600
581
For InMemoryGraphIndex the estimate is exact.
602
return len(self._nodes) + sum(backing.key_count() for backing in
583
return len(self._keys) + sum(backing.key_count() for backing in
603
584
self._backing_indices if backing is not None)
605
586
def validate(self):
606
587
"""In memory index's have no known corruption at the moment."""
609
class _LeafNode(dict):
590
class _LeafNode(object):
610
591
"""A leaf node for a serialised B+Tree index."""
612
__slots__ = ('min_key', 'max_key', '_keys')
593
__slots__ = ('keys',)
614
595
def __init__(self, bytes, key_length, ref_list_length):
615
596
"""Parse bytes to create a leaf node object."""
616
597
# splitlines mangles the \r delimiters.. don't use it.
617
key_list = _btree_serializer._parse_leaf_lines(bytes,
618
key_length, ref_list_length)
620
self.min_key = key_list[0][0]
621
self.max_key = key_list[-1][0]
623
self.min_key = self.max_key = None
624
super(_LeafNode, self).__init__(key_list)
625
self._keys = dict(self)
628
"""Return a sorted list of (key, (value, refs)) items"""
634
"""Return a sorted list of all keys."""
598
self.keys = dict(_btree_serializer._parse_leaf_lines(bytes,
599
key_length, ref_list_length))
640
602
class _InternalNode(object):
1106
1043
output.append(cur_out)
1109
def _walk_through_internal_nodes(self, keys):
1110
"""Take the given set of keys, and find the corresponding LeafNodes.
1112
:param keys: An unsorted iterable of keys to search for
1113
:return: (nodes, index_and_keys)
1114
nodes is a dict mapping {index: LeafNode}
1115
keys_at_index is a list of tuples of [(index, [keys for Leaf])]
1117
# 6 seconds spent in miss_torture using the sorted() line.
1118
# Even with out of order disk IO it seems faster not to sort it when
1119
# large queries are being made.
1120
keys_at_index = [(0, sorted(keys))]
1122
for row_pos, next_row_start in enumerate(self._row_offsets[1:-1]):
1123
node_indexes = [idx for idx, s_keys in keys_at_index]
1124
nodes = self._get_internal_nodes(node_indexes)
1126
next_nodes_and_keys = []
1127
for node_index, sub_keys in keys_at_index:
1128
node = nodes[node_index]
1129
positions = self._multi_bisect_right(sub_keys, node.keys)
1130
node_offset = next_row_start + node.offset
1131
next_nodes_and_keys.extend([(node_offset + pos, s_keys)
1132
for pos, s_keys in positions])
1133
keys_at_index = next_nodes_and_keys
1134
# We should now be at the _LeafNodes
1135
node_indexes = [idx for idx, s_keys in keys_at_index]
1137
# TODO: We may *not* want to always read all the nodes in one
1138
# big go. Consider setting a max size on this.
1139
nodes = self._get_leaf_nodes(node_indexes)
1140
return nodes, keys_at_index
1142
1046
def iter_entries(self, keys):
1143
1047
"""Iterate over keys within the index.
1182
1086
needed_keys = keys
1183
1087
if not needed_keys:
1185
nodes, nodes_and_keys = self._walk_through_internal_nodes(needed_keys)
1089
# 6 seconds spent in miss_torture using the sorted() line.
1090
# Even with out of order disk IO it seems faster not to sort it when
1091
# large queries are being made.
1092
needed_keys = sorted(needed_keys)
1094
nodes_and_keys = [(0, needed_keys)]
1096
for row_pos, next_row_start in enumerate(self._row_offsets[1:-1]):
1097
node_indexes = [idx for idx, s_keys in nodes_and_keys]
1098
nodes = self._get_internal_nodes(node_indexes)
1100
next_nodes_and_keys = []
1101
for node_index, sub_keys in nodes_and_keys:
1102
node = nodes[node_index]
1103
positions = self._multi_bisect_right(sub_keys, node.keys)
1104
node_offset = next_row_start + node.offset
1105
next_nodes_and_keys.extend([(node_offset + pos, s_keys)
1106
for pos, s_keys in positions])
1107
nodes_and_keys = next_nodes_and_keys
1108
# We should now be at the _LeafNodes
1109
node_indexes = [idx for idx, s_keys in nodes_and_keys]
1111
# TODO: We may *not* want to always read all the nodes in one
1112
# big go. Consider setting a max size on this.
1114
nodes = self._get_leaf_nodes(node_indexes)
1186
1115
for node_index, sub_keys in nodes_and_keys:
1187
1116
if not sub_keys:
1189
1118
node = nodes[node_index]
1190
1119
for next_sub_key in sub_keys:
1191
if next_sub_key in node:
1192
value, refs = node[next_sub_key]
1120
if next_sub_key in node.keys:
1121
value, refs = node.keys[next_sub_key]
1193
1122
if self.node_ref_lists:
1194
1123
yield (self, next_sub_key, value, refs)
1196
1125
yield (self, next_sub_key, value)
1198
def _find_ancestors(self, keys, ref_list_num, parent_map, missing_keys):
1199
"""Find the parent_map information for the set of keys.
1201
This populates the parent_map dict and missing_keys set based on the
1202
queried keys. It also can fill out an arbitrary number of parents that
1203
it finds while searching for the supplied keys.
1205
It is unlikely that you want to call this directly. See
1206
"CombinedGraphIndex.find_ancestry()" for a more appropriate API.
1208
:param keys: A keys whose ancestry we want to return
1209
Every key will either end up in 'parent_map' or 'missing_keys'.
1210
:param ref_list_num: This index in the ref_lists is the parents we
1212
:param parent_map: {key: parent_keys} for keys that are present in this
1213
index. This may contain more entries than were in 'keys', that are
1214
reachable ancestors of the keys requested.
1215
:param missing_keys: keys which are known to be missing in this index.
1216
This may include parents that were not directly requested, but we
1217
were able to determine that they are not present in this index.
1218
:return: search_keys parents that were found but not queried to know
1219
if they are missing or present. Callers can re-query this index for
1220
those keys, and they will be placed into parent_map or missing_keys
1222
if not self.key_count():
1223
# We use key_count() to trigger reading the root node and
1224
# determining info about this BTreeGraphIndex
1225
# If we don't have any keys, then everything is missing
1226
missing_keys.update(keys)
1228
if ref_list_num >= self.node_ref_lists:
1229
raise ValueError('No ref list %d, index has %d ref lists'
1230
% (ref_list_num, self.node_ref_lists))
1232
# The main trick we are trying to accomplish is that when we find a
1233
# key listing its parents, we expect that the parent key is also likely
1234
# to sit on the same page. Allowing us to expand parents quickly
1235
# without suffering the full stack of bisecting, etc.
1236
nodes, nodes_and_keys = self._walk_through_internal_nodes(keys)
1238
# These are parent keys which could not be immediately resolved on the
1239
# page where the child was present. Note that we may already be
1240
# searching for that key, and it may actually be present [or known
1241
# missing] on one of the other pages we are reading.
1243
# We could try searching for them in the immediate previous or next
1244
# page. If they occur "later" we could put them in a pending lookup
1245
# set, and then for each node we read thereafter we could check to
1246
# see if they are present.
1247
# However, we don't know the impact of keeping this list of things
1248
# that I'm going to search for every node I come across from here on
1250
# It doesn't handle the case when the parent key is missing on a
1251
# page that we *don't* read. So we already have to handle being
1252
# re-entrant for that.
1253
# Since most keys contain a date string, they are more likely to be
1254
# found earlier in the file than later, but we would know that right
1255
# away (key < min_key), and wouldn't keep searching it on every other
1256
# page that we read.
1257
# Mostly, it is an idea, one which should be benchmarked.
1258
parents_not_on_page = set()
1260
for node_index, sub_keys in nodes_and_keys:
1263
# sub_keys is all of the keys we are looking for that should exist
1264
# on this page, if they aren't here, then they won't be found
1265
node = nodes[node_index]
1266
parents_to_check = set()
1267
for next_sub_key in sub_keys:
1268
if next_sub_key not in node:
1269
# This one is just not present in the index at all
1270
missing_keys.add(next_sub_key)
1272
value, refs = node[next_sub_key]
1273
parent_keys = refs[ref_list_num]
1274
parent_map[next_sub_key] = parent_keys
1275
parents_to_check.update(parent_keys)
1276
# Don't look for things we've already found
1277
parents_to_check = parents_to_check.difference(parent_map)
1278
# this can be used to test the benefit of having the check loop
1280
# parents_not_on_page.update(parents_to_check)
1282
while parents_to_check:
1283
next_parents_to_check = set()
1284
for key in parents_to_check:
1286
value, refs = node[key]
1287
parent_keys = refs[ref_list_num]
1288
parent_map[key] = parent_keys
1289
next_parents_to_check.update(parent_keys)
1291
# This parent either is genuinely missing, or should be
1292
# found on another page. Perf test whether it is better
1293
# to check if this node should fit on this page or not.
1294
# in the 'everything-in-one-pack' scenario, this *not*
1295
# doing the check is 237ms vs 243ms.
1296
# So slightly better, but I assume the standard 'lots
1297
# of packs' is going to show a reasonable improvement
1298
# from the check, because it avoids 'going around
1299
# again' for everything that is in another index
1300
# parents_not_on_page.add(key)
1301
# Missing for some reason
1302
if key < node.min_key:
1303
# in the case of bzr.dev, 3.4k/5.3k misses are
1304
# 'earlier' misses (65%)
1305
parents_not_on_page.add(key)
1306
elif key > node.max_key:
1307
# This parent key would be present on a different
1309
parents_not_on_page.add(key)
1311
# assert key != node.min_key and key != node.max_key
1312
# If it was going to be present, it would be on
1313
# *this* page, so mark it missing.
1314
missing_keys.add(key)
1315
parents_to_check = next_parents_to_check.difference(parent_map)
1316
# Might want to do another .difference() from missing_keys
1317
# parents_not_on_page could have been found on a different page, or be
1318
# known to be missing. So cull out everything that has already been
1320
search_keys = parents_not_on_page.difference(
1321
parent_map).difference(missing_keys)
1324
1127
def iter_entries_prefix(self, keys):
1325
1128
"""Iterate over keys within the index using prefix matching.