Line data Source code
1 : use std::ops::Range;
2 :
3 : // NOTE the `im` crate has 20x more downloads and also has
4 : // persistent/immutable BTree. But it's bugged so rpds is a
5 : // better choice <https://github.com/neondatabase/neon/issues/3395>
6 : use rpds::RedBlackTreeMapSync;
7 :
8 : /// Data structure that can efficiently:
9 : /// - find the latest layer by lsn.end at a given key
10 : /// - iterate the latest layers in a key range
11 : /// - insert layers in non-decreasing lsn.start order
12 : ///
13 : /// For a detailed explanation and justification of this approach, see:
14 : /// <https://neon.tech/blog/persistent-structures-in-neons-wal-indexing>
15 : ///
16 : /// NOTE The struct is parameterized over Value for easier
17 : /// testing, but in practice it's some sort of layer.
18 : pub struct LayerCoverage<Value> {
19 : /// For every change in coverage (as we sweep the key space)
20 : /// we store (lsn.end, value).
21 : ///
22 : /// NOTE We use an immutable/persistent tree so that we can keep historic
23 : /// versions of this coverage without cloning the whole thing and
24 : /// incurring quadratic memory cost. See HistoricLayerCoverage.
25 : ///
26 : /// NOTE We use the Sync version of the map because we want Self to
27 : /// be Sync. Using nonsync might be faster, if we can work with
28 : /// that.
29 : nodes: RedBlackTreeMapSync<i128, Option<(u64, Value)>>,
30 : }
31 :
32 : impl<T: Clone> Default for LayerCoverage<T> {
33 1124 : fn default() -> Self {
34 1124 : Self::new()
35 1124 : }
36 : }
37 :
38 : impl<Value: Clone> LayerCoverage<Value> {
39 1124 : pub fn new() -> Self {
40 1124 : Self {
41 1124 : nodes: RedBlackTreeMapSync::default(),
42 1124 : }
43 1124 : }
44 :
45 : /// Helper function to subdivide the key range without changing any values
46 : ///
47 : /// This operation has no semantic effect by itself. It only helps us pin in
48 : /// place the part of the coverage we don't want to change when inserting.
49 : ///
50 : /// As an analogy, think of a polygon. If you add a vertex along one of the
51 : /// segments, the polygon is still the same, but it behaves differently when
52 : /// we move or delete one of the other points.
53 : ///
54 : /// Complexity: O(log N)
55 2036 : fn add_node(&mut self, key: i128) {
56 2036 : let value = match self.nodes.range(..=key).last() {
57 472 : Some((_, Some(v))) => Some(v.clone()),
58 1256 : Some((_, None)) => None,
59 308 : None => None,
60 : };
61 2036 : self.nodes.insert_mut(key, value);
62 2036 : }
63 :
64 : /// Insert a layer.
65 : ///
66 : /// Complexity: worst case O(N), in practice O(log N). See NOTE in implementation.
67 1018 : pub fn insert(&mut self, key: Range<i128>, lsn: Range<u64>, value: Value) {
68 1018 : // Add nodes at endpoints
69 1018 : //
70 1018 : // NOTE The order of lines is important. We add nodes at the start
71 1018 : // and end of the key range **before updating any nodes** in order
72 1018 : // to pin down the current coverage outside of the relevant key range.
73 1018 : // Only the coverage inside the layer's key range should change.
74 1018 : self.add_node(key.start);
75 1018 : self.add_node(key.end);
76 1018 :
77 1018 : // Raise the height where necessary
78 1018 : //
79 1018 : // NOTE This loop is worst case O(N), but amortized O(log N) in the special
80 1018 : // case when rectangles have no height. In practice I don't think we'll see
81 1018 : // the kind of layer intersections needed to trigger O(N) behavior. The worst
82 1018 : // case is N/2 horizontal layers overlapped with N/2 vertical layers in a
83 1018 : // grid pattern.
84 1018 : let mut to_update = Vec::new();
85 1018 : let mut to_remove = Vec::new();
86 1018 : let mut prev_covered = false;
87 1152 : for (k, node) in self.nodes.range(key) {
88 1152 : let needs_cover = match node {
89 598 : None => true,
90 554 : Some((h, _)) => h < &lsn.end,
91 : };
92 1152 : if needs_cover {
93 1142 : match prev_covered {
94 118 : true => to_remove.push(*k),
95 1024 : false => to_update.push(*k),
96 : }
97 10 : }
98 1152 : prev_covered = needs_cover;
99 : }
100 : // TODO check if the nodes inserted at key.start and key.end are safe
101 : // to remove. It's fine to keep them but they could be redundant.
102 2042 : for k in to_update {
103 1024 : self.nodes.insert_mut(k, Some((lsn.end, value.clone())));
104 1024 : }
105 1136 : for k in to_remove {
106 118 : self.nodes.remove_mut(&k);
107 118 : }
108 1018 : }
109 :
110 : /// Get the latest (by lsn.end) layer at a given key
111 : ///
112 : /// Complexity: O(log N)
113 401356 : pub fn query(&self, key: i128) -> Option<Value> {
114 401356 : self.nodes
115 401356 : .range(..=key)
116 401356 : .next_back()?
117 : .1
118 310864 : .as_ref()
119 310864 : .map(|(_, v)| v.clone())
120 401356 : }
121 :
122 : /// Iterate the changes in layer coverage in a given range. You will likely
123 : /// want to start with self.query(key.start), and then follow up with self.range
124 : ///
125 : /// Complexity: O(log N + result_size)
126 7116 : pub fn range(&self, key: Range<i128>) -> impl '_ + Iterator<Item = (i128, Option<Value>)> {
127 7116 : self.nodes
128 7116 : .range(key)
129 13164 : .map(|(k, v)| (*k, v.as_ref().map(|x| x.1.clone())))
130 7116 : }
131 :
132 : /// Returns an iterator which includes all coverage changes for layers that intersect
133 : /// with the provided range.
134 7112 : pub fn range_overlaps(
135 7112 : &self,
136 7112 : key_range: &Range<i128>,
137 7112 : ) -> impl Iterator<Item = (i128, Option<Value>)> + '_
138 7112 : where
139 7112 : Value: Eq,
140 7112 : {
141 7112 : let first_change = self.query(key_range.start);
142 7112 : match first_change {
143 3398 : Some(change) => {
144 3398 : // If the start of the range is covered, we have to deal with two cases:
145 3398 : // 1. Start of the range is aligned with the start of a layer.
146 3398 : // In this case the return of `self.range` will contain the layer which aligns with the start of the key range.
147 3398 : // We advance said iterator to avoid duplicating the first change.
148 3398 : // 2. Start of the range is not aligned with the start of a layer.
149 3398 : let range = key_range.start..key_range.end;
150 3398 : let mut range_coverage = self.range(range).peekable();
151 3398 : if range_coverage
152 3398 : .peek()
153 3398 : .is_some_and(|c| c.1.as_ref() == Some(&change))
154 418 : {
155 418 : range_coverage.next();
156 2980 : }
157 3398 : itertools::Either::Left(
158 3398 : std::iter::once((key_range.start, Some(change))).chain(range_coverage),
159 3398 : )
160 : }
161 : None => {
162 3714 : let range = key_range.start..key_range.end;
163 3714 : let coverage = self.range(range);
164 3714 : itertools::Either::Right(coverage)
165 : }
166 : }
167 7112 : }
168 : /// O(1) clone
169 3096 : pub fn clone(&self) -> Self {
170 3096 : Self {
171 3096 : nodes: self.nodes.clone(),
172 3096 : }
173 3096 : }
174 : }
175 :
176 : /// Image and delta coverage at a specific LSN.
177 : pub struct LayerCoverageTuple<Value> {
178 : pub image_coverage: LayerCoverage<Value>,
179 : pub delta_coverage: LayerCoverage<Value>,
180 : }
181 :
182 : impl<T: Clone> Default for LayerCoverageTuple<T> {
183 562 : fn default() -> Self {
184 562 : Self {
185 562 : image_coverage: LayerCoverage::default(),
186 562 : delta_coverage: LayerCoverage::default(),
187 562 : }
188 562 : }
189 : }
190 :
191 : impl<Value: Clone> LayerCoverageTuple<Value> {
192 1548 : pub fn clone(&self) -> Self {
193 1548 : Self {
194 1548 : image_coverage: self.image_coverage.clone(),
195 1548 : delta_coverage: self.delta_coverage.clone(),
196 1548 : }
197 1548 : }
198 : }
|
