Line data Source code
1 : //! An LSM tree consists of multiple levels, each exponential larger than the
2 : //! previous level. And each level consists of be multiple "tiers". With tiered
3 : //! compaction, a level is compacted when it has accumulated more than N tiers,
4 : //! forming one tier on the next level.
5 : //!
6 : //! In the pageserver, we don't explicitly track the levels and tiers. Instead,
7 : //! we identify them by looking at the shapes of the layers. It's an easy task
8 : //! for a human, but it's not straightforward to come up with the exact
9 : //! rules. Especially if there are cases like interrupted, half-finished
10 : //! compactions, or highly skewed data distributions that have let us "skip"
11 : //! some levels. It's not critical to classify all cases correctly; at worst we
12 : //! delay some compaction work, and suffer from more read amplification, or we
13 : //! perform some unnecessary compaction work.
14 : //!
15 : //! `identify_level` performs that shape-matching.
16 : //!
17 : //! It returns a Level struct, which has `depth()` function to count the number
18 : //! of "tiers" in the level. The tier count is the max depth of stacked layers
19 : //! within the level. That's a good measure, because the point of compacting is
20 : //! to reduce read amplification, and the depth is what determines that.
21 : //!
22 : //! One interesting effect of this is that if we generate very small delta
23 : //! layers at L0, e.g. because the L0 layers are flushed by timeout rather than
24 : //! because they reach the target size, the L0 compaction will combine them to
25 : //! one larger file. But if the combined file is still smaller than the target
26 : //! file size, the file will still be considered to be part of L0 at the next
27 : //! iteration.
28 :
29 : use anyhow::bail;
30 : use std::collections::BTreeSet;
31 : use std::ops::Range;
32 : use utils::lsn::Lsn;
33 :
34 : use crate::interface::*;
35 :
36 : use tracing::{info, trace};
37 :
38 : pub struct Level<L> {
39 : pub lsn_range: Range<Lsn>,
40 : pub layers: Vec<L>,
41 : }
42 :
43 : /// Identify an LSN > `end_lsn` that partitions the LSN space, so that there are
44 : /// no layers that cross the boundary LSN.
45 : ///
46 : /// A further restriction is that all layers in the returned partition cover at
47 : /// most 'lsn_max_size' LSN bytes.
48 16 : pub async fn identify_level<K, L>(
49 16 : all_layers: Vec<L>,
50 16 : end_lsn: Lsn,
51 16 : lsn_max_size: u64,
52 16 : ) -> anyhow::Result<Option<Level<L>>>
53 16 : where
54 16 : K: CompactionKey,
55 16 : L: CompactionLayer<K> + Clone,
56 16 : {
57 16 : // filter out layers that are above the `end_lsn`, they are completely irrelevant.
58 16 : let mut layers = Vec::new();
59 84 : for l in all_layers {
60 70 : if l.lsn_range().start < end_lsn && l.lsn_range().end > end_lsn {
61 : // shouldn't happen. Indicates that the caller passed a bogus
62 : // end_lsn.
63 2 : bail!("identify_level() called with end_lsn that does not partition the LSN space: end_lsn {} intersects with layer {}", end_lsn, l.short_id());
64 68 : }
65 68 : // include image layers sitting exacty at `end_lsn`.
66 68 : let is_image = !l.is_delta();
67 68 : if (is_image && l.lsn_range().start > end_lsn)
68 68 : || (!is_image && l.lsn_range().start >= end_lsn)
69 : {
70 10 : continue;
71 58 : }
72 58 : layers.push(l);
73 : }
74 : // All the remaining layers either belong to this level, or are below it.
75 0 : info!(
76 0 : "identify level at {}, size {}, num layers below: {}",
77 0 : end_lsn,
78 0 : lsn_max_size,
79 0 : layers.len()
80 0 : );
81 14 : if layers.is_empty() {
82 0 : return Ok(None);
83 14 : }
84 14 :
85 14 : // Walk the ranges in LSN order.
86 14 : //
87 14 : // ----- end_lsn
88 14 : // |
89 14 : // |
90 14 : // v
91 14 : //
92 200 : layers.sort_by_key(|l| l.lsn_range().end);
93 14 : let mut candidate_start_lsn = end_lsn;
94 14 : let mut candidate_layers: Vec<L> = Vec::new();
95 14 : let mut current_best_start_lsn = end_lsn;
96 14 : let mut current_best_layers: Vec<L> = Vec::new();
97 14 : let mut iter = layers.into_iter();
98 : loop {
99 60 : let Some(l) = iter.next_back() else {
100 : // Reached end. Accept the last candidate
101 10 : current_best_start_lsn = candidate_start_lsn;
102 10 : current_best_layers.extend_from_slice(&std::mem::take(&mut candidate_layers));
103 10 : break;
104 : };
105 0 : trace!(
106 0 : "inspecting {} for candidate {}, current best {}",
107 0 : l.short_id(),
108 0 : candidate_start_lsn,
109 0 : current_best_start_lsn
110 0 : );
111 :
112 50 : let r = l.lsn_range();
113 50 :
114 50 : // Image layers don't restrict our choice of cutoff LSN
115 50 : if l.is_delta() {
116 : // Is this candidate workable? In other words, are there any
117 : // delta layers that span across this LSN
118 : //
119 : // Valid: Not valid:
120 : // + +
121 : // | | +
122 : // + <- candidate + | <- candidate
123 : // + +
124 : // |
125 : // +
126 48 : if r.end <= candidate_start_lsn {
127 42 : // Hooray, there are no crossing LSNs. And we have visited
128 42 : // through all the layers within candidate..end_lsn. The
129 42 : // current candidate can be accepted.
130 42 : current_best_start_lsn = r.end;
131 42 : current_best_layers.extend_from_slice(&std::mem::take(&mut candidate_layers));
132 42 : candidate_start_lsn = r.start;
133 42 : }
134 :
135 : // Is it small enough to be considered part of this level?
136 48 : if r.end.0 - r.start.0 > lsn_max_size {
137 : // Too large, this layer belongs to next level. Stop.
138 0 : trace!(
139 0 : "too large {}, size {} vs {}",
140 0 : l.short_id(),
141 0 : r.end.0 - r.start.0,
142 0 : lsn_max_size
143 0 : );
144 4 : break;
145 44 : }
146 44 :
147 44 : // If this crosses the candidate lsn, push it down.
148 44 : if r.start < candidate_start_lsn {
149 0 : trace!(
150 0 : "layer {} prevents from stopping at {}",
151 0 : l.short_id(),
152 0 : candidate_start_lsn
153 0 : );
154 4 : candidate_start_lsn = r.start;
155 40 : }
156 2 : }
157 :
158 : // Include this layer in our candidate
159 46 : candidate_layers.push(l);
160 : }
161 :
162 14 : Ok(if current_best_start_lsn == end_lsn {
163 : // empty level
164 0 : None
165 : } else {
166 14 : Some(Level {
167 14 : lsn_range: current_best_start_lsn..end_lsn,
168 14 : layers: current_best_layers,
169 14 : })
170 : })
171 16 : }
172 :
173 : // helper struct used in depth()
174 : struct Event<K> {
175 : key: K,
176 : layer_idx: usize,
177 : start: bool,
178 : }
179 :
180 : impl<L> Level<L> {
181 : /// Count the number of deltas stacked on each other.
182 14 : pub fn depth<K>(&self) -> u64
183 14 : where
184 14 : K: CompactionKey,
185 14 : L: CompactionLayer<K>,
186 14 : {
187 14 : let mut events: Vec<Event<K>> = Vec::new();
188 40 : for (idx, l) in self.layers.iter().enumerate() {
189 40 : events.push(Event {
190 40 : key: l.key_range().start,
191 40 : layer_idx: idx,
192 40 : start: true,
193 40 : });
194 40 : events.push(Event {
195 40 : key: l.key_range().end,
196 40 : layer_idx: idx,
197 40 : start: false,
198 40 : });
199 40 : }
200 228 : events.sort_by_key(|e| (e.key, e.start));
201 14 :
202 14 : // Sweep the key space left to right. Stop at each distinct key, and
203 14 : // count the number of deltas on top of the highest image at that key.
204 14 : //
205 14 : // This is a little enefficient, as we walk through the active_set on
206 14 : // every key. We could increment/decrement a counter on each step
207 14 : // instead, but that'd require a bit more complex bookkeeping.
208 14 : let mut active_set: BTreeSet<(Lsn, bool, usize)> = BTreeSet::new();
209 14 : let mut max_depth = 0;
210 14 : let mut events_iter = events.iter().peekable();
211 94 : while let Some(e) = events_iter.next() {
212 80 : let l = &self.layers[e.layer_idx];
213 80 : let is_image = !l.is_delta();
214 80 :
215 80 : // update the active set
216 80 : if e.start {
217 40 : active_set.insert((l.lsn_range().end, is_image, e.layer_idx));
218 40 : } else {
219 40 : active_set.remove(&(l.lsn_range().end, is_image, e.layer_idx));
220 40 : }
221 :
222 : // recalculate depth if this was the last event at this point
223 80 : let more_events_at_this_key = events_iter
224 80 : .peek()
225 80 : .map_or(false, |next_e| next_e.key == e.key);
226 80 : if !more_events_at_this_key {
227 46 : let mut active_depth = 0;
228 46 : for (_end_lsn, is_image, _idx) in active_set.iter().rev() {
229 46 : if *is_image {
230 4 : break;
231 42 : }
232 42 : active_depth += 1;
233 : }
234 46 : if active_depth > max_depth {
235 16 : max_depth = active_depth;
236 30 : }
237 34 : }
238 : }
239 14 : max_depth
240 14 : }
241 : }
242 :
243 : #[cfg(test)]
244 : mod tests {
245 : use super::*;
246 : use crate::simulator::{Key, MockDeltaLayer, MockImageLayer, MockLayer};
247 : use std::sync::{Arc, Mutex};
248 :
249 40 : fn delta(key_range: Range<Key>, lsn_range: Range<Lsn>) -> MockLayer {
250 40 : MockLayer::Delta(Arc::new(MockDeltaLayer {
251 40 : key_range,
252 40 : lsn_range,
253 40 : // identify_level() doesn't pay attention to the rest of the fields
254 40 : file_size: 0,
255 40 : deleted: Mutex::new(false),
256 40 : records: vec![],
257 40 : }))
258 40 : }
259 :
260 2 : fn image(key_range: Range<Key>, lsn: Lsn) -> MockLayer {
261 2 : MockLayer::Image(Arc::new(MockImageLayer {
262 2 : key_range,
263 2 : lsn_range: lsn..(lsn + 1),
264 2 : // identify_level() doesn't pay attention to the rest of the fields
265 2 : file_size: 0,
266 2 : deleted: Mutex::new(false),
267 2 : }))
268 2 : }
269 :
270 2 : #[tokio::test]
271 2 : async fn test_identify_level() -> anyhow::Result<()> {
272 2 : let layers = vec![
273 2 : delta(Key::MIN..Key::MAX, Lsn(0x8000)..Lsn(0x9000)),
274 2 : delta(Key::MIN..Key::MAX, Lsn(0x5000)..Lsn(0x7000)),
275 2 : delta(Key::MIN..Key::MAX, Lsn(0x4000)..Lsn(0x5000)),
276 2 : delta(Key::MIN..Key::MAX, Lsn(0x3000)..Lsn(0x4000)),
277 2 : delta(Key::MIN..Key::MAX, Lsn(0x2000)..Lsn(0x3000)),
278 2 : delta(Key::MIN..Key::MAX, Lsn(0x1000)..Lsn(0x2000)),
279 2 : ];
280 2 :
281 2 : // All layers fit in the max file size
282 2 : let level = identify_level(layers.clone(), Lsn(0x10000), 0x2000)
283 2 : .await?
284 2 : .unwrap();
285 2 : assert_eq!(level.depth(), 6);
286 2 :
287 2 : // Same LSN with smaller max file size. The second layer from the top is larger
288 2 : // and belongs to next level.
289 2 : let level = identify_level(layers.clone(), Lsn(0x10000), 0x1000)
290 2 : .await?
291 2 : .unwrap();
292 2 : assert_eq!(level.depth(), 1);
293 2 :
294 2 : // Call with a smaller LSN
295 2 : let level = identify_level(layers.clone(), Lsn(0x3000), 0x1000)
296 2 : .await?
297 2 : .unwrap();
298 2 : assert_eq!(level.depth(), 2);
299 2 :
300 2 : // Call with an LSN that doesn't partition the space
301 2 : let result = identify_level(layers, Lsn(0x6000), 0x1000).await;
302 2 : assert!(result.is_err());
303 2 : Ok(())
304 2 : }
305 :
306 2 : #[tokio::test]
307 2 : async fn test_overlapping_lsn_ranges() -> anyhow::Result<()> {
308 2 : // The files LSN ranges overlap, so even though there are more files that
309 2 : // fit under the file size, they are not included in the level because they
310 2 : // overlap so that we'd need to include the oldest file, too, which is
311 2 : // larger
312 2 : let layers = vec![
313 2 : delta(Key::MIN..Key::MAX, Lsn(0x4000)..Lsn(0x5000)),
314 2 : delta(Key::MIN..Key::MAX, Lsn(0x3000)..Lsn(0x4000)), // overlap
315 2 : delta(Key::MIN..Key::MAX, Lsn(0x2500)..Lsn(0x3500)), // overlap
316 2 : delta(Key::MIN..Key::MAX, Lsn(0x2000)..Lsn(0x3000)), // overlap
317 2 : delta(Key::MIN..Key::MAX, Lsn(0x1000)..Lsn(0x2500)), // larger
318 2 : ];
319 2 :
320 2 : let level = identify_level(layers.clone(), Lsn(0x10000), 0x1000)
321 2 : .await?
322 2 : .unwrap();
323 2 : assert_eq!(level.depth(), 1);
324 2 :
325 2 : Ok(())
326 2 : }
327 :
328 2 : #[tokio::test]
329 2 : async fn test_depth_nonoverlapping() -> anyhow::Result<()> {
330 2 : // The key ranges don't overlap, so depth is only 1.
331 2 : let layers = vec![
332 2 : delta(4000..5000, Lsn(0x6000)..Lsn(0x7000)),
333 2 : delta(3000..4000, Lsn(0x7000)..Lsn(0x8000)),
334 2 : delta(1000..2000, Lsn(0x8000)..Lsn(0x9000)),
335 2 : ];
336 2 :
337 2 : let level = identify_level(layers.clone(), Lsn(0x10000), 0x2000)
338 2 : .await?
339 2 : .unwrap();
340 2 : assert_eq!(level.layers.len(), 3);
341 2 : assert_eq!(level.depth(), 1);
342 2 :
343 2 : // Staggered. The 1st and 3rd layer don't overlap with each other.
344 2 : let layers = vec![
345 2 : delta(1000..2000, Lsn(0x8000)..Lsn(0x9000)),
346 2 : delta(1500..2500, Lsn(0x7000)..Lsn(0x8000)),
347 2 : delta(2000..3000, Lsn(0x6000)..Lsn(0x7000)),
348 2 : ];
349 2 :
350 2 : let level = identify_level(layers.clone(), Lsn(0x10000), 0x2000)
351 2 : .await?
352 2 : .unwrap();
353 2 : assert_eq!(level.layers.len(), 3);
354 2 : assert_eq!(level.depth(), 2);
355 2 : Ok(())
356 2 : }
357 :
358 2 : #[tokio::test]
359 2 : async fn test_depth_images() -> anyhow::Result<()> {
360 2 : let layers: Vec<MockLayer> = vec![
361 2 : delta(1000..2000, Lsn(0x8000)..Lsn(0x9000)),
362 2 : delta(1500..2500, Lsn(0x7000)..Lsn(0x8000)),
363 2 : delta(2000..3000, Lsn(0x6000)..Lsn(0x7000)),
364 2 : // This covers the same key range as the 2nd delta layer. The depth
365 2 : // in that key range is therefore 0.
366 2 : image(1500..2500, Lsn(0x9000)),
367 2 : ];
368 2 :
369 2 : let level = identify_level(layers.clone(), Lsn(0x10000), 0x2000)
370 2 : .await?
371 2 : .unwrap();
372 2 : assert_eq!(level.layers.len(), 4);
373 2 : assert_eq!(level.depth(), 1);
374 2 : Ok(())
375 2 : }
376 : }
|
