604
590
class _LeafNode(object):
605
591
"""A leaf node for a serialised B+Tree index."""
607
__slots__ = ('keys', 'min_key', 'max_key')
609
593
def __init__(self, bytes, key_length, ref_list_length):
610
594
"""Parse bytes to create a leaf node object."""
611
595
# 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)
596
self.keys = dict(_btree_serializer._parse_leaf_lines(bytes,
597
key_length, ref_list_length))
622
600
class _InternalNode(object):
623
601
"""An internal node for a serialised B+Tree index."""
625
__slots__ = ('keys', 'offset')
627
603
def __init__(self, bytes):
628
604
"""Parse bytes to create an internal node object."""
629
605
# splitlines mangles the \r delimiters.. don't use it.
670
639
self._file = None
671
640
self._recommended_pages = self._compute_recommended_pages()
672
641
self._root_node = None
673
self._base_offset = offset
674
642
# Default max size is 100,000 leave values
675
643
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)
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)
685
650
self._key_count = None
686
651
self._row_lengths = None
687
652
self._row_offsets = None # Start of each row, [-1] is the end
1087
1039
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
1042
def iter_entries(self, keys):
1124
1043
"""Iterate over keys within the index.
1163
1082
needed_keys = keys
1164
1083
if not needed_keys:
1166
nodes, nodes_and_keys = self._walk_through_internal_nodes(needed_keys)
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)
1167
1111
for node_index, sub_keys in nodes_and_keys:
1168
1112
if not sub_keys:
1177
1121
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
1123
def iter_entries_prefix(self, keys):
1307
1124
"""Iterate over keys within the index using prefix matching.
added
removed
bzrlib/btree_index.py
