1
# Copyright (C) 2008 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 bisect import bisect_right
35
from bzrlib.index import _OPTION_NODE_REFS, _OPTION_KEY_ELEMENTS, _OPTION_LEN
36
from bzrlib.transport import get_transport
39
_BTSIGNATURE = "B+Tree Graph Index 2\n"
40
_OPTION_ROW_LENGTHS = "row_lengths="
41
_LEAF_FLAG = "type=leaf\n"
42
_INTERNAL_FLAG = "type=internal\n"
43
_INTERNAL_OFFSET = "offset="
45
_RESERVED_HEADER_BYTES = 120
48
# 4K per page: 4MB - 1000 entries
49
_NODE_CACHE_SIZE = 1000
52
class _BuilderRow(object):
53
"""The stored state accumulated while writing out a row in the index.
55
:ivar spool: A temporary file used to accumulate nodes for this row
57
:ivar nodes: The count of nodes emitted so far.
61
"""Create a _BuilderRow."""
63
self.spool = tempfile.TemporaryFile()
66
def finish_node(self, pad=True):
67
byte_lines, _, padding = self.writer.finish()
70
self.spool.write("\x00" * _RESERVED_HEADER_BYTES)
72
if not pad and padding:
74
skipped_bytes = padding
75
self.spool.writelines(byte_lines)
76
remainder = (self.spool.tell() + skipped_bytes) % _PAGE_SIZE
78
raise AssertionError("incorrect node length: %d, %d"
79
% (self.spool.tell(), remainder))
84
class _InternalBuilderRow(_BuilderRow):
85
"""The stored state accumulated while writing out internal rows."""
87
def finish_node(self, pad=True):
89
raise AssertionError("Must pad internal nodes only.")
90
_BuilderRow.finish_node(self)
93
class _LeafBuilderRow(_BuilderRow):
94
"""The stored state accumulated while writing out a leaf rows."""
97
class BTreeBuilder(index.GraphIndexBuilder):
98
"""A Builder for B+Tree based Graph indices.
100
The resulting graph has the structure:
102
_SIGNATURE OPTIONS NODES
103
_SIGNATURE := 'B+Tree Graph Index 1' NEWLINE
104
OPTIONS := REF_LISTS KEY_ELEMENTS LENGTH
105
REF_LISTS := 'node_ref_lists=' DIGITS NEWLINE
106
KEY_ELEMENTS := 'key_elements=' DIGITS NEWLINE
107
LENGTH := 'len=' DIGITS NEWLINE
108
ROW_LENGTHS := 'row_lengths' DIGITS (COMMA DIGITS)*
109
NODES := NODE_COMPRESSED*
110
NODE_COMPRESSED:= COMPRESSED_BYTES{4096}
111
NODE_RAW := INTERNAL | LEAF
112
INTERNAL := INTERNAL_FLAG POINTERS
113
LEAF := LEAF_FLAG ROWS
114
KEY_ELEMENT := Not-whitespace-utf8
115
KEY := KEY_ELEMENT (NULL KEY_ELEMENT)*
117
ROW := KEY NULL ABSENT? NULL REFERENCES NULL VALUE NEWLINE
119
REFERENCES := REFERENCE_LIST (TAB REFERENCE_LIST){node_ref_lists - 1}
120
REFERENCE_LIST := (REFERENCE (CR REFERENCE)*)?
122
VALUE := no-newline-no-null-bytes
125
def __init__(self, reference_lists=0, key_elements=1, spill_at=100000):
126
"""See GraphIndexBuilder.__init__.
128
:param spill_at: Optional parameter controlling the maximum number
129
of nodes that BTreeBuilder will hold in memory.
131
index.GraphIndexBuilder.__init__(self, reference_lists=reference_lists,
132
key_elements=key_elements)
133
self._spill_at = spill_at
134
self._backing_indices = []
135
# A map of {key: (node_refs, value)}
137
# Indicate it hasn't been built yet
138
self._nodes_by_key = None
139
self._optimize_for_size = False
141
def add_node(self, key, value, references=()):
142
"""Add a node to the index.
144
If adding the node causes the builder to reach its spill_at threshold,
145
disk spilling will be triggered.
147
:param key: The key. keys are non-empty tuples containing
148
as many whitespace-free utf8 bytestrings as the key length
149
defined for this index.
150
:param references: An iterable of iterables of keys. Each is a
151
reference to another key.
152
:param value: The value to associate with the key. It may be any
153
bytes as long as it does not contain \0 or \n.
155
# we don't care about absent_references
156
node_refs, _ = self._check_key_ref_value(key, references, value)
157
if key in self._nodes:
158
raise errors.BadIndexDuplicateKey(key, self)
159
self._nodes[key] = (node_refs, value)
161
if self._nodes_by_key is not None and self._key_length > 1:
162
self._update_nodes_by_key(key, value, node_refs)
163
if len(self._keys) < self._spill_at:
165
self._spill_mem_keys_to_disk()
167
def _spill_mem_keys_to_disk(self):
168
"""Write the in memory keys down to disk to cap memory consumption.
170
If we already have some keys written to disk, we will combine them so
171
as to preserve the sorted order. The algorithm for combining uses
172
powers of two. So on the first spill, write all mem nodes into a
173
single index. On the second spill, combine the mem nodes with the nodes
174
on disk to create a 2x sized disk index and get rid of the first index.
175
On the third spill, create a single new disk index, which will contain
176
the mem nodes, and preserve the existing 2x sized index. On the fourth,
177
combine mem with the first and second indexes, creating a new one of
178
size 4x. On the fifth create a single new one, etc.
180
if self._combine_backing_indices:
181
(new_backing_file, size,
182
backing_pos) = self._spill_mem_keys_and_combine()
184
new_backing_file, size = self._spill_mem_keys_without_combining()
185
dir_path, base_name = osutils.split(new_backing_file.name)
186
# Note: The transport here isn't strictly needed, because we will use
187
# direct access to the new_backing._file object
188
new_backing = BTreeGraphIndex(get_transport(dir_path),
190
# GC will clean up the file
191
new_backing._file = new_backing_file
192
if self._combine_backing_indices:
193
if len(self._backing_indices) == backing_pos:
194
self._backing_indices.append(None)
195
self._backing_indices[backing_pos] = new_backing
196
for backing_pos in range(backing_pos):
197
self._backing_indices[backing_pos] = None
199
self._backing_indices.append(new_backing)
202
self._nodes_by_key = None
204
def _spill_mem_keys_without_combining(self):
205
return self._write_nodes(self._iter_mem_nodes(), allow_optimize=False)
207
def _spill_mem_keys_and_combine(self):
208
iterators_to_combine = [self._iter_mem_nodes()]
210
for pos, backing in enumerate(self._backing_indices):
214
iterators_to_combine.append(backing.iter_all_entries())
215
backing_pos = pos + 1
216
new_backing_file, size = \
217
self._write_nodes(self._iter_smallest(iterators_to_combine),
218
allow_optimize=False)
219
return new_backing_file, size, backing_pos
221
def add_nodes(self, nodes):
222
"""Add nodes to the index.
224
:param nodes: An iterable of (key, node_refs, value) entries to add.
226
if self.reference_lists:
227
for (key, value, node_refs) in nodes:
228
self.add_node(key, value, node_refs)
230
for (key, value) in nodes:
231
self.add_node(key, value)
233
def _iter_mem_nodes(self):
234
"""Iterate over the nodes held in memory."""
236
if self.reference_lists:
237
for key in sorted(nodes):
238
references, value = nodes[key]
239
yield self, key, value, references
241
for key in sorted(nodes):
242
references, value = nodes[key]
243
yield self, key, value
245
def _iter_smallest(self, iterators_to_combine):
246
if len(iterators_to_combine) == 1:
247
for value in iterators_to_combine[0]:
251
for iterator in iterators_to_combine:
253
current_values.append(iterator.next())
254
except StopIteration:
255
current_values.append(None)
258
# Decorate candidates with the value to allow 2.4's min to be used.
259
candidates = [(item[1][1], item) for item
260
in enumerate(current_values) if item[1] is not None]
261
if not len(candidates):
263
selected = min(candidates)
264
# undecorate back to (pos, node)
265
selected = selected[1]
266
if last == selected[1][1]:
267
raise errors.BadIndexDuplicateKey(last, self)
268
last = selected[1][1]
269
# Yield, with self as the index
270
yield (self,) + selected[1][1:]
273
current_values[pos] = iterators_to_combine[pos].next()
274
except StopIteration:
275
current_values[pos] = None
277
def _add_key(self, string_key, line, rows, allow_optimize=True):
278
"""Add a key to the current chunk.
280
:param string_key: The key to add.
281
:param line: The fully serialised key and value.
282
:param allow_optimize: If set to False, prevent setting the optimize
283
flag when writing out. This is used by the _spill_mem_keys_to_disk
286
if rows[-1].writer is None:
287
# opening a new leaf chunk;
288
for pos, internal_row in enumerate(rows[:-1]):
289
# flesh out any internal nodes that are needed to
290
# preserve the height of the tree
291
if internal_row.writer is None:
293
if internal_row.nodes == 0:
294
length -= _RESERVED_HEADER_BYTES # padded
296
optimize_for_size = self._optimize_for_size
298
optimize_for_size = False
299
internal_row.writer = chunk_writer.ChunkWriter(length, 0,
300
optimize_for_size=optimize_for_size)
301
internal_row.writer.write(_INTERNAL_FLAG)
302
internal_row.writer.write(_INTERNAL_OFFSET +
303
str(rows[pos + 1].nodes) + "\n")
306
if rows[-1].nodes == 0:
307
length -= _RESERVED_HEADER_BYTES # padded
308
rows[-1].writer = chunk_writer.ChunkWriter(length,
309
optimize_for_size=self._optimize_for_size)
310
rows[-1].writer.write(_LEAF_FLAG)
311
if rows[-1].writer.write(line):
312
# this key did not fit in the node:
313
rows[-1].finish_node()
314
key_line = string_key + "\n"
316
for row in reversed(rows[:-1]):
317
# Mark the start of the next node in the node above. If it
318
# doesn't fit then propagate upwards until we find one that
320
if row.writer.write(key_line):
323
# We've found a node that can handle the pointer.
326
# If we reached the current root without being able to mark the
327
# division point, then we need a new root:
330
if 'index' in debug.debug_flags:
331
trace.mutter('Inserting new global row.')
332
new_row = _InternalBuilderRow()
334
rows.insert(0, new_row)
335
# This will be padded, hence the -100
336
new_row.writer = chunk_writer.ChunkWriter(
337
_PAGE_SIZE - _RESERVED_HEADER_BYTES,
339
optimize_for_size=self._optimize_for_size)
340
new_row.writer.write(_INTERNAL_FLAG)
341
new_row.writer.write(_INTERNAL_OFFSET +
342
str(rows[1].nodes - 1) + "\n")
343
new_row.writer.write(key_line)
344
self._add_key(string_key, line, rows, allow_optimize=allow_optimize)
346
def _write_nodes(self, node_iterator, allow_optimize=True):
347
"""Write node_iterator out as a B+Tree.
349
:param node_iterator: An iterator of sorted nodes. Each node should
350
match the output given by iter_all_entries.
351
:param allow_optimize: If set to False, prevent setting the optimize
352
flag when writing out. This is used by the _spill_mem_keys_to_disk
354
:return: A file handle for a temporary file containing a B+Tree for
357
# The index rows - rows[0] is the root, rows[1] is the layer under it
360
# forward sorted by key. In future we may consider topological sorting,
361
# at the cost of table scans for direct lookup, or a second index for
364
# A stack with the number of nodes of each size. 0 is the root node
365
# and must always be 1 (if there are any nodes in the tree).
366
self.row_lengths = []
367
# Loop over all nodes adding them to the bottom row
368
# (rows[-1]). When we finish a chunk in a row,
369
# propagate the key that didn't fit (comes after the chunk) to the
370
# row above, transitively.
371
for node in node_iterator:
373
# First key triggers the first row
374
rows.append(_LeafBuilderRow())
376
string_key, line = _btree_serializer._flatten_node(node,
377
self.reference_lists)
378
self._add_key(string_key, line, rows, allow_optimize=allow_optimize)
379
for row in reversed(rows):
380
pad = (type(row) != _LeafBuilderRow)
381
row.finish_node(pad=pad)
382
result = tempfile.NamedTemporaryFile(prefix='bzr-index-')
383
lines = [_BTSIGNATURE]
384
lines.append(_OPTION_NODE_REFS + str(self.reference_lists) + '\n')
385
lines.append(_OPTION_KEY_ELEMENTS + str(self._key_length) + '\n')
386
lines.append(_OPTION_LEN + str(key_count) + '\n')
387
row_lengths = [row.nodes for row in rows]
388
lines.append(_OPTION_ROW_LENGTHS + ','.join(map(str, row_lengths)) + '\n')
389
result.writelines(lines)
390
position = sum(map(len, lines))
392
if position > _RESERVED_HEADER_BYTES:
393
raise AssertionError("Could not fit the header in the"
394
" reserved space: %d > %d"
395
% (position, _RESERVED_HEADER_BYTES))
396
# write the rows out:
398
reserved = _RESERVED_HEADER_BYTES # reserved space for first node
401
# copy nodes to the finalised file.
402
# Special case the first node as it may be prefixed
403
node = row.spool.read(_PAGE_SIZE)
404
result.write(node[reserved:])
405
result.write("\x00" * (reserved - position))
406
position = 0 # Only the root row actually has an offset
407
copied_len = osutils.pumpfile(row.spool, result)
408
if copied_len != (row.nodes - 1) * _PAGE_SIZE:
409
if type(row) != _LeafBuilderRow:
410
raise AssertionError("Incorrect amount of data copied"
411
" expected: %d, got: %d"
412
% ((row.nodes - 1) * _PAGE_SIZE,
420
"""Finalise the index.
422
:return: A file handle for a temporary file containing the nodes added
425
return self._write_nodes(self.iter_all_entries())[0]
427
def iter_all_entries(self):
428
"""Iterate over all keys within the index
430
:return: An iterable of (index, key, value, reference_lists). There is
431
no defined order for the result iteration - it will be in the most
432
efficient order for the index (in this case dictionary hash order).
434
if 'evil' in debug.debug_flags:
435
trace.mutter_callsite(3,
436
"iter_all_entries scales with size of history.")
437
# Doing serial rather than ordered would be faster; but this shouldn't
438
# be getting called routinely anyway.
439
iterators = [self._iter_mem_nodes()]
440
for backing in self._backing_indices:
441
if backing is not None:
442
iterators.append(backing.iter_all_entries())
443
if len(iterators) == 1:
445
return self._iter_smallest(iterators)
447
def iter_entries(self, keys):
448
"""Iterate over keys within the index.
450
:param keys: An iterable providing the keys to be retrieved.
451
:return: An iterable of (index, key, value, reference_lists). There is no
452
defined order for the result iteration - it will be in the most
453
efficient order for the index (keys iteration order in this case).
456
local_keys = keys.intersection(self._keys)
457
if self.reference_lists:
458
for key in local_keys:
459
node = self._nodes[key]
460
yield self, key, node[1], node[0]
462
for key in local_keys:
463
node = self._nodes[key]
464
yield self, key, node[1]
465
# Find things that are in backing indices that have not been handled
467
if not self._backing_indices:
468
return # We won't find anything there either
469
# Remove all of the keys that we found locally
470
keys.difference_update(local_keys)
471
for backing in self._backing_indices:
476
for node in backing.iter_entries(keys):
478
yield (self,) + node[1:]
480
def iter_entries_prefix(self, keys):
481
"""Iterate over keys within the index using prefix matching.
483
Prefix matching is applied within the tuple of a key, not to within
484
the bytestring of each key element. e.g. if you have the keys ('foo',
485
'bar'), ('foobar', 'gam') and do a prefix search for ('foo', None) then
486
only the former key is returned.
488
:param keys: An iterable providing the key prefixes to be retrieved.
489
Each key prefix takes the form of a tuple the length of a key, but
490
with the last N elements 'None' rather than a regular bytestring.
491
The first element cannot be 'None'.
492
:return: An iterable as per iter_all_entries, but restricted to the
493
keys with a matching prefix to those supplied. No additional keys
494
will be returned, and every match that is in the index will be
497
# XXX: To much duplication with the GraphIndex class; consider finding
498
# a good place to pull out the actual common logic.
502
for backing in self._backing_indices:
505
for node in backing.iter_entries_prefix(keys):
506
yield (self,) + node[1:]
507
if self._key_length == 1:
511
raise errors.BadIndexKey(key)
512
if len(key) != self._key_length:
513
raise errors.BadIndexKey(key)
515
node = self._nodes[key]
518
if self.reference_lists:
519
yield self, key, node[1], node[0]
521
yield self, key, node[1]
526
raise errors.BadIndexKey(key)
527
if len(key) != self._key_length:
528
raise errors.BadIndexKey(key)
529
# find what it refers to:
530
key_dict = self._get_nodes_by_key()
532
# find the subdict to return
534
while len(elements) and elements[0] is not None:
535
key_dict = key_dict[elements[0]]
538
# a non-existant lookup.
543
key_dict = dicts.pop(-1)
544
# can't be empty or would not exist
545
item, value = key_dict.iteritems().next()
546
if type(value) == dict:
548
dicts.extend(key_dict.itervalues())
551
for value in key_dict.itervalues():
552
yield (self, ) + value
554
yield (self, ) + key_dict
556
def _get_nodes_by_key(self):
557
if self._nodes_by_key is None:
559
if self.reference_lists:
560
for key, (references, value) in self._nodes.iteritems():
561
key_dict = nodes_by_key
562
for subkey in key[:-1]:
563
key_dict = key_dict.setdefault(subkey, {})
564
key_dict[key[-1]] = key, value, references
566
for key, (references, value) in self._nodes.iteritems():
567
key_dict = nodes_by_key
568
for subkey in key[:-1]:
569
key_dict = key_dict.setdefault(subkey, {})
570
key_dict[key[-1]] = key, value
571
self._nodes_by_key = nodes_by_key
572
return self._nodes_by_key
575
"""Return an estimate of the number of keys in this index.
577
For InMemoryGraphIndex the estimate is exact.
579
return len(self._keys) + sum(backing.key_count() for backing in
580
self._backing_indices if backing is not None)
583
"""In memory index's have no known corruption at the moment."""
586
class _LeafNode(object):
587
"""A leaf node for a serialised B+Tree index."""
589
__slots__ = ('keys',)
591
def __init__(self, bytes, key_length, ref_list_length):
592
"""Parse bytes to create a leaf node object."""
593
# splitlines mangles the \r delimiters.. don't use it.
594
self.keys = dict(_btree_serializer._parse_leaf_lines(bytes,
595
key_length, ref_list_length))
598
class _InternalNode(object):
599
"""An internal node for a serialised B+Tree index."""
601
__slots__ = ('keys', 'offset')
603
def __init__(self, bytes):
604
"""Parse bytes to create an internal node object."""
605
# splitlines mangles the \r delimiters.. don't use it.
606
self.keys = self._parse_lines(bytes.split('\n'))
608
def _parse_lines(self, lines):
610
self.offset = int(lines[1][7:])
611
for line in lines[2:]:
614
nodes.append(tuple(map(intern, line.split('\0'))))
618
class BTreeGraphIndex(object):
619
"""Access to nodes via the standard GraphIndex interface for B+Tree's.
621
Individual nodes are held in a LRU cache. This holds the root node in
622
memory except when very large walks are done.
625
def __init__(self, transport, name, size):
626
"""Create a B+Tree index object on the index name.
628
:param transport: The transport to read data for the index from.
629
:param name: The file name of the index on transport.
630
:param size: Optional size of the index in bytes. This allows
631
compatibility with the GraphIndex API, as well as ensuring that
632
the initial read (to read the root node header) can be done
633
without over-reading even on empty indices, and on small indices
634
allows single-IO to read the entire index.
636
self._transport = transport
640
self._recommended_pages = self._compute_recommended_pages()
641
self._root_node = None
642
# Default max size is 100,000 leave values
643
self._leaf_value_cache = None # lru_cache.LRUCache(100*1000)
644
self._leaf_node_cache = lru_cache.LRUCache(_NODE_CACHE_SIZE)
645
# We could limit this, but even a 300k record btree has only 3k leaf
646
# nodes, and only 20 internal nodes. So the default of 100 nodes in an
647
# LRU would mean we always cache everything anyway, no need to pay the
649
self._internal_node_cache = fifo_cache.FIFOCache(100)
650
self._key_count = None
651
self._row_lengths = None
652
self._row_offsets = None # Start of each row, [-1] is the end
654
def __eq__(self, other):
655
"""Equal when self and other were created with the same parameters."""
657
type(self) == type(other) and
658
self._transport == other._transport and
659
self._name == other._name and
660
self._size == other._size)
662
def __ne__(self, other):
663
return not self.__eq__(other)
665
def _get_and_cache_nodes(self, nodes):
666
"""Read nodes and cache them in the lru.
668
The nodes list supplied is sorted and then read from disk, each node
669
being inserted it into the _node_cache.
671
Note: Asking for more nodes than the _node_cache can contain will
672
result in some of the results being immediately discarded, to prevent
673
this an assertion is raised if more nodes are asked for than are
676
:return: A dict of {node_pos: node}
679
start_of_leaves = None
680
for node_pos, node in self._read_nodes(sorted(nodes)):
681
if node_pos == 0: # Special case
682
self._root_node = node
684
if start_of_leaves is None:
685
start_of_leaves = self._row_offsets[-2]
686
if node_pos < start_of_leaves:
687
self._internal_node_cache.add(node_pos, node)
689
self._leaf_node_cache.add(node_pos, node)
690
found[node_pos] = node
693
def _compute_recommended_pages(self):
694
"""Convert transport's recommended_page_size into btree pages.
696
recommended_page_size is in bytes, we want to know how many _PAGE_SIZE
697
pages fit in that length.
699
recommended_read = self._transport.recommended_page_size()
700
recommended_pages = int(math.ceil(recommended_read /
702
return recommended_pages
704
def _compute_total_pages_in_index(self):
705
"""How many pages are in the index.
707
If we have read the header we will use the value stored there.
708
Otherwise it will be computed based on the length of the index.
710
if self._size is None:
711
raise AssertionError('_compute_total_pages_in_index should not be'
712
' called when self._size is None')
713
if self._root_node is not None:
714
# This is the number of pages as defined by the header
715
return self._row_offsets[-1]
716
# This is the number of pages as defined by the size of the index. They
717
# should be indentical.
718
total_pages = int(math.ceil(self._size / float(_PAGE_SIZE)))
721
def _expand_offsets(self, offsets):
722
"""Find extra pages to download.
724
The idea is that we always want to make big-enough requests (like 64kB
725
for http), so that we don't waste round trips. So given the entries
726
that we already have cached and the new pages being downloaded figure
727
out what other pages we might want to read.
729
See also doc/developers/btree_index_prefetch.txt for more details.
731
:param offsets: The offsets to be read
732
:return: A list of offsets to download
734
if 'index' in debug.debug_flags:
735
trace.mutter('expanding: %s\toffsets: %s', self._name, offsets)
737
if len(offsets) >= self._recommended_pages:
738
# Don't add more, we are already requesting more than enough
739
if 'index' in debug.debug_flags:
740
trace.mutter(' not expanding large request (%s >= %s)',
741
len(offsets), self._recommended_pages)
743
if self._size is None:
744
# Don't try anything, because we don't know where the file ends
745
if 'index' in debug.debug_flags:
746
trace.mutter(' not expanding without knowing index size')
748
total_pages = self._compute_total_pages_in_index()
749
cached_offsets = self._get_offsets_to_cached_pages()
750
# If reading recommended_pages would read the rest of the index, just
752
if total_pages - len(cached_offsets) <= self._recommended_pages:
753
# Read whatever is left
755
expanded = [x for x in xrange(total_pages)
756
if x not in cached_offsets]
758
expanded = range(total_pages)
759
if 'index' in debug.debug_flags:
760
trace.mutter(' reading all unread pages: %s', expanded)
763
if self._root_node is None:
764
# ATM on the first read of the root node of a large index, we don't
765
# bother pre-reading any other pages. This is because the
766
# likelyhood of actually reading interesting pages is very low.
767
# See doc/developers/btree_index_prefetch.txt for a discussion, and
768
# a possible implementation when we are guessing that the second
769
# layer index is small
770
final_offsets = offsets
772
tree_depth = len(self._row_lengths)
773
if len(cached_offsets) < tree_depth and len(offsets) == 1:
774
# We haven't read enough to justify expansion
775
# If we are only going to read the root node, and 1 leaf node,
776
# then it isn't worth expanding our request. Once we've read at
777
# least 2 nodes, then we are probably doing a search, and we
778
# start expanding our requests.
779
if 'index' in debug.debug_flags:
780
trace.mutter(' not expanding on first reads')
782
final_offsets = self._expand_to_neighbors(offsets, cached_offsets,
785
final_offsets = sorted(final_offsets)
786
if 'index' in debug.debug_flags:
787
trace.mutter('expanded: %s', final_offsets)
790
def _expand_to_neighbors(self, offsets, cached_offsets, total_pages):
791
"""Expand requests to neighbors until we have enough pages.
793
This is called from _expand_offsets after policy has determined that we
795
We only want to expand requests within a given layer. We cheat a little
796
bit and assume all requests will be in the same layer. This is true
797
given the current design, but if it changes this algorithm may perform
800
:param offsets: requested offsets
801
:param cached_offsets: offsets for pages we currently have cached
802
:return: A set() of offsets after expansion
804
final_offsets = set(offsets)
806
new_tips = set(final_offsets)
807
while len(final_offsets) < self._recommended_pages and new_tips:
811
first, end = self._find_layer_first_and_end(pos)
814
and previous not in cached_offsets
815
and previous not in final_offsets
816
and previous >= first):
817
next_tips.add(previous)
819
if (after < total_pages
820
and after not in cached_offsets
821
and after not in final_offsets
824
# This would keep us from going bigger than
825
# recommended_pages by only expanding the first offsets.
826
# However, if we are making a 'wide' request, it is
827
# reasonable to expand all points equally.
828
# if len(final_offsets) > recommended_pages:
830
final_offsets.update(next_tips)
834
def external_references(self, ref_list_num):
835
if self._root_node is None:
836
self._get_root_node()
837
if ref_list_num + 1 > self.node_ref_lists:
838
raise ValueError('No ref list %d, index has %d ref lists'
839
% (ref_list_num, self.node_ref_lists))
842
for node in self.iter_all_entries():
844
refs.update(node[3][ref_list_num])
847
def _find_layer_first_and_end(self, offset):
848
"""Find the start/stop nodes for the layer corresponding to offset.
850
:return: (first, end)
851
first is the first node in this layer
852
end is the first node of the next layer
855
for roffset in self._row_offsets:
862
def _get_offsets_to_cached_pages(self):
863
"""Determine what nodes we already have cached."""
864
cached_offsets = set(self._internal_node_cache.keys())
865
cached_offsets.update(self._leaf_node_cache.keys())
866
if self._root_node is not None:
867
cached_offsets.add(0)
868
return cached_offsets
870
def _get_root_node(self):
871
if self._root_node is None:
872
# We may not have a root node yet
873
self._get_internal_nodes([0])
874
return self._root_node
876
def _get_nodes(self, cache, node_indexes):
879
for idx in node_indexes:
880
if idx == 0 and self._root_node is not None:
881
found[0] = self._root_node
884
found[idx] = cache[idx]
889
needed = self._expand_offsets(needed)
890
found.update(self._get_and_cache_nodes(needed))
893
def _get_internal_nodes(self, node_indexes):
894
"""Get a node, from cache or disk.
896
After getting it, the node will be cached.
898
return self._get_nodes(self._internal_node_cache, node_indexes)
900
def _cache_leaf_values(self, nodes):
901
"""Cache directly from key => value, skipping the btree."""
902
if self._leaf_value_cache is not None:
903
for node in nodes.itervalues():
904
for key, value in node.keys.iteritems():
905
if key in self._leaf_value_cache:
906
# Don't add the rest of the keys, we've seen this node
909
self._leaf_value_cache[key] = value
911
def _get_leaf_nodes(self, node_indexes):
912
"""Get a bunch of nodes, from cache or disk."""
913
found = self._get_nodes(self._leaf_node_cache, node_indexes)
914
self._cache_leaf_values(found)
917
def iter_all_entries(self):
918
"""Iterate over all keys within the index.
920
:return: An iterable of (index, key, value) or (index, key, value, reference_lists).
921
The former tuple is used when there are no reference lists in the
922
index, making the API compatible with simple key:value index types.
923
There is no defined order for the result iteration - it will be in
924
the most efficient order for the index.
926
if 'evil' in debug.debug_flags:
927
trace.mutter_callsite(3,
928
"iter_all_entries scales with size of history.")
929
if not self.key_count():
931
if self._row_offsets[-1] == 1:
932
# There is only the root node, and we read that via key_count()
933
if self.node_ref_lists:
934
for key, (value, refs) in sorted(self._root_node.keys.items()):
935
yield (self, key, value, refs)
937
for key, (value, refs) in sorted(self._root_node.keys.items()):
938
yield (self, key, value)
940
start_of_leaves = self._row_offsets[-2]
941
end_of_leaves = self._row_offsets[-1]
942
needed_offsets = range(start_of_leaves, end_of_leaves)
943
if needed_offsets == [0]:
944
# Special case when we only have a root node, as we have already
946
nodes = [(0, self._root_node)]
948
nodes = self._read_nodes(needed_offsets)
949
# We iterate strictly in-order so that we can use this function
950
# for spilling index builds to disk.
951
if self.node_ref_lists:
952
for _, node in nodes:
953
for key, (value, refs) in sorted(node.keys.items()):
954
yield (self, key, value, refs)
956
for _, node in nodes:
957
for key, (value, refs) in sorted(node.keys.items()):
958
yield (self, key, value)
961
def _multi_bisect_right(in_keys, fixed_keys):
962
"""Find the positions where each 'in_key' would fit in fixed_keys.
964
This is equivalent to doing "bisect_right" on each in_key into
967
:param in_keys: A sorted list of keys to match with fixed_keys
968
:param fixed_keys: A sorted list of keys to match against
969
:return: A list of (integer position, [key list]) tuples.
974
# no pointers in the fixed_keys list, which means everything must
976
return [(0, in_keys)]
978
# TODO: Iterating both lists will generally take M + N steps
979
# Bisecting each key will generally take M * log2 N steps.
980
# If we had an efficient way to compare, we could pick the method
981
# based on which has the fewer number of steps.
982
# There is also the argument that bisect_right is a compiled
983
# function, so there is even more to be gained.
984
# iter_steps = len(in_keys) + len(fixed_keys)
985
# bisect_steps = len(in_keys) * math.log(len(fixed_keys), 2)
986
if len(in_keys) == 1: # Bisect will always be faster for M = 1
987
return [(bisect_right(fixed_keys, in_keys[0]), in_keys)]
988
# elif bisect_steps < iter_steps:
990
# for key in in_keys:
991
# offsets.setdefault(bisect_right(fixed_keys, key),
993
# return [(o, offsets[o]) for o in sorted(offsets)]
994
in_keys_iter = iter(in_keys)
995
fixed_keys_iter = enumerate(fixed_keys)
996
cur_in_key = in_keys_iter.next()
997
cur_fixed_offset, cur_fixed_key = fixed_keys_iter.next()
999
class InputDone(Exception): pass
1000
class FixedDone(Exception): pass
1005
# TODO: Another possibility is that rather than iterating on each side,
1006
# we could use a combination of bisecting and iterating. For
1007
# example, while cur_in_key < fixed_key, bisect to find its
1008
# point, then iterate all matching keys, then bisect (restricted
1009
# to only the remainder) for the next one, etc.
1012
if cur_in_key < cur_fixed_key:
1014
cur_out = (cur_fixed_offset, cur_keys)
1015
output.append(cur_out)
1016
while cur_in_key < cur_fixed_key:
1017
cur_keys.append(cur_in_key)
1019
cur_in_key = in_keys_iter.next()
1020
except StopIteration:
1022
# At this point cur_in_key must be >= cur_fixed_key
1023
# step the cur_fixed_key until we pass the cur key, or walk off
1025
while cur_in_key >= cur_fixed_key:
1027
cur_fixed_offset, cur_fixed_key = fixed_keys_iter.next()
1028
except StopIteration:
1031
# We consumed all of the input, nothing more to do
1034
# There was some input left, but we consumed all of fixed, so we
1035
# have to add one more for the tail
1036
cur_keys = [cur_in_key]
1037
cur_keys.extend(in_keys_iter)
1038
cur_out = (len(fixed_keys), cur_keys)
1039
output.append(cur_out)
1042
def iter_entries(self, keys):
1043
"""Iterate over keys within the index.
1045
:param keys: An iterable providing the keys to be retrieved.
1046
:return: An iterable as per iter_all_entries, but restricted to the
1047
keys supplied. No additional keys will be returned, and every
1048
key supplied that is in the index will be returned.
1050
# 6 seconds spent in miss_torture using the sorted() line.
1051
# Even with out of order disk IO it seems faster not to sort it when
1052
# large queries are being made.
1053
# However, now that we are doing multi-way bisecting, we need the keys
1054
# in sorted order anyway. We could change the multi-way code to not
1055
# require sorted order. (For example, it bisects for the first node,
1056
# does an in-order search until a key comes before the current point,
1057
# which it then bisects for, etc.)
1058
keys = frozenset(keys)
1062
if not self.key_count():
1066
if self._leaf_value_cache is None:
1070
value = self._leaf_value_cache.get(key, None)
1071
if value is not None:
1072
# This key is known not to be here, skip it
1074
if self.node_ref_lists:
1075
yield (self, key, value, refs)
1077
yield (self, key, value)
1079
needed_keys.append(key)
1085
# 6 seconds spent in miss_torture using the sorted() line.
1086
# Even with out of order disk IO it seems faster not to sort it when
1087
# large queries are being made.
1088
needed_keys = sorted(needed_keys)
1090
nodes_and_keys = [(0, needed_keys)]
1092
for row_pos, next_row_start in enumerate(self._row_offsets[1:-1]):
1093
node_indexes = [idx for idx, s_keys in nodes_and_keys]
1094
nodes = self._get_internal_nodes(node_indexes)
1096
next_nodes_and_keys = []
1097
for node_index, sub_keys in nodes_and_keys:
1098
node = nodes[node_index]
1099
positions = self._multi_bisect_right(sub_keys, node.keys)
1100
node_offset = next_row_start + node.offset
1101
next_nodes_and_keys.extend([(node_offset + pos, s_keys)
1102
for pos, s_keys in positions])
1103
nodes_and_keys = next_nodes_and_keys
1104
# We should now be at the _LeafNodes
1105
node_indexes = [idx for idx, s_keys in nodes_and_keys]
1107
# TODO: We may *not* want to always read all the nodes in one
1108
# big go. Consider setting a max size on this.
1110
nodes = self._get_leaf_nodes(node_indexes)
1111
for node_index, sub_keys in nodes_and_keys:
1114
node = nodes[node_index]
1115
for next_sub_key in sub_keys:
1116
if next_sub_key in node.keys:
1117
value, refs = node.keys[next_sub_key]
1118
if self.node_ref_lists:
1119
yield (self, next_sub_key, value, refs)
1121
yield (self, next_sub_key, value)
1123
def iter_entries_prefix(self, keys):
1124
"""Iterate over keys within the index using prefix matching.
1126
Prefix matching is applied within the tuple of a key, not to within
1127
the bytestring of each key element. e.g. if you have the keys ('foo',
1128
'bar'), ('foobar', 'gam') and do a prefix search for ('foo', None) then
1129
only the former key is returned.
1131
WARNING: Note that this method currently causes a full index parse
1132
unconditionally (which is reasonably appropriate as it is a means for
1133
thunking many small indices into one larger one and still supplies
1134
iter_all_entries at the thunk layer).
1136
:param keys: An iterable providing the key prefixes to be retrieved.
1137
Each key prefix takes the form of a tuple the length of a key, but
1138
with the last N elements 'None' rather than a regular bytestring.
1139
The first element cannot be 'None'.
1140
:return: An iterable as per iter_all_entries, but restricted to the
1141
keys with a matching prefix to those supplied. No additional keys
1142
will be returned, and every match that is in the index will be
1145
keys = sorted(set(keys))
1148
# Load if needed to check key lengths
1149
if self._key_count is None:
1150
self._get_root_node()
1151
# TODO: only access nodes that can satisfy the prefixes we are looking
1152
# for. For now, to meet API usage (as this function is not used by
1153
# current bzrlib) just suck the entire index and iterate in memory.
1155
if self.node_ref_lists:
1156
if self._key_length == 1:
1157
for _1, key, value, refs in self.iter_all_entries():
1158
nodes[key] = value, refs
1161
for _1, key, value, refs in self.iter_all_entries():
1162
key_value = key, value, refs
1163
# For a key of (foo, bar, baz) create
1164
# _nodes_by_key[foo][bar][baz] = key_value
1165
key_dict = nodes_by_key
1166
for subkey in key[:-1]:
1167
key_dict = key_dict.setdefault(subkey, {})
1168
key_dict[key[-1]] = key_value
1170
if self._key_length == 1:
1171
for _1, key, value in self.iter_all_entries():
1175
for _1, key, value in self.iter_all_entries():
1176
key_value = key, value
1177
# For a key of (foo, bar, baz) create
1178
# _nodes_by_key[foo][bar][baz] = key_value
1179
key_dict = nodes_by_key
1180
for subkey in key[:-1]:
1181
key_dict = key_dict.setdefault(subkey, {})
1182
key_dict[key[-1]] = key_value
1183
if self._key_length == 1:
1187
raise errors.BadIndexKey(key)
1188
if len(key) != self._key_length:
1189
raise errors.BadIndexKey(key)
1191
if self.node_ref_lists:
1192
value, node_refs = nodes[key]
1193
yield self, key, value, node_refs
1195
yield self, key, nodes[key]
1202
raise errors.BadIndexKey(key)
1203
if len(key) != self._key_length:
1204
raise errors.BadIndexKey(key)
1205
# find what it refers to:
1206
key_dict = nodes_by_key
1207
elements = list(key)
1208
# find the subdict whose contents should be returned.
1210
while len(elements) and elements[0] is not None:
1211
key_dict = key_dict[elements[0]]
1214
# a non-existant lookup.
1219
key_dict = dicts.pop(-1)
1220
# can't be empty or would not exist
1221
item, value = key_dict.iteritems().next()
1222
if type(value) == dict:
1224
dicts.extend(key_dict.itervalues())
1227
for value in key_dict.itervalues():
1228
# each value is the key:value:node refs tuple
1230
yield (self, ) + value
1232
# the last thing looked up was a terminal element
1233
yield (self, ) + key_dict
1235
def key_count(self):
1236
"""Return an estimate of the number of keys in this index.
1238
For BTreeGraphIndex the estimate is exact as it is contained in the
1241
if self._key_count is None:
1242
self._get_root_node()
1243
return self._key_count
1245
def _compute_row_offsets(self):
1246
"""Fill out the _row_offsets attribute based on _row_lengths."""
1249
for row in self._row_lengths:
1250
offsets.append(row_offset)
1252
offsets.append(row_offset)
1253
self._row_offsets = offsets
1255
def _parse_header_from_bytes(self, bytes):
1256
"""Parse the header from a region of bytes.
1258
:param bytes: The data to parse.
1259
:return: An offset, data tuple such as readv yields, for the unparsed
1260
data. (which may be of length 0).
1262
signature = bytes[0:len(self._signature())]
1263
if not signature == self._signature():
1264
raise errors.BadIndexFormatSignature(self._name, BTreeGraphIndex)
1265
lines = bytes[len(self._signature()):].splitlines()
1266
options_line = lines[0]
1267
if not options_line.startswith(_OPTION_NODE_REFS):
1268
raise errors.BadIndexOptions(self)
1270
self.node_ref_lists = int(options_line[len(_OPTION_NODE_REFS):])
1272
raise errors.BadIndexOptions(self)
1273
options_line = lines[1]
1274
if not options_line.startswith(_OPTION_KEY_ELEMENTS):
1275
raise errors.BadIndexOptions(self)
1277
self._key_length = int(options_line[len(_OPTION_KEY_ELEMENTS):])
1279
raise errors.BadIndexOptions(self)
1280
options_line = lines[2]
1281
if not options_line.startswith(_OPTION_LEN):
1282
raise errors.BadIndexOptions(self)
1284
self._key_count = int(options_line[len(_OPTION_LEN):])
1286
raise errors.BadIndexOptions(self)
1287
options_line = lines[3]
1288
if not options_line.startswith(_OPTION_ROW_LENGTHS):
1289
raise errors.BadIndexOptions(self)
1291
self._row_lengths = map(int, [length for length in
1292
options_line[len(_OPTION_ROW_LENGTHS):].split(',')
1295
raise errors.BadIndexOptions(self)
1296
self._compute_row_offsets()
1298
# calculate the bytes we have processed
1299
header_end = (len(signature) + sum(map(len, lines[0:4])) + 4)
1300
return header_end, bytes[header_end:]
1302
def _read_nodes(self, nodes):
1303
"""Read some nodes from disk into the LRU cache.
1305
This performs a readv to get the node data into memory, and parses each
1306
node, then yields it to the caller. The nodes are requested in the
1307
supplied order. If possible doing sort() on the list before requesting
1308
a read may improve performance.
1310
:param nodes: The nodes to read. 0 - first node, 1 - second node etc.
1313
# may be the byte string of the whole file
1315
# list of (offset, length) regions of the file that should, evenually
1316
# be read in to data_ranges, either from 'bytes' or from the transport
1319
offset = index * _PAGE_SIZE
1322
# Root node - special case
1324
size = min(_PAGE_SIZE, self._size)
1326
# The only case where we don't know the size, is for very
1327
# small indexes. So we read the whole thing
1328
bytes = self._transport.get_bytes(self._name)
1329
self._size = len(bytes)
1330
# the whole thing should be parsed out of 'bytes'
1331
ranges.append((0, len(bytes)))
1334
if offset > self._size:
1335
raise AssertionError('tried to read past the end'
1336
' of the file %s > %s'
1337
% (offset, self._size))
1338
size = min(size, self._size - offset)
1339
ranges.append((offset, size))
1342
elif bytes is not None:
1343
# already have the whole file
1344
data_ranges = [(start, bytes[start:start+_PAGE_SIZE])
1345
for start in xrange(0, len(bytes), _PAGE_SIZE)]
1346
elif self._file is None:
1347
data_ranges = self._transport.readv(self._name, ranges)
1350
for offset, size in ranges:
1351
self._file.seek(offset)
1352
data_ranges.append((offset, self._file.read(size)))
1353
for offset, data in data_ranges:
1355
# extract the header
1356
offset, data = self._parse_header_from_bytes(data)
1359
bytes = zlib.decompress(data)
1360
if bytes.startswith(_LEAF_FLAG):
1361
node = _LeafNode(bytes, self._key_length, self.node_ref_lists)
1362
elif bytes.startswith(_INTERNAL_FLAG):
1363
node = _InternalNode(bytes)
1365
raise AssertionError("Unknown node type for %r" % bytes)
1366
yield offset / _PAGE_SIZE, node
1368
def _signature(self):
1369
"""The file signature for this index type."""
1373
"""Validate that everything in the index can be accessed."""
1374
# just read and parse every node.
1375
self._get_root_node()
1376
if len(self._row_lengths) > 1:
1377
start_node = self._row_offsets[1]
1379
# We shouldn't be reading anything anyway
1381
node_end = self._row_offsets[-1]
1382
for node in self._read_nodes(range(start_node, node_end)):
1387
from bzrlib import _btree_serializer_pyx as _btree_serializer
1389
from bzrlib import _btree_serializer_py as _btree_serializer
added
removed
bzrlib/btree_index.py
