1039
1075
output.append(cur_out)
1078
def _walk_through_internal_nodes(self, keys):
1079
"""Take the given set of keys, and find the corresponding LeafNodes.
1081
:param keys: An unsorted iterable of keys to search for
1082
:return: (nodes, index_and_keys)
1083
nodes is a dict mapping {index: LeafNode}
1084
keys_at_index is a list of tuples of [(index, [keys for Leaf])]
1086
# 6 seconds spent in miss_torture using the sorted() line.
1087
# Even with out of order disk IO it seems faster not to sort it when
1088
# large queries are being made.
1089
keys_at_index = [(0, sorted(keys))]
1091
for row_pos, next_row_start in enumerate(self._row_offsets[1:-1]):
1092
node_indexes = [idx for idx, s_keys in keys_at_index]
1093
nodes = self._get_internal_nodes(node_indexes)
1095
next_nodes_and_keys = []
1096
for node_index, sub_keys in keys_at_index:
1097
node = nodes[node_index]
1098
positions = self._multi_bisect_right(sub_keys, node.keys)
1099
node_offset = next_row_start + node.offset
1100
next_nodes_and_keys.extend([(node_offset + pos, s_keys)
1101
for pos, s_keys in positions])
1102
keys_at_index = next_nodes_and_keys
1103
# We should now be at the _LeafNodes
1104
node_indexes = [idx for idx, s_keys in keys_at_index]
1106
# TODO: We may *not* want to always read all the nodes in one
1107
# big go. Consider setting a max size on this.
1108
nodes = self._get_leaf_nodes(node_indexes)
1109
return nodes, keys_at_index
1042
1111
def iter_entries(self, keys):
1043
1112
"""Iterate over keys within the index.
1082
1151
needed_keys = keys
1083
1152
if not 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)
1154
nodes, nodes_and_keys = self._walk_through_internal_nodes(needed_keys)
1111
1155
for node_index, sub_keys in nodes_and_keys:
1112
1156
if not sub_keys:
1121
1165
yield (self, next_sub_key, value)
1167
def _find_ancestors(self, keys, ref_list_num, parent_map, missing_keys):
1168
"""Find the parent_map information for the set of keys.
1170
This populates the parent_map dict and missing_keys set based on the
1171
queried keys. It also can fill out an arbitrary number of parents that
1172
it finds while searching for the supplied keys.
1174
It is unlikely that you want to call this directly. See
1175
"CombinedGraphIndex.find_ancestry()" for a more appropriate API.
1177
:param keys: A keys whose ancestry we want to return
1178
Every key will either end up in 'parent_map' or 'missing_keys'.
1179
:param ref_list_num: This index in the ref_lists is the parents we
1181
:param parent_map: {key: parent_keys} for keys that are present in this
1182
index. This may contain more entries than were in 'keys', that are
1183
reachable ancestors of the keys requested.
1184
:param missing_keys: keys which are known to be missing in this index.
1185
This may include parents that were not directly requested, but we
1186
were able to determine that they are not present in this index.
1187
:return: search_keys parents that were found but not queried to know
1188
if they are missing or present. Callers can re-query this index for
1189
those keys, and they will be placed into parent_map or missing_keys
1191
if not self.key_count():
1192
# We use key_count() to trigger reading the root node and
1193
# determining info about this BTreeGraphIndex
1194
# If we don't have any keys, then everything is missing
1195
missing_keys.update(keys)
1197
if ref_list_num >= self.node_ref_lists:
1198
raise ValueError('No ref list %d, index has %d ref lists'
1199
% (ref_list_num, self.node_ref_lists))
1201
# The main trick we are trying to accomplish is that when we find a
1202
# key listing its parents, we expect that the parent key is also likely
1203
# to sit on the same page. Allowing us to expand parents quickly
1204
# without suffering the full stack of bisecting, etc.
1205
nodes, nodes_and_keys = self._walk_through_internal_nodes(keys)
1207
# These are parent keys which could not be immediately resolved on the
1208
# page where the child was present. Note that we may already be
1209
# searching for that key, and it may actually be present [or known
1210
# missing] on one of the other pages we are reading.
1212
# We could try searching for them in the immediate previous or next
1213
# page. If they occur "later" we could put them in a pending lookup
1214
# set, and then for each node we read thereafter we could check to
1215
# see if they are present.
1216
# However, we don't know the impact of keeping this list of things
1217
# that I'm going to search for every node I come across from here on
1219
# It doesn't handle the case when the parent key is missing on a
1220
# page that we *don't* read. So we already have to handle being
1221
# re-entrant for that.
1222
# Since most keys contain a date string, they are more likely to be
1223
# found earlier in the file than later, but we would know that right
1224
# away (key < min_key), and wouldn't keep searching it on every other
1225
# page that we read.
1226
# Mostly, it is an idea, one which should be benchmarked.
1227
parents_not_on_page = set()
1229
for node_index, sub_keys in nodes_and_keys:
1232
# sub_keys is all of the keys we are looking for that should exist
1233
# on this page, if they aren't here, then they won't be found
1234
node = nodes[node_index]
1235
node_keys = node.keys
1236
parents_to_check = set()
1237
for next_sub_key in sub_keys:
1238
if next_sub_key not in node_keys:
1239
# This one is just not present in the index at all
1240
missing_keys.add(next_sub_key)
1242
value, refs = node_keys[next_sub_key]
1243
parent_keys = refs[ref_list_num]
1244
parent_map[next_sub_key] = parent_keys
1245
parents_to_check.update(parent_keys)
1246
# Don't look for things we've already found
1247
parents_to_check = parents_to_check.difference(parent_map)
1248
# this can be used to test the benefit of having the check loop
1250
# parents_not_on_page.update(parents_to_check)
1252
while parents_to_check:
1253
next_parents_to_check = set()
1254
for key in parents_to_check:
1255
if key in node_keys:
1256
value, refs = node_keys[key]
1257
parent_keys = refs[ref_list_num]
1258
parent_map[key] = parent_keys
1259
next_parents_to_check.update(parent_keys)
1261
# This parent either is genuinely missing, or should be
1262
# found on another page. Perf test whether it is better
1263
# to check if this node should fit on this page or not.
1264
# in the 'everything-in-one-pack' scenario, this *not*
1265
# doing the check is 237ms vs 243ms.
1266
# So slightly better, but I assume the standard 'lots
1267
# of packs' is going to show a reasonable improvement
1268
# from the check, because it avoids 'going around
1269
# again' for everything that is in another index
1270
# parents_not_on_page.add(key)
1271
# Missing for some reason
1272
if key < node.min_key:
1273
# in the case of bzr.dev, 3.4k/5.3k misses are
1274
# 'earlier' misses (65%)
1275
parents_not_on_page.add(key)
1276
elif key > node.max_key:
1277
# This parent key would be present on a different
1279
parents_not_on_page.add(key)
1281
# assert key != node.min_key and key != node.max_key
1282
# If it was going to be present, it would be on
1283
# *this* page, so mark it missing.
1284
missing_keys.add(key)
1285
parents_to_check = next_parents_to_check.difference(parent_map)
1286
# Might want to do another .difference() from missing_keys
1287
# parents_not_on_page could have been found on a different page, or be
1288
# known to be missing. So cull out everything that has already been
1290
search_keys = parents_not_on_page.difference(
1291
parent_map).difference(missing_keys)
1123
1294
def iter_entries_prefix(self, keys):
1124
1295
"""Iterate over keys within the index using prefix matching.