_HEADER_BYTES,
345
optimize_for_size=self._optimize_for_size)
346
new_row.writer.write(_INTERNAL_FLAG)
347
new_row.writer.write(_INTERNAL_OFFSET +
348
str(rows[1].nodes - 1) + "\n")
349
new_row.writer.write(key_line)
350
self._add_key(string_key, line, rows, allow_optimize=allow_optimize)
352
def _write_nodes(self, node_iterator, allow_optimize=True):
353
"""Write node_iterator out as a B+Tree.
355
:param node_iterator: An iterator of sorted nodes. Each node should
356
match the output given by iter_all_entries.
357
:param allow_optimize: If set to False, prevent setting the optimize
358
flag when writing out. This is used by the _spill_mem_keys_to_disk
360
:return: A file handle for a temporary file containing a B+Tree for
363
# The index rows - rows[0] is the root, rows[1] is the layer under it
366
# forward sorted by key. In future we may consider topological sorting,
367
# at the cost of table scans for direct lookup, or a second index for
370
# A stack with the number of nodes of each size. 0 is the root node
371
# and must always be 1 (if there are any nodes in the tree).
372
self.row_lengths = []
373
# Loop over all nodes adding them to the bottom row
374
# (rows[-1]). When we finish a chunk in a row,
375
# propagate the key that didn't fit (comes after the chunk) to the
376
# row above, transitively.
377
for node in node_iterator:
379
# First key triggers the first row
380
rows.append(_LeafBuilderRow())
382
string_key, line = _btree_serializer._flatten_node(node,
383
self.reference_lists)
384
self._add_key(string_key, line, rows, allow_optimize=allow_optimize)
385
for row in reversed(rows):
386
pad = (type(row) != _LeafBuilderRow)
387
row.finish_node(pad=pad)
388
lines = [_BTSIGNATURE]
389
lines.append(_OPTION_NODE_REFS + str(self.reference_lists) + '\n')
390
lines.append(_OPTION_KEY_ELEMENTS + str(self._key_length) + '\n')
391
lines.append(_OPTION_LEN + str(key_count) + '\n')
392
row_lengths = [row.nodes for row in rows]
393
lines.append(_OPTION_ROW_LENGTHS + ','.join(map(str, row_lengths)) + '\n')
394
if row_lengths and row_lengths[-1] > 1:
395
result = tempfile.NamedTemporaryFile(prefix='bzr-index-')
397
result = cStringIO.StringIO()
398
result.writelines(lines)
399
position = sum(map(len, lines))
401
if position > _RESERVED_HEADER_BYTES:
402
raise AssertionError("Could not fit the header in the"
403
" reserved space: %d > %d"
404
% (position, _RESERVED_HEADER_BYTES))
405
# write the rows out:
407
reserved = _RESERVED_HEADER_BYTES # reserved space for first node
410
# copy nodes to the finalised file.
411
# Special case the first node as it may be prefixed
412
node = row.spool.read(_PAGE_SIZE)
413
result.write(node[reserved:])
414
if len(node) == _PAGE_SIZE:
415
result.write("\x00" * (reserved - position))
416
position = 0 # Only the root row actually has an offset
417
copied_len = osutils.pumpfile(row.spool, result)
418
if copied_len != (row.nodes - 1) * _PAGE_SIZE:
419
if type(row) != _LeafBuilderRow:
420
raise AssertionError("Incorrect amount of data copied"
421
" expected: %d, got: %d"
422
% ((row.nodes - 1) * _PAGE_SIZE,
430
"""Finalise the index.
432
:return: A file handle for a temporary file containing the nodes added
435
return self._write_nodes(self.iter_all_entries())[0]
437
def iter_all_entries(self):
438
"""Iterate over all keys within the index
440
:return: An iterable of (index, key, value, reference_lists). There is
441
no defined order for the result iteration - it will be in the most
442
efficient order for the index (in this case dictionary hash order).
444
if 'evil' in debug.debug_flags:
445
trace.mutter_callsite(3,
446
"iter_all_entries scales with size of history.")
447
# Doing serial rather than ordered would be faster; but this shouldn't
448
# be getting called routinely anyway.
449
iterators = [self._iter_mem_nodes()]
450
for backing in self._backing_indices:
451
if backing is not None:
452
iterators.append(backing.iter_all_entries())
453
if len(iterators) == 1:
455
return self._iter_smallest(iterators)
457
def iter_entries(self, keys):
458
"""Iterate over keys within the index.
460
:param keys: An iterable providing the keys to be retrieved.
461
:return: An iterable of (index, key, value, reference_lists). There is no
462
defined order for the result iteration - it will be in the most
463
efficient order for the index (keys iteration order in this case).
466
# Note: We don't use keys.intersection() here. If you read the C api,
467
# set.intersection(other) special cases when other is a set and
468
# will iterate the smaller of the two and lookup in the other.
469
# It does *not* do this for any other type (even dict, unlike
470
# some other set functions.) Since we expect keys is generally <<
471
# self._nodes, it is faster to iterate over it in a list
474
local_keys = [key for key in keys if key in nodes]
475
if self.reference_lists:
476
for key in local_keys:
478
yield self, key, node[1], node[0]
480
for key in local_keys:
482
yield self, key, node[1]
483
# Find things that are in backing indices that have not been handled
485
if not self._backing_indices:
486
return # We won't find anything there either
487
# Remove all of the keys that we found locally
488
keys.difference_update(local_keys)
489
for backing in self._backing_indices:
494
for node in backing.iter_entries(keys):
496
yield (self,) + node[1:]
498
def iter_entries_prefix(self, keys):
499
"""Iterate over keys within the index using prefix matching.
501
Prefix matching is applied within the tuple of a key, not to within
502
the bytestring of each key element. e.g. if you have the keys ('foo',
503
'bar'), ('foobar', 'gam') and do a prefix search for ('foo', None) then
504
only the former key is returned.
506
:param keys: An iterable providing the key prefixes to be retrieved.
507
Each key prefix takes the form of a tuple the length of a key, but
508
with the last N elements 'None' rather than a regular bytestring.
509
The first element cannot be 'None'.
510
:return: An iterable as per iter_all_entries, but restricted to the
511
keys with a matching prefix to those supplied. No additional keys
512
will be returned, and every match that is in the index will be
515
# XXX: To much duplication with the GraphIndex class; consider finding
516
# a good place to pull out the actual common logic.
520
for backing in self._backing_indices:
523
for node in backing.iter_entries_prefix(keys):
524
yield (self,) + node[1:]
525
if self._key_length == 1:
529
raise errors.BadIndexKey(key)
530
if len(key) != self._key_length:
531
raise errors.BadIndexKey(key)
533
node = self._nodes[key]
536
if self.reference_lists:
537
yield self, key, node[1], node[0]
539
yield self, key, node[1]
544
raise errors.BadIndexKey(key)
545
if len(key) != self._key_length:
546
raise errors.BadIndexKey(key)
547
# find what it refers to:
548
key_dict = self._get_nodes_by_key()
550
# find the subdict to return
552
while len(elements) and elements[0] is not None:
553
key_dict = key_dict[elements[0]]
556
# a non-existant lookup.
561
key_dict = dicts.pop(-1)
562
# can't be empty or would not exist
563
item, value = key_dict.iteritems().next()
564
if type(value) == dict:
566
dicts.extend(key_dict.itervalues())
569
for value in key_dict.itervalues():
570
yield (self, ) + tuple(value)
572
yield (self, ) + key_dict
574
def _get_nodes_by_key(self):
575
if self._nodes_by_key is None:
577
if self.reference_lists:
578
for key, (references, value) in self._nodes.iteritems():
579
key_dict = nodes_by_key
580
for subkey in key[:-1]:
581
key_dict = key_dict.setdefault(subkey, {})
582
key_dict[key[-1]] = key, value, references
584
for key, (references, value) in self._nodes.iteritems():
585
key_dict = nodes_by_key
586
for subkey in key[:-1]:
587
key_dict = key_dict.setdefault(subkey, {})
588
key_dict[key[-1]] = key, value
589
self._nodes_by_key = nodes_by_key
590
return self._nodes_by_key
593
"""Return an estimate of the number of keys in this index.
595
For InMemoryGraphIndex the estimate is exact.
597
return len(self._nodes) + sum(backing.key_count() for backing in
598
self._backing_indices if backing is not None)
601
"""In memory index's have no known corruption at the moment."""
604
class _LeafNode(object):
605
"""A leaf node for a serialised B+Tree index."""
607
__slots__ = ('keys', 'min_key', 'max_key')
609
def __init__(self, bytes, key_length, ref_list_length):
610
"""Parse bytes to create a leaf node object."""
611
# splitlines mangles the \r delimiters.. don't use it.
612
key_list = _btree_serializer._parse_leaf_lines(bytes,
613
key_length, ref_list_length)
615
self.min_key = key_list[0][0]
616
self.max_key = key_list[-1][0]
618
self.min_key = self.max_key = None
619
self.keys = dict(key_list)
622
class _InternalNode(object):
623
"""An internal node for a serialised B+Tree index."""
625
__slots__ = ('keys', 'offset')
627
def __init__(self, bytes):
628
"""Parse bytes to create an internal node object."""
629
# splitlines mangles the \r delimiters.. don't use it.
630
self.keys = self._parse_lines(bytes.split('\n'))
632
def _parse_lines(self, lines):
634
self.offset = int(lines[1][7:])
635
as_st = static_tuple.StaticTuple.from_sequence
636
for line in lines[2:]:
639
nodes.append(as_st(map(intern, line.split('\0'))).intern())
643
class BTreeGraphIndex(object):
644
"""Access to nodes via the standard GraphIndex interface for B+Tree's.
646
Individual nodes are held in a LRU cache. This holds the root node in
647
memory except when very large walks are done.
650
def __init__(self, transport, name, size, unlimited_cache=False,
652
"""Create a B+Tree index object on the index name.
654
:param transport: The transport to read data for the index from.
655
:param name: The file name of the index on transport.
656
:param size: Optional size of the index in bytes. This allows
657
compatibility with the GraphIndex API, as well as ensuring that
658
the initial read (to read the root node header) can be done
659
without over-reading even on empty indices, and on small indices
660
allows single-IO to read the entire index.
661
:param unlimited_cache: If set to True, then instead of using an
662
LRUCache with size _NODE_CACHE_SIZE, we will use a dict and always
663
cache all leaf nodes.
664
:param offset: The start of the btree index data isn't byte 0 of the
665
file. Instead it starts at some point later.
667
self._transport = transport
671
self._recommended_pages = self._compute_recommended_pages()
672
self._root_node = None
673
self._base_offset = offset
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
type(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 xrange(total_pages)
791
if x not in cached_offsets]
793
expanded = 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.keys())
913
cached_offsets.update(self._leaf_node_cache.keys())
914
if self._root_node is not None:
915
cached_offsets.add(0)
916
return cached_offsets
918
def _get_root_node(self):
919
if self._root_node is None:
920
# We may not have a root node yet
921
self._get_internal_nodes([0])
922
return self._root_node
924
def _get_nodes(self, cache, node_indexes):
927
for idx in node_indexes:
928
if idx == 0 and self._root_node is not None:
929
found[0] = self._root_node
932
found[idx] = cache[idx]
937
needed = self._expand_offsets(needed)
938
found.update(self._get_and_cache_nodes(needed))
941
def _get_internal_nodes(self, node_indexes):
942
"""Get a node, from cache or disk.
944
After getting it, the node will be cached.
946
return self._get_nodes(self._internal_node_cache, node_indexes)
948
def _cache_leaf_values(self, nodes):
949
"""Cache directly from key => value, skipping the btree."""
950
if self._leaf_value_cache is not None:
951
for node in nodes.itervalues():
952
for key, value in node.keys.iteritems():
953
if key in self._leaf_value_cache:
954
# Don't add the rest of the keys, we've seen this node
957
self._leaf_value_cache[key] = value
959
def _get_leaf_nodes(self, node_indexes):
960
"""Get a bunch of nodes, from cache or disk."""
961
found = self._get_nodes(self._leaf_node_cache, node_indexes)
962
self._cache_leaf_values(found)
965
def iter_all_entries(self):
966
"""Iterate over all keys within the index.
968
:return: An iterable of (index, key, value) or (index, key, value, reference_lists).
969
The former tuple is used when there are no reference lists in the
970
index, making the API compatible with simple key:value index types.
971
There is no defined order for the result iteration - it will be in
972
the most efficient order for the index.
974
if 'evil' in debug.debug_flags:
975
trace.mutter_callsite(3,
976
"iter_all_entries scales with size of history.")
977
if not self.key_count():
979
if self._row_offsets[-1] == 1:
980
# There is only the root node, and we read that via key_count()
981
if self.node_ref_lists:
982
for key, (value, refs) in sorted(self._root_node.keys.items()):
983
yield (self, key, value, refs)
985
for key, (value, refs) in sorted(self._root_node.keys.items()):
986
yield (self, key, value)
988
start_of_leaves = self._row_offsets[-2]
989
end_of_leaves = self._row_offsets[-1]
990
needed_offsets = range(start_of_leaves, end_of_leaves)
991
if needed_offsets == [0]:
992
# Special case when we only have a root node, as we have already
994
nodes = [(0, self._root_node)]
996
nodes = self._read_nodes(needed_offsets)
997
# We iterate strictly in-order so that we can use this function
998
# for spilling index builds to disk.
999
if self.node_ref_lists:
1000
for _, node in nodes:
1001
for key, (value, refs) in sorted(node.keys.items()):
1002
yield (self, key, value, refs)
1004
for _, node in nodes:
1005
for key, (value, refs) in sorted(node.keys.items()):
1006
yield (self, key, value)
1009
def _multi_bisect_right(in_keys, fixed_keys):
1010
"""Find the positions where each 'in_key' would fit in fixed_keys.
1012
This is equivalent to doing "bisect_right" on each in_key into
1015
:param in_keys: A sorted list of keys to match with fixed_keys
1016
:param fixed_keys: A sorted list of keys to match against
1017
:return: A list of (integer position, [key list]) tuples.
1022
# no pointers in the fixed_keys list, which means everything must
1024
return [(0, in_keys)]
1026
# TODO: Iterating both lists will generally take M + N steps
1027
# Bisecting each key will generally take M * log2 N steps.
1028
# If we had an efficient way to compare, we could pick the method
1029
# based on which has the fewer number of steps.
1030
# There is also the argument that bisect_right is a compiled
1031
# function, so there is even more to be gained.
1032
# iter_steps = len(in_keys) + len(fixed_keys)
1033
# bisect_steps = len(in_keys) * math.log(len(fixed_keys), 2)
1034
if len(in_keys) == 1: # Bisect will always be faster for M = 1
1035
return [(bisect_right(fixed_keys, in_keys[0]), in_keys)]
1036
# elif bisect_steps < iter_steps:
1038
# for key in in_keys:
1039
# offsets.setdefault(bisect_right(fixed_keys, key),
1041
# return [(o, offsets[o]) for o in sorted(offsets)]
1042
in_keys_iter = iter(in_keys)
1043
fixed_keys_iter = enumerate(fixed_keys)
1044
cur_in_key = in_keys_iter.next()
1045
cur_fixed_offset, cur_fixed_key = fixed_keys_iter.next()
1047
class InputDone(Exception): pass
1048
class FixedDone(Exception): pass
1053
# TODO: Another possibility is that rather than iterating on each side,
1054
# we could use a combination of bisecting and iterating. For
1055
# example, while cur_in_key < fixed_key, bisect to find its
1056
# point, then iterate all matching keys, then bisect (restricted
1057
# to only the remainder) for the next one, etc.
1060
if cur_in_key < cur_fixed_key:
1062
cur_out = (cur_fixed_offset, cur_keys)
1063
output.append(cur_out)
1064
while cur_in_key < cur_fixed_key:
1065
cur_keys.append(cur_in_key)
1067
cur_in_key = in_keys_iter.next()
1068
except StopIteration:
1070
# At this point cur_in_key must be >= cur_fixed_key
1071
# step the cur_fixed_key until we pass the cur key, or walk off
1073
while cur_in_key >= cur_fixed_key:
1075
cur_fixed_offset, cur_fixed_key = fixed_keys_iter.next()
1076
except StopIteration:
1079
# We consumed all of the input, nothing more to do
1082
# There was some input left, but we consumed all of fixed, so we
1083
# have to add one more for the tail
1084
cur_keys = [cur_in_key]
1085
cur_keys.extend(in_keys_iter)
1086
cur_out = (len(fixed_keys), cur_keys)
1087
output.append(cur_out)
1090
def _walk_through_internal_nodes(self, keys):
1091
"""Take the given set of keys, and find the corresponding LeafNodes.
1093
:param keys: An unsorted iterable of keys to search for
1094
:return: (nodes, index_and_keys)
1095
nodes is a dict mapping {index: LeafNode}
1096
keys_at_index is a list of tuples of [(index, [keys for Leaf])]
1098
# 6 seconds spent in miss_torture using the sorted() line.
1099
# Even with out of order disk IO it seems faster not to sort it when
1100
# large queries are being made.
1101
keys_at_index = [(0, sorted(keys))]
1103
for row_pos, next_row_start in enumerate(self._row_offsets[1:-1]):
1104
node_indexes = [idx for idx, s_keys in keys_at_index]
1105
nodes = self._get_internal_nodes(node_indexes)
1107
next_nodes_and_keys = []
1108
for node_index, sub_keys in keys_at_index:
1109
node = nodes[node_index]
1110
positions = self._multi_bisect_right(sub_keys, node.keys)
1111
node_offset = next_row_start + node.offset
1112
next_nodes_and_keys.extend([(node_offset + pos, s_keys)
1113
for pos, s_keys in positions])
1114
keys_at_index = next_nodes_and_keys
1115
# We should now be at the _LeafNodes
1116
node_indexes = [idx for idx, s_keys in keys_at_index]
1118
# TODO: We may *not* want to always read all the nodes in one
1119
# big go. Consider setting a max size on this.
1120
nodes = self._get_leaf_nodes(node_indexes)
1121
return nodes, keys_at_index
1123
def iter_entries(self, keys):
1124
"""Iterate over keys within the index.
1126
:param keys: An iterable providing the keys to be retrieved.
1127
:return: An iterable as per iter_all_entries, but restricted to the
1128
keys supplied. No additional keys will be returned, and every
1129
key supplied that is in the index will be returned.
1131
# 6 seconds spent in miss_torture using the sorted() line.
1132
# Even with out of order disk IO it seems faster not to sort it when
1133
# large queries are being made.
1134
# However, now that we are doing multi-way bisecting, we need the keys
1135
# in sorted order anyway. We could change the multi-way code to not
1136
# require sorted order. (For example, it bisects for the first node,
1137
# does an in-order search until a key comes before the current point,
1138
# which it then bisects for, etc.)
1139
keys = frozenset(keys)
1143
if not self.key_count():
1147
if self._leaf_value_cache is None:
1151
value = self._leaf_value_cache.get(key, None)
1152
if value is not None:
1153
# This key is known not to be here, skip it
1155
if self.node_ref_lists:
1156
yield (self, key, value, refs)
1158
yield (self, key, value)
1160
needed_keys.append(key)
1166
nodes, nodes_and_keys = self._walk_through_internal_nodes(needed_keys)
1167
for node_index, sub_keys in nodes_and_keys:
1170
node = nodes[node_index]
1171
for next_sub_key in sub_keys:
1172
if next_sub_key in node.keys:
1173
value, refs = node.keys[next_sub_key]
1174
if self.node_ref_lists:
1175
yield (self, next_sub_key, value, refs)
1177
yield (self, next_sub_key, value)
1179
def _find_ancestors(self, keys, ref_list_num, parent_map, missing_keys):
1180
"""Find the parent_map information for the set of keys.
1182
This populates the parent_map dict and missing_keys set based on the
1183
queried keys. It also can fill out an arbitrary number of parents that
1184
it finds while searching for the supplied keys.
1186
It is unlikely that you want to call this directly. See
1187
"CombinedGraphIndex.find_ancestry()" for a more appropriate API.
1189
:param keys: A keys whose ancestry we want to return
1190
Every key will either end up in 'parent_map' or 'missing_keys'.
1191
:param ref_list_num: This index in the ref_lists is the parents we
1193
:param parent_map: {key: parent_keys} for keys that are present in this
1194
index. This may contain more entries than were in 'keys', that are
1195
reachable ancestors of the keys requested.
1196
:param missing_keys: keys which are known to be missing in this index.
1197
This may include parents that were not directly requested, but we
1198
were able to determine that they are not present in this index.
1199
:return: search_keys parents that were found but not queried to know
1200
if they are missing or present. Callers can re-query this index for
1201
those keys, and they will be placed into parent_map or missing_keys
1203
if not self.key_count():
1204
# We use key_count() to trigger reading the root node and
1205
# determining info about this BTreeGraphIndex
1206
# If we don't have any keys, then everything is missing
1207
missing_keys.update(keys)
1209
if ref_list_num >= self.node_ref_lists:
1210
raise ValueError('No ref list %d, index has %d ref lists'
1211
% (ref_list_num, self.node_ref_lists))
1213
# The main trick we are trying to accomplish is that when we find a
1214
# key listing its parents, we expect that the parent key is also likely
1215
# to sit on the same page. Allowing us to expand parents quickly
1216
# without suffering the full stack of bisecting, etc.
1217
nodes, nodes_and_keys = self._walk_through_internal_nodes(keys)
1219
# These are parent keys which could not be immediately resolved on the
1220
# page where the child was present. Note that we may already be
1221
# searching for that key, and it may actually be present [or known
1222
# missing] on one of the other pages we are reading.
1224
# We could try searching for them in the immediate previous or next
1225
# page. If they occur "later" we could put them in a pending lookup
1226
# set, and then for each node we read thereafter we could check to
1227
# see if they are present.
1228
# However, we don't know the impact of keeping this list of things
1229
# that I'm going to search for every node I come across from here on
1231
# It doesn't handle the case when the parent key is missing on a
1232
# page that we *don't* read. So we already have to handle being
1233
# re-entrant for that.
1234
# Since most keys contain a date string, they are more likely to be
1235
# found earlier in the file than later, but we would know that right
1236
# away (key < min_key), and wouldn't keep searching it on every other
1237
# page that we read.
1238
# Mostly, it is an idea, one which should be benchmarked.
1239
parents_not_on_page = set()
1241
for node_index, sub_keys in nodes_and_keys:
1244
# sub_keys is all of the keys we are looking for that should exist
1245
# on this page, if they aren't here, then they won't be found
1246
node = nodes[node_index]
1247
node_keys = node.keys
1248
parents_to_check = set()
1249
for next_sub_key in sub_keys:
1250
if next_sub_key not in node_keys:
1251
# This one is just not present in the index at all
1252
missing_keys.add(next_sub_key)
1254
value, refs = node_keys[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:
1267
if key in node_keys:
1268
value, refs = node_keys[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 bzrlib) 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:
1370
raise errors.BadIndexKey(key)
1371
if len(key) != self._key_length:
1372
raise errors.BadIndexKey(key)
1374
if self.node_ref_lists:
1375
value, node_refs = nodes[key]
1376
yield self, key, value, node_refs
1378
yield self, key, nodes[key]
1385
raise errors.BadIndexKey(key)
1386
if len(key) != self._key_length:
1387
raise errors.BadIndexKey(key)
1388
# find what it refers to:
1389
key_dict = nodes_by_key
1390
elements = list(key)
1391
# find the subdict whose contents should be returned.
1393
while len(elements) and elements[0] is not None:
1394
key_dict = key_dict[elements[0]]
1397
# a non-existant lookup.
1402
key_dict = dicts.pop(-1)
1403
# can't be empty or would not exist
1404
item, value = key_dict.iteritems().next()
1405
if type(value) == dict:
1407
dicts.extend(key_dict.itervalues())
1410
for value in key_dict.itervalues():
1411
# each value is the key:value:node refs tuple
1413
yield (self, ) + value
1415
# the last thing looked up was a terminal element
1416
yield (self, ) + key_dict
1418
def key_count(self):
1419
"""Return an estimate of the number of keys in this index.
1421
For BTreeGraphIndex the estimate is exact as it is contained in the
1424
if self._key_count is None:
1425
self._get_root_node()
1426
return self._key_count
1428
def _compute_row_offsets(self):
1429
"""Fill out the _row_offsets attribute based on _row_lengths."""
1432
for row in self._row_lengths:
1433
offsets.append(row_offset)
1435
offsets.append(row_offset)
1436
self._row_offsets = offsets
1438
def _parse_header_from_bytes(self, bytes):
1439
"""Parse the header from a region of bytes.
1441
:param bytes: The data to parse.
1442
:return: An offset, data tuple such as readv yields, for the unparsed
1443
data. (which may be of length 0).
1445
signature = bytes[0:len(self._signature())]
1446
if not signature == self._signature():
1447
raise errors.BadIndexFormatSignature(self._name, BTreeGraphIndex)
1448
lines = bytes[len(self._signature()):].splitlines()
1449
options_line = lines[0]
1450
if not options_line.startswith(_OPTION_NODE_REFS):
1451
raise errors.BadIndexOptions(self)
1453
self.node_ref_lists = int(options_line[len(_OPTION_NODE_REFS):])
1455
raise errors.BadIndexOptions(self)
1456
options_line = lines[1]
1457
if not options_line.startswith(_OPTION_KEY_ELEMENTS):
1458
raise errors.BadIndexOptions(self)
1460
self._key_length = int(options_line[len(_OPTION_KEY_ELEMENTS):])
1462
raise errors.BadIndexOptions(self)
1463
options_line = lines[2]
1464
if not options_line.startswith(_OPTION_LEN):
1465
raise errors.BadIndexOptions(self)
1467
self._key_count = int(options_line[len(_OPTION_LEN):])
1469
raise errors.BadIndexOptions(self)
1470
options_line = lines[3]
1471
if not options_line.startswith(_OPTION_ROW_LENGTHS):
1472
raise errors.BadIndexOptions(self)
1474
self._row_lengths = map(int, [length for length in
1475
options_line[len(_OPTION_ROW_LENGTHS):].split(',')
1478
raise errors.BadIndexOptions(self)
1479
self._compute_row_offsets()
1481
# calculate the bytes we have processed
1482
header_end = (len(signature) + sum(map(len, lines[0:4])) + 4)
1483
return header_end, bytes[header_end:]
1485
def _read_nodes(self, nodes):
1486
"""Read some nodes from disk into the LRU cache.
1488
This performs a readv to get the node data into memory, and parses each
1489
node, then yields it to the caller. The nodes are requested in the
1490
supplied order. If possible doing sort() on the list before requesting
1491
a read may improve performance.
1493
:param nodes: The nodes to read. 0 - first node, 1 - second node etc.
1496
# may be the byte string of the whole file
1498
# list of (offset, length) regions of the file that should, evenually
1499
# be read in to data_ranges, either from 'bytes' or from the transport
1501
base_offset = self._base_offset
1503
offset = (index * _PAGE_SIZE)
1506
# Root node - special case
1508
size = min(_PAGE_SIZE, self._size)
1510
# The only case where we don't know the size, is for very
1511
# small indexes. So we read the whole thing
1512
bytes = self._transport.get_bytes(self._name)
1513
num_bytes = len(bytes)
1514
self._size = num_bytes - base_offset
1515
# the whole thing should be parsed out of 'bytes'
1516
ranges = [(start, min(_PAGE_SIZE, num_bytes - start))
1517
for start in xrange(base_offset, num_bytes, _PAGE_SIZE)]
1520
if offset > self._size:
1521
raise AssertionError('tried to read past the end'
1522
' of the file %s > %s'
1523
% (offset, self._size))
1524
size = min(size, self._size - offset)
1525
ranges.append((base_offset + offset, size))
1528
elif bytes is not None:
1529
# already have the whole file
1530
data_ranges = [(start, bytes[start:start+size])
1531
for start, size in ranges]
1532
elif self._file is None:
1533
data_ranges = self._transport.readv(self._name, ranges)
1536
for offset, size in ranges:
1537
self._file.seek(offset)
1538
data_ranges.append((offset, self._file.read(size)))
1539
for offset, data in data_ranges:
1540
offset -= base_offset
1542
# extract the header
1543
offset, data = self._parse_header_from_bytes(data)
1546
bytes = zlib.decompress(data)
1547
if bytes.startswith(_LEAF_FLAG):
1548
node = _LeafNode(bytes, self._key_length, self.node_ref_lists)
1549
elif bytes.startswith(_INTERNAL_FLAG):
1550
node = _InternalNode(bytes)
1552
raise AssertionError("Unknown node type for %r" % bytes)
1553
yield offset / _PAGE_SIZE, node
1555
def _signature(self):
1556
"""The file signature for this index type."""
1560
"""Validate that everything in the index can be accessed."""
1561
# just read and parse every node.
1562
self._get_root_node()
1563
if len(self._row_lengths) > 1:
1564
start_node = self._row_offsets[1]
1566
# We shouldn't be reading anything anyway
1568
node_end = self._row_offsets[-1]
1569
for node in self._read_nodes(range(start_node, node_end)):
1574
from bzrlib import _btree_serializer_pyx as _btree_serializer
1575
except ImportError, e:
1576
osutils.failed_to_load_extension(e)
1577
from bzrlib import _btree_serializer_py as _btree_serializer
added
removed
bzrlib/btree_index.py
