rete_core/index.rs
1//! Permutation index set (SPO / POS / OSP) over integer triples (SPEC.md §6).
2//!
3//! The three orders together answer every triple-pattern shape: for any subset
4//! of bound `(s, p, o)` components, at least one permutation sorts those bound
5//! components into a leading prefix, turning the lookup into a range scan.
6//!
7//! Each permutation is **tiled** (format v0.2): consecutive runs of whole
8//! a-groups packed to a byte budget ([`INDEX_TILE_BUDGET`]), each tile a
9//! self-contained [`TripleBlock`]. A bound leading id routes to exactly one
10//! tile (binary search over the tile ranges), then jumps to its a-group via
11//! the tile's lazily-built [`GroupDirectory`]; unbound scans chain all tiles
12//! with per-tile zone-map pruning. On disk each tile is compressed
13//! independently, so a ranged reader can fetch just the tiles a query needs.
14//! v0.1 single-block sections are still read (one tile per permutation).
15
16use std::sync::Arc;
17
18use crate::chunk_cache::{CacheEntry, ChunkCache};
19use crate::triples::{GroupDirectory, Triple, TripleBlock, TripleBlockBuilder};
20
21/// A triple pattern: `None` is an unbound variable, `Some(id)` a bound term.
22pub type Pattern = (Option<u32>, Option<u32>, Option<u32>);
23
24/// The most tiles one probe-prefetch window faults ([`GraphIndex::prefetch_probe_tiles`]).
25pub(crate) const PROBE_WINDOW_TILES: usize = 1024;
26
27/// The tiles a window of upcoming probes will route to, gathered one pattern
28/// at a time ([`GraphIndex::probe_batch_add`]) so the window can grow until it
29/// is full, then faulted together ([`GraphIndex::probe_batch_fault`]).
30///
31/// Bounded in tiles ([`PROBE_WINDOW_TILES`]) and in stored bytes (an eighth
32/// of the tile cache, within 256 KiB–64 MiB): faulting every tile of a long probe
33/// list at once overran the cache, which evicted the early tiles before their
34/// probes ran, and each then came back as a request of its own, one round trip
35/// at a time (G50a).
36#[derive(Default)]
37pub(crate) struct ProbeBatch {
38 want: [std::collections::BTreeSet<usize>; NUM_PERMS],
39 tiles: usize,
40 bytes: u64,
41 /// Patterns accepted so far.
42 patterns: usize,
43}
44
45/// Default per-tile budget in (uncompressed) encoded bytes. Tiles are the
46/// independently-compressed, independently-fetchable units of a permutation
47/// section: ~64 KiB matches both zstd's sweet spot and one HTTP range read.
48/// The budget a build uses is [`crate::write_opts::tile_budget`] (`--tile-kib`).
49pub const INDEX_TILE_BUDGET: usize = crate::write_opts::DEFAULT_TILE_BUDGET_BYTES;
50
51/// Geometric tile-prefetch ramp for an unbound (multi-tile) scan (see
52/// [`GraphIndex::scan_iter`]). The first coalesced batch faults this many
53/// tiles; each subsequent batch doubles up to [`PREFETCH_WINDOW_MAX`]. Small
54/// enough that `LIMIT 1` fetches only a few tiles, large enough that a full
55/// scan still coalesces into a handful of range reads.
56const PREFETCH_WINDOW_START: usize = 4;
57const PREFETCH_WINDOW_MAX: usize = 512;
58
59/// Fetches one tile's (uncompressed) block image on demand: the bridge that
60/// lets a remote `GraphIndex` fault tiles in over a `RangeReader` without this
61/// module knowing about I/O. `None` = the fetch failed; the index records the
62/// failure ([`GraphIndex::load_incomplete`]) and the scan sees an empty tile —
63/// callers running over remote data MUST check the flag after evaluating.
64pub type TileLoader = Box<dyn Fn(usize, usize) -> Option<Vec<u8>> + Send + Sync>;
65
66/// Fetches **many** tiles of one section in one round trip: given the section
67/// and the (ascending) indices of the tiles wanted, returns each tile's
68/// uncompressed block image in the same order. The ranged reader implements
69/// this by coalescing byte-adjacent tile ranges into single range reads, so a
70/// full-section scan costs a handful of requests instead of one per tile.
71/// `None` = the batch failed as a whole; callers fall back to the per-tile
72/// [`TileLoader`] (which records per-tile failures).
73pub type TileBulkLoader = Box<dyn Fn(usize, &[usize]) -> Option<Vec<Vec<u8>>> + Send + Sync>;
74
75/// One tile of a permutation section: a fully self-contained [`TripleBlock`]
76/// over a consecutive run of whole a-groups, plus the leading-id range it
77/// covers (for routing without parsing or fetching) and its lazily-built group
78/// directory (for intra-tile point lookups). The block image itself may be
79/// local (built/decoded eagerly) or faulted in by the index's [`TileLoader`].
80pub struct Tile {
81 min_a: u32,
82 max_a: u32,
83 /// Optional **tile synopsis**: the inclusive min/max of the two non-leading
84 /// columns `(min_b, max_b, min_c, max_c)`, read from the directory trailer of
85 /// a synopsis-carrying file. Lets a scan prune this tile by a bound secondary
86 /// component *before* faulting it (remote reads). `None` = no synopsis (an
87 /// older file, or a locally-built/opened tile) — then nothing is pruned early
88 /// and the in-tile zone map prunes after the tile is in hand, as before.
89 pub(crate) syn: Option<(u32, u32, u32, u32)>,
90 /// Encoded byte length of this tile (compressed on-disk size for a remote
91 /// tile, in-memory image size for a local one; 0 = unknown). Feeds the join
92 /// planner's fatness gates without faulting any data.
93 len: u32,
94 /// Build-time staging only: a locally-built or locally-decoded tile's body
95 /// waits here until [`GraphIndex::from_sections`] moves it into the shared
96 /// [`ChunkCache`] (keyed by `(section, tile)`) and clears it. `None`
97 /// afterward, and for a remote tile whose body faults in through the loader.
98 /// The long-lived retention is the cache, not this field.
99 staged: Option<Vec<u8>>,
100}
101
102impl Tile {
103 fn local(min_a: u32, max_a: u32, bytes: Vec<u8>) -> Self {
104 let len = bytes.len().min(u32::MAX as usize) as u32;
105 Tile {
106 min_a,
107 max_a,
108 syn: None,
109 len,
110 staged: Some(bytes),
111 }
112 }
113
114 /// A remote tile's descriptor with its stored length.
115 pub(crate) fn remote_with_len(
116 min_a: u32,
117 max_a: u32,
118 syn: Option<(u32, u32, u32, u32)>,
119 len: u32,
120 ) -> Self {
121 Tile {
122 len,
123 ..Tile::remote(min_a, max_a, syn)
124 }
125 }
126
127 fn remote(min_a: u32, max_a: u32, syn: Option<(u32, u32, u32, u32)>) -> Self {
128 Tile {
129 min_a,
130 max_a,
131 syn,
132 len: 0,
133 staged: None,
134 }
135 }
136
137 /// Encoded byte length (see the field doc); falls back to a still-staged
138 /// body's size when the directory didn't provide one.
139 pub(crate) fn encoded_len(&self) -> u64 {
140 if self.len > 0 {
141 self.len as u64
142 } else {
143 self.staged.as_ref().map_or(0, |d| d.len() as u64)
144 }
145 }
146
147 /// Leading-component (permuted `a`) range this tile covers, inclusive.
148 pub fn leading_range(&self) -> (u32, u32) {
149 (self.min_a, self.max_a)
150 }
151
152 /// Could this tile hold a triple with the given bound non-leading components,
153 /// per its synopsis? `true` when there is no synopsis (can't rule it out) or
154 /// the bound `b`/`c` fall inside the recorded ranges. Conservative: a `false`
155 /// is a *proven* miss (the in-tile zone map would reject the same tile), so
156 /// skipping the fetch never drops a result.
157 pub(crate) fn syn_admits(&self, pb: Option<u32>, pc: Option<u32>) -> bool {
158 match self.syn {
159 None => true,
160 Some((min_b, max_b, min_c, max_c)) => {
161 let ok = |v: Option<u32>, lo: u32, hi: u32| v.is_none_or(|x| lo <= x && x <= hi);
162 ok(pb, min_b, max_b) && ok(pc, min_c, max_c)
163 }
164 }
165 }
166}
167
168/// Fetches stored leaf blocks of a two-level directory (format `0x06`, SPEC.md
169/// §4.4): absolute `(offset, len)` ranges in, their bytes out, in order — one
170/// batched read. `None` = the batch failed.
171pub type LeafFetch = Arc<dyn Fn(&[(u64, u64)]) -> Option<Vec<Vec<u8>>> + Send + Sync>;
172
173/// A two-level directory whose leaves fault on demand (a tile directory, a
174/// chunk directory): what [`ensure_leaves_across`] needs to fetch the leaves of
175/// several directories in **one** batched read instead of one read each.
176pub(crate) trait LeafDir {
177 /// Leaf `li`'s absolute stored range, when it exists and is not resident.
178 fn missing_leaf(&self, li: usize) -> Option<(u64, u64)>;
179 /// Decode `stored` as leaf `li` and keep it (recording a failure).
180 fn install_leaf(&self, li: usize, stored: &[u8]);
181 /// Record that a batch carrying this directory's leaves failed.
182 fn fail_leaves(&self);
183 /// The batched reader the directory's leaves come through.
184 fn leaf_fetch(&self) -> &LeafFetch;
185}
186
187/// Fault the not-yet-resident leaves `lis` of each directory in `dirs` in one
188/// batched read, through the first directory's fetcher. Every directory of one
189/// open reads through the same reader, so this is a single round trip where one
190/// read per directory was one per directory in sequence (a lookup that names
191/// terms in three dictionary sections, a pattern estimated on four
192/// permutations). A failed batch marks every directory it carried.
193pub(crate) fn ensure_leaves_across(dirs: &[(&dyn LeafDir, &[usize])]) {
194 let mut ranges: Vec<(u64, u64)> = Vec::new();
195 let mut slots: Vec<(usize, usize)> = Vec::new();
196 for (di, (dir, lis)) in dirs.iter().enumerate() {
197 // The same leaf can come many times over (a join probing 460 k subjects
198 // of one leaf): fetch each once.
199 let mut lis: Vec<usize> = lis.to_vec();
200 lis.sort_unstable();
201 lis.dedup();
202 for li in lis {
203 if let Some(r) = dir.missing_leaf(li) {
204 ranges.push(r);
205 slots.push((di, li));
206 }
207 }
208 }
209 let Some((first, _)) = dirs.first() else {
210 return;
211 };
212 if ranges.is_empty() {
213 return;
214 }
215 match (first.leaf_fetch())(&ranges) {
216 Some(blobs) if blobs.len() == slots.len() => {
217 for (&(di, li), b) in slots.iter().zip(&blobs) {
218 dirs[di].0.install_leaf(li, b);
219 }
220 }
221 _ => {
222 for &(di, _) in &slots {
223 dirs[di].0.fail_leaves();
224 }
225 }
226 }
227}
228
229/// One leaf of a permutation's tile directory, as the `0x06` root describes it.
230#[derive(Debug, Clone)]
231pub struct TileLeaf {
232 /// Index of the leaf's first tile in the permutation.
233 pub first: usize,
234 /// Tiles in the leaf.
235 pub n: usize,
236 /// Leading-id range of the leaf (first tile's min, last tile's max).
237 pub min_a: u32,
238 /// See `min_a`.
239 pub max_a: u32,
240 /// Absolute range of the stored leaf block.
241 pub leaf: (u64, u64),
242 /// Absolute offset of the leaf's first tile.
243 pub tile_base: u64,
244}
245
246/// A permutation's tile directory whose entries live in leaf blocks faulted
247/// on demand and kept for the session (format `0x06`). Routing narrows by the
248/// leaves' leading-id ranges first, so a bound lookup faults one leaf.
249/// A faulted tile-directory leaf: each tile and its absolute offset.
250type TileCell = std::sync::OnceLock<Box<[(Tile, u64)]>>;
251
252pub struct LeafedTiles {
253 leaves: Vec<TileLeaf>,
254 cells: Vec<TileCell>,
255 total: usize,
256 total_bytes: u64,
257 fetch: LeafFetch,
258 codec: u8,
259 /// Decompression bound on a leaf.
260 max_block: u64,
261 failed: std::sync::atomic::AtomicBool,
262}
263
264/// The tile standing in for one whose leaf could not be read: it routes
265/// nowhere and its synopsis admits nothing. The failure is recorded.
266static UNREADABLE_TILE: Tile = Tile {
267 min_a: u32::MAX,
268 max_a: u32::MAX,
269 syn: Some((1, 0, 1, 0)),
270 len: 0,
271 staged: None,
272};
273
274impl LeafDir for LeafedTiles {
275 fn missing_leaf(&self, li: usize) -> Option<(u64, u64)> {
276 (li < self.cells.len() && self.cells[li].get().is_none()).then(|| self.leaves[li].leaf)
277 }
278
279 fn install_leaf(&self, li: usize, stored: &[u8]) {
280 match self.decode_leaf(li, stored) {
281 Some(v) => {
282 let _ = self.cells[li].set(v);
283 }
284 None => self
285 .failed
286 .store(true, std::sync::atomic::Ordering::Relaxed),
287 }
288 }
289
290 fn fail_leaves(&self) {
291 self.failed
292 .store(true, std::sync::atomic::Ordering::Relaxed);
293 }
294
295 fn leaf_fetch(&self) -> &LeafFetch {
296 &self.fetch
297 }
298}
299
300impl LeafedTiles {
301 /// Bound every leaf decompression at `max` bytes (from the `0x06` root).
302 pub fn with_max_block(mut self, max: u64) -> Self {
303 self.max_block = max;
304 self
305 }
306
307 /// A leafed directory of `total` tiles and `total_bytes` stored bytes.
308 pub fn new(
309 leaves: Vec<TileLeaf>,
310 total: usize,
311 total_bytes: u64,
312 fetch: LeafFetch,
313 codec: u8,
314 ) -> Self {
315 let cells = leaves.iter().map(|_| std::sync::OnceLock::new()).collect();
316 LeafedTiles {
317 leaves,
318 cells,
319 total,
320 total_bytes,
321 fetch,
322 codec,
323 max_block: u64::MAX,
324 failed: std::sync::atomic::AtomicBool::new(false),
325 }
326 }
327
328 /// How many tiles hold leading id `a`, from the root alone and without
329 /// reading anything (the `rete cost` preview). A resident leaf is counted
330 /// exactly; one that is not is estimated from what the root records of it:
331 /// a leaf whose whole range is `a` holds `n` such tiles, and a leaf shared
332 /// with other ids holds its tiles' share of the range (at least one). Skewed
333 /// ids make the shared-leaf figure rough; [`Self::admitted_estimate`] reads
334 /// instead.
335 pub(crate) fn admitted_unfaulted(&self, a: u32) -> f64 {
336 let mut tiles = 0.0;
337 for (li, lf) in self.leaves.iter().enumerate() {
338 if a < lf.min_a || a > lf.max_a {
339 continue;
340 }
341 tiles += match self.cells[li].get() {
342 Some(v) => v
343 .iter()
344 .filter(|(t, _)| t.min_a <= a && a <= t.max_a)
345 .count() as f64,
346 None if lf.min_a == lf.max_a => lf.n as f64,
347 None => {
348 let width = (lf.max_a as f64 - lf.min_a as f64) + 1.0;
349 (lf.n as f64 / width).max(1.0)
350 }
351 };
352 }
353 tiles
354 }
355
356 /// The `(first tile's min_b, last tile's max_b)` of leaf `li`, faulting it;
357 /// `None` when the leaf cannot be read or a tile carries no synopsis.
358 fn leaf_b_span(&self, li: usize) -> Option<(u32, u32)> {
359 self.ensure_leaves(&[li]);
360 let v = self.cells[li].get()?;
361 let first = v.first()?.0.syn?;
362 let last = v.last()?.0.syn?;
363 Some((first.0, last.1))
364 }
365
366 /// The leaves of the run of tiles holding leading id `a` that [`admitted_estimate`](Self::admitted_estimate)
367 /// counts exactly: the first and the last, when shared with other ids.
368 pub(crate) fn run_end_leaves(&self, a: u32) -> Vec<usize> {
369 let first = self.leaves.partition_point(|lf| lf.max_a < a);
370 let mut last = first;
371 while last < self.leaves.len() && self.leaves[last].min_a <= a {
372 last += 1;
373 }
374 let shared = |li: usize| !(self.leaves[li].min_a == a && self.leaves[li].max_a == a);
375 let mut out = Vec::new();
376 if first < last {
377 if shared(first) {
378 out.push(first);
379 }
380 if last - 1 != first && shared(last - 1) {
381 out.push(last - 1);
382 }
383 }
384 out
385 }
386
387 /// Exact count of leaf `li`'s tiles that hold `a` and admit `pb`/`pc`,
388 /// faulting the leaf; the unfaulted figure if it cannot be read.
389 fn count_in_leaf(&self, li: usize, a: u32, pb: Option<u32>, pc: Option<u32>) -> f64 {
390 self.ensure_leaves(&[li]);
391 match self.cells[li].get() {
392 Some(v) => v
393 .iter()
394 .filter(|(t, _)| t.min_a <= a && a <= t.max_a && t.syn_admits(pb, pc))
395 .count() as f64,
396 None => {
397 let lf = &self.leaves[li];
398 let width = (lf.max_a as f64 - lf.min_a as f64) + 1.0;
399 (lf.n as f64 / width).max(1.0)
400 }
401 }
402 }
403
404 /// The tiles that hold leading id `a` and admit the bound second and third
405 /// keys `pb`/`pc`, counted from the root plus a few leaf reads (all kept
406 /// for the session; the lookup that follows needs them anyway).
407 ///
408 /// The planner's row estimates stand on this, so it must neither say "one
409 /// tile" for a predicate that fills a leaf nor ignore a bound class:
410 /// `?s rdf:type <small class>` is a few tiles of an 851 M-row predicate.
411 /// - A leaf shared with other ids (at most two per id) is read and counted
412 /// exactly, synopsis included.
413 /// - Leaves wholly inside `a` cost nothing when only `a` is bound (each is
414 /// `n` tiles). With `pb` bound they are ordered by `pb` (tiles are cut in
415 /// `(a, b, c)` order), so the admitted tiles are a contiguous run that two
416 /// binary searches over the leaves' `b` spans bound: O(log leaves) reads,
417 /// then the run's end leaves are counted exactly and the leaves between
418 /// them are all `n`.
419 pub(crate) fn admitted_estimate(&self, a: u32, pb: Option<u32>, pc: Option<u32>) -> f64 {
420 let first = self.leaves.partition_point(|lf| lf.max_a < a);
421 let mut last = first;
422 while last < self.leaves.len() && self.leaves[last].min_a <= a {
423 last += 1;
424 }
425 if first >= last {
426 return 0.0;
427 }
428 let mut tiles = 0.0;
429 // Pure leaves are contiguous: every shared one is first or last.
430 let pure = |li: usize| self.leaves[li].min_a == a && self.leaves[li].max_a == a;
431 let (mut p0, mut p1) = (first, last);
432 if !pure(first) {
433 tiles += self.count_in_leaf(first, a, pb, pc);
434 p0 += 1;
435 }
436 if p1 > p0 && !pure(p1 - 1) {
437 tiles += self.count_in_leaf(p1 - 1, a, pb, pc);
438 p1 -= 1;
439 }
440 if p0 >= p1 {
441 return tiles;
442 }
443 // Wholly-`a` leaves p0..p1.
444 let Some(b) = pb else {
445 return tiles + self.leaves[p0..p1].iter().map(|l| l.n as f64).sum::<f64>();
446 };
447 // First leaf whose last tile reaches `b`; first leaf that starts past it.
448 let search = |past: bool| -> Option<usize> {
449 let (mut lo, mut hi) = (p0, p1);
450 while lo < hi {
451 let mid = lo + (hi - lo) / 2;
452 let (min_b, max_b) = self.leaf_b_span(mid)?;
453 if if past { min_b > b } else { max_b >= b } {
454 hi = mid;
455 } else {
456 lo = mid + 1;
457 }
458 }
459 Some(lo)
460 };
461 let (Some(f), Some(g)) = (search(false), search(true)) else {
462 // A leaf could not be read: no narrowing, count them all.
463 return tiles + self.leaves[p0..p1].iter().map(|l| l.n as f64).sum::<f64>();
464 };
465 if f >= g {
466 return tiles;
467 }
468 // Leaves f..g may hold `b`; those strictly between the ends are all `b`.
469 tiles += self.count_in_leaf(f, a, pb, pc);
470 if g - 1 > f {
471 tiles += self.count_in_leaf(g - 1, a, pb, pc);
472 tiles += self.leaves[f + 1..g - 1]
473 .iter()
474 .map(|l| l.n as f64)
475 .sum::<f64>();
476 }
477 tiles
478 }
479
480 fn leaf_of(&self, ti: usize) -> usize {
481 self.leaves
482 .partition_point(|l| l.first <= ti)
483 .saturating_sub(1)
484 }
485
486 fn decode_leaf(&self, li: usize, stored: &[u8]) -> Option<Box<[(Tile, u64)]>> {
487 let l = &self.leaves[li];
488 let raw = crate::file::decompress_bounded(self.codec, stored, self.max_block).ok()?;
489 let ents = crate::v06::decode_tile_leaf(&raw, l.n).ok()?;
490 let mut at = l.tile_base;
491 let mut out = Vec::with_capacity(ents.len());
492 for e in ents {
493 let mut t = Tile::remote(e.min_a, e.max_a, Some(e.syn));
494 t.len = e.clen.min(u32::MAX as u64) as u32;
495 out.push((t, at));
496 at = at.checked_add(e.clen)?;
497 }
498 Some(out.into_boxed_slice())
499 }
500
501 /// Fault every not-yet-resident leaf of `lis` in one batched read.
502 pub fn ensure_leaves(&self, lis: &[usize]) {
503 ensure_leaves_across(&[(self, lis)]);
504 }
505
506 /// Fault the leaves covering tiles `[start, end)`.
507 pub fn ensure_span(&self, start: usize, end: usize) {
508 if start >= end || start >= self.total {
509 return;
510 }
511 let a = self.leaf_of(start);
512 let b = self.leaf_of(end.min(self.total) - 1);
513 self.ensure_leaves(&(a..=b).collect::<Vec<_>>());
514 }
515
516 fn entry(&self, ti: usize) -> Option<&(Tile, u64)> {
517 if ti >= self.total {
518 return None;
519 }
520 let li = self.leaf_of(ti);
521 if self.cells[li].get().is_none() {
522 self.ensure_leaves(&[li]);
523 }
524 self.cells[li].get()?.get(ti - self.leaves[li].first)
525 }
526
527 /// Tile `ti`'s absolute stored range, faulting its leaf.
528 pub fn range(&self, ti: usize) -> Option<(u64, u64)> {
529 self.entry(ti).map(|(t, off)| (*off, t.len as u64))
530 }
531
532 /// Did a leaf fault fail?
533 pub fn load_failed(&self) -> bool {
534 self.failed.load(std::sync::atomic::Ordering::Relaxed)
535 }
536
537 /// Forget a recorded leaf-fault failure.
538 pub fn reset_load_failure(&self) {
539 self.failed
540 .store(false, std::sync::atomic::Ordering::Relaxed);
541 }
542
543 /// Bytes held by the leaves faulted so far.
544 fn resident_bytes(&self) -> u64 {
545 let n: usize = self
546 .cells
547 .iter()
548 .filter_map(|c| c.get())
549 .map(|v| v.len())
550 .sum();
551 (n * (std::mem::size_of::<Tile>() + 8)) as u64
552 + (self.leaves.len() * std::mem::size_of::<TileLeaf>()) as u64
553 }
554}
555
556/// A tile list produced on first use by a closure (see [`TileList::Deferred`]).
557pub struct DeferredTiles {
558 cell: std::sync::OnceLock<Vec<Tile>>,
559 /// Stored bytes the planner may assume before the descriptors exist.
560 approx_bytes: u64,
561 load: Box<dyn Fn() -> Option<Vec<Tile>> + Send + Sync>,
562 failed: std::sync::atomic::AtomicBool,
563}
564
565impl DeferredTiles {
566 /// Descriptors built by `load` on first use and kept. A failed load is recorded
567 /// (see [`GraphIndex::load_incomplete`]) and reads as no tiles until retried.
568 pub fn new(approx_bytes: u64, load: Box<dyn Fn() -> Option<Vec<Tile>> + Send + Sync>) -> Self {
569 DeferredTiles {
570 cell: std::sync::OnceLock::new(),
571 approx_bytes,
572 load,
573 failed: std::sync::atomic::AtomicBool::new(false),
574 }
575 }
576
577 fn get(&self) -> &[Tile] {
578 if let Some(v) = self.cell.get() {
579 return v;
580 }
581 match (self.load)() {
582 Some(v) => {
583 // A racing thread may have built it first; both are the same.
584 let _ = self.cell.set(v);
585 self.cell.get().map_or(&[][..], Vec::as_slice)
586 }
587 None => {
588 // Not cached: the next use retries.
589 self.failed
590 .store(true, std::sync::atomic::Ordering::Relaxed);
591 &[]
592 }
593 }
594 }
595}
596
597/// A permutation's tile directory: resident in full (`0x05`, built indexes),
598/// or two-level with leaves faulted on demand (`0x06`).
599pub enum TileList {
600 /// Every tile descriptor resident.
601 Full(Vec<Tile>),
602 /// Root resident, leaves on demand.
603 Leafed(Arc<LeafedTiles>),
604 /// Descriptors built on first use (a quad-index view's slice of one order's
605 /// directory: an order a query never touches is never fetched).
606 Deferred(Arc<DeferredTiles>),
607}
608
609impl Default for TileList {
610 fn default() -> Self {
611 TileList::Full(Vec::new())
612 }
613}
614
615impl std::ops::Index<usize> for TileList {
616 type Output = Tile;
617 fn index(&self, ti: usize) -> &Tile {
618 match self {
619 TileList::Full(v) => &v[ti],
620 TileList::Deferred(d) => d.get().get(ti).unwrap_or(&UNREADABLE_TILE),
621 TileList::Leafed(l) => match l.entry(ti) {
622 Some((t, _)) => t,
623 None => {
624 l.failed.store(true, std::sync::atomic::Ordering::Relaxed);
625 &UNREADABLE_TILE
626 }
627 },
628 }
629 }
630}
631
632impl TileList {
633 /// Tiles known without building anything (0 for a deferred list that has
634 /// not been built).
635 fn known_len(&self) -> usize {
636 match self {
637 TileList::Deferred(d) => d.cell.get().map_or(0, Vec::len),
638 other => other.len(),
639 }
640 }
641
642 /// Number of tiles.
643 pub fn len(&self) -> usize {
644 match self {
645 TileList::Full(v) => v.len(),
646 TileList::Deferred(d) => d.get().len(),
647 TileList::Leafed(l) => l.total,
648 }
649 }
650
651 /// No tiles?
652 pub fn is_empty(&self) -> bool {
653 self.len() == 0
654 }
655
656 /// The resident descriptors of a full list (`None` for a leafed one).
657 pub fn as_full(&self) -> Option<&[Tile]> {
658 match self {
659 TileList::Full(v) => Some(v),
660 TileList::Deferred(d) => Some(d.get()),
661 TileList::Leafed(_) => None,
662 }
663 }
664
665 /// Every tile, in order, faulting all leaves first (one batch).
666 pub fn iter(&self) -> impl Iterator<Item = &Tile> {
667 if let TileList::Leafed(l) = self {
668 l.ensure_span(0, l.total);
669 }
670 (0..self.len()).map(move |i| &self[i])
671 }
672
673 /// The first tile, if any.
674 pub fn first(&self) -> Option<&Tile> {
675 (!self.is_empty()).then(|| &self[0])
676 }
677
678 /// The last tile, if any.
679 pub fn last(&self) -> Option<&Tile> {
680 (!self.is_empty()).then(|| &self[self.len() - 1])
681 }
682
683 /// Fault the leaves covering `[start, end)` in one batch (a no-op on a full
684 /// list).
685 pub fn ensure_span(&self, start: usize, end: usize) {
686 if let TileList::Leafed(l) = self {
687 l.ensure_span(start, end);
688 }
689 }
690
691 /// Index of the first tile whose `max_a >= a` (`len()` if none). A leafed
692 /// list narrows to the one leaf that can hold it before faulting.
693 pub fn first_max_ge(&self, a: u32) -> usize {
694 match self {
695 TileList::Full(v) => v.partition_point(|t| t.max_a < a),
696 TileList::Deferred(d) => d.get().partition_point(|t| t.max_a < a),
697 TileList::Leafed(l) => {
698 let li = l.leaves.partition_point(|lf| lf.max_a < a);
699 if li >= l.leaves.len() {
700 return l.total;
701 }
702 let lf = &l.leaves[li];
703 l.ensure_leaves(&[li]);
704 // Binary search inside the faulted leaf: stepping with `self[i]`
705 // costs a leaf lookup per tile, and a lookup runs this per probe
706 // row (94,700 probes x a leaf of thousands of tiles was 126 s).
707 match l.cells[li].get() {
708 Some(v) => lf.first + v.partition_point(|(t, _)| t.max_a < a),
709 None => lf.first,
710 }
711 }
712 }
713 }
714
715 /// Index of the first tile whose `min_a > a` (`len()` if none): the end of
716 /// the run of tiles that can hold leading id `a`. Faults one leaf at most.
717 pub fn first_min_gt(&self, a: u32) -> usize {
718 match self {
719 TileList::Full(v) => v.partition_point(|t| t.min_a <= a),
720 TileList::Deferred(d) => d.get().partition_point(|t| t.min_a <= a),
721 TileList::Leafed(l) => {
722 let next = l.leaves.partition_point(|lf| lf.min_a <= a);
723 if next == 0 {
724 return 0;
725 }
726 let li = next - 1;
727 let lf = &l.leaves[li];
728 l.ensure_leaves(&[li]);
729 match l.cells[li].get() {
730 Some(v) => lf.first + v.partition_point(|(t, _)| t.min_a <= a),
731 None => lf.first,
732 }
733 }
734 }
735 }
736
737 /// Stored bytes of every tile (from the root on a leafed list).
738 pub fn total_bytes(&self) -> u64 {
739 match self {
740 TileList::Full(v) => v.iter().map(Tile::encoded_len).sum(),
741 TileList::Deferred(d) => d.get().iter().map(Tile::encoded_len).sum(),
742 TileList::Leafed(l) => l.total_bytes,
743 }
744 }
745
746 /// Stored bytes of the tiles whose leading range holds `a`, **without
747 /// faulting**: exact where the leaf is resident, else the mean tile size times the
748 /// tiles the root says hold it — the planner's cost
749 /// estimate must not cost a round trip.
750 pub fn routed_bytes(&self, a: u32) -> u64 {
751 match self {
752 TileList::Full(v) => v
753 .iter()
754 .filter(|t| t.min_a <= a && a <= t.max_a)
755 .map(Tile::encoded_len)
756 .sum(),
757 // No fetch for an estimate: exact once built, else a rough size.
758 TileList::Deferred(d) => match d.cell.get() {
759 Some(v) => v
760 .iter()
761 .filter(|t| t.min_a <= a && a <= t.max_a)
762 .map(Tile::encoded_len)
763 .sum(),
764 None => d.approx_bytes,
765 },
766 TileList::Leafed(l) => {
767 let lis: Vec<usize> = l
768 .leaves
769 .iter()
770 .enumerate()
771 .filter(|(_, lf)| lf.min_a <= a && a <= lf.max_a)
772 .map(|(i, _)| i)
773 .collect();
774 if lis.is_empty() {
775 return 0;
776 }
777 if lis.iter().all(|&li| l.cells[li].get().is_some()) {
778 lis.iter()
779 .flat_map(|&li| l.cells[li].get().into_iter().flat_map(|v| v.iter()))
780 .filter(|(t, _)| t.min_a <= a && a <= t.max_a)
781 .map(|(t, _)| t.encoded_len())
782 .sum()
783 } else {
784 // The mean tile size times the tiles the root says hold
785 // `a`: a predicate filling many tiles is many tiles' bytes.
786 let mean = l.total_bytes as f64 / l.total.max(1) as f64;
787 (l.admitted_unfaulted(a).max(1.0) * mean) as u64
788 }
789 }
790 }
791 }
792
793 /// Is the leaf holding tile `ti` resident (always, for a full list)?
794 pub fn resident_at(&self, ti: usize) -> bool {
795 match self {
796 TileList::Full(_) => true,
797 TileList::Deferred(d) => d.cell.get().is_some(),
798 TileList::Leafed(l) => ti < l.total && l.cells[l.leaf_of(ti)].get().is_some(),
799 }
800 }
801
802 /// The absolute stored range of tile `ti` (leafed lists only).
803 pub fn range(&self, ti: usize) -> Option<(u64, u64)> {
804 match self {
805 TileList::Full(_) | TileList::Deferred(_) => None,
806 TileList::Leafed(l) => l.range(ti),
807 }
808 }
809
810 fn load_failed(&self) -> bool {
811 match self {
812 TileList::Leafed(l) => l.load_failed(),
813 TileList::Deferred(d) => d.failed.load(std::sync::atomic::Ordering::Relaxed),
814 TileList::Full(_) => false,
815 }
816 }
817
818 fn reset_load_failure(&self) {
819 match self {
820 TileList::Leafed(l) => l.reset_load_failure(),
821 TileList::Deferred(d) => d.failed.store(false, std::sync::atomic::Ordering::Relaxed),
822 TileList::Full(_) => {}
823 }
824 }
825
826 fn directory_bytes(&self) -> u64 {
827 match self {
828 TileList::Full(v) => (v.capacity() * std::mem::size_of::<Tile>()) as u64,
829 TileList::Deferred(d) => d
830 .cell
831 .get()
832 .map_or(0, |v| (v.capacity() * std::mem::size_of::<Tile>()) as u64),
833 TileList::Leafed(l) => l.resident_bytes(),
834 }
835 }
836}
837
838/// What a routed scan of one pattern will touch, from the tile directories
839/// alone — see [`GraphIndex::scan_plan`].
840#[derive(Debug, Clone, Copy, PartialEq, Eq)]
841pub struct ScanPlan {
842 /// The permutation the scan routes to (always inside [`PermSet::CORE`]).
843 pub permutation: IndexPermutation,
844 /// Tiles in that permutation's section — the denominator.
845 pub tiles_total: usize,
846 /// Tiles left after routing on the bound leading component.
847 pub tiles_routed: usize,
848 /// Tiles left after the synopsis prune — what the scan will actually fetch.
849 pub tiles_admitted: usize,
850 /// Encoded (on-disk) bytes of the admitted tiles: the index-side upper
851 /// bound on the scan's range reads.
852 pub tile_bytes: u64,
853 /// Encoded bytes of the whole section — what an unpruned scan would read.
854 pub section_bytes: u64,
855}
856
857/// Which stored permutation to scan. The full **six** orders (a SPARQL engine's
858/// classic set, as in QLever): together they sort the triples on *every* prefix
859/// of `(s, p, o)` columns, so for any bound-component prefix and any free
860/// component there is a permutation that routes on the prefix **and** yields the
861/// stream sorted on that free component — the precondition for a merge join.
862#[derive(Debug, Clone, Copy, PartialEq, Eq)]
863pub enum IndexPermutation {
864 Spo,
865 Sop,
866 Pso,
867 Pos,
868 Osp,
869 Ops,
870}
871
872/// The number of permutations the format can address (the width of every
873/// per-permutation array). How many a *given file* stores is its [`PermSet`].
874pub(crate) const NUM_PERMS: usize = 6;
875
876/// The six permutations in section order — the canonical iteration list. The
877/// original three (SPO, POS, OSP) lead, so a default scan's tie-break (and its
878/// provenance) is unchanged from the 3-permutation format; the three added orders
879/// (SOP, PSO, OPS) follow and exist to give a merge join a stream sorted on the
880/// join key.
881pub(crate) const ALL_PERMS: [IndexPermutation; NUM_PERMS] = [
882 IndexPermutation::Spo,
883 IndexPermutation::Pos,
884 IndexPermutation::Osp,
885 IndexPermutation::Sop,
886 IndexPermutation::Pso,
887 IndexPermutation::Ops,
888];
889
890/// **Which** permutations a file stores — a 6-bit set over
891/// [`IndexPermutation::section_index`], recorded in the header so a reader
892/// never has to assume six.
893///
894/// Every legal set is a superset of [`PermSet::CORE`] = {SPO, POS, OSP}. That
895/// is not a convention: those three **tie the longest bound prefix on all eight
896/// triple-pattern shapes**, so [`GraphIndex::best_permutation`] never selects
897/// outside them and a lean file routes every pattern exactly where a full one
898/// does — same tiles, same rows. What the other three (SOP, PSO, OPS) add is a
899/// *sort order*: they let [`GraphIndex::permutation_sorted_on`] hand a
900/// sort-merge join a co-sorted stream for three of the twelve
901/// (bound-set, join-column) shapes it otherwise has to hash-join. Absent, the
902/// merge seed declines and the hash path answers — the cost is a fast path,
903/// never correctness (`bgp::try_merge_join` already returns `Option`).
904///
905/// `CORE` is the smallest legal set, [`PermSet::ALL`] is the default and what
906/// every file written before this existed carries.
907#[derive(Debug, Clone, Copy, PartialEq, Eq, Hash)]
908pub struct PermSet(u8);
909
910impl Default for PermSet {
911 fn default() -> Self {
912 PermSet::ALL
913 }
914}
915
916impl PermSet {
917 /// All six permutations — the default, and what a `0` header byte means.
918 pub const ALL: PermSet = PermSet(0b0011_1111);
919 /// The three routing permutations {SPO, POS, OSP}: every pattern shape
920 /// routes as well as it does with six, no merge-join sort orders.
921 pub const CORE: PermSet = PermSet(0b0000_0111);
922
923 /// Parse a raw mask. `Err` when it is not a superset of [`PermSet::CORE`]
924 /// or addresses a permutation this build has no section index for — a
925 /// pattern would then route to a permutation the file does not carry, which
926 /// is the one failure mode that must never be silent.
927 pub fn from_bits(bits: u8) -> Result<PermSet, &'static str> {
928 if bits & !PermSet::ALL.0 != 0 {
929 return Err("permutation mask addresses more than six permutations");
930 }
931 if bits & PermSet::CORE.0 != PermSet::CORE.0 {
932 return Err("permutation mask must contain SPO, POS and OSP");
933 }
934 Ok(PermSet(bits))
935 }
936
937 /// The raw 6-bit mask.
938 pub fn bits(self) -> u8 {
939 self.0
940 }
941
942 /// Is `perm` stored in this file?
943 pub fn contains(self, perm: IndexPermutation) -> bool {
944 self.0 & (1 << perm.section_index()) != 0
945 }
946
947 /// How many permutation sections the file's index container holds.
948 pub fn len(self) -> usize {
949 self.0.count_ones() as usize
950 }
951
952 /// Never true: every legal set contains [`PermSet::CORE`]. Present because
953 /// a public `len` without it is a clippy error, and the honest answer is
954 /// more useful than an `allow`.
955 pub fn is_empty(self) -> bool {
956 false
957 }
958
959 /// The stored permutations in [`ALL_PERMS`] (= container) order.
960 pub fn iter(self) -> impl Iterator<Item = IndexPermutation> {
961 ALL_PERMS.into_iter().filter(move |p| self.contains(*p))
962 }
963
964 /// `perm`'s position **inside the index container**, which is its rank
965 /// among the stored permutations — not [`IndexPermutation::section_index`],
966 /// which is its slot in the format's fixed six-wide addressing. The two
967 /// coincide only for [`PermSet::ALL`].
968 pub fn position(self, perm: IndexPermutation) -> Option<usize> {
969 self.contains(perm)
970 .then(|| (self.0 & ((1u8 << perm.section_index()) - 1)).count_ones() as usize)
971 }
972
973 /// Names in container order, e.g. `["SPO", "POS", "OSP"]`.
974 pub fn names(self) -> Vec<&'static str> {
975 self.iter().map(|p| p.name()).collect()
976 }
977
978 /// Does this set carry the three merge-join orders?
979 pub fn has_merge_orders(self) -> bool {
980 self == PermSet::ALL
981 }
982}
983
984impl IndexPermutation {
985 /// Stable display name used in CLI diagnostics and provenance records.
986 pub fn name(self) -> &'static str {
987 match self {
988 IndexPermutation::Spo => "SPO",
989 IndexPermutation::Sop => "SOP",
990 IndexPermutation::Pso => "PSO",
991 IndexPermutation::Pos => "POS",
992 IndexPermutation::Osp => "OSP",
993 IndexPermutation::Ops => "OPS",
994 }
995 }
996
997 /// Section number inside the stored permutation container.
998 pub fn section_index(self) -> usize {
999 match self {
1000 IndexPermutation::Spo => 0,
1001 IndexPermutation::Pos => 1,
1002 IndexPermutation::Osp => 2,
1003 IndexPermutation::Sop => 3,
1004 IndexPermutation::Pso => 4,
1005 IndexPermutation::Ops => 5,
1006 }
1007 }
1008
1009 /// The canonical `(s, p, o)` roles in this permutation's `(a, b, c)` slots,
1010 /// as indices (0=s, 1=p, 2=o). The single source of truth for `forward` /
1011 /// `back` / `order_pattern`.
1012 pub(crate) const fn roles(self) -> [usize; 3] {
1013 match self {
1014 IndexPermutation::Spo => [0, 1, 2],
1015 IndexPermutation::Sop => [0, 2, 1],
1016 IndexPermutation::Pso => [1, 0, 2],
1017 IndexPermutation::Pos => [1, 2, 0],
1018 IndexPermutation::Osp => [2, 0, 1],
1019 IndexPermutation::Ops => [2, 1, 0],
1020 }
1021 }
1022
1023 /// Map a canonical `(s, p, o)` triple into this permutation's `(a, b, c)`.
1024 pub(crate) fn forward(self, t: Triple) -> Triple {
1025 let c = [t.0, t.1, t.2];
1026 let r = self.roles();
1027 (c[r[0]], c[r[1]], c[r[2]])
1028 }
1029
1030 /// Map a stored `(a, b, c)` back to canonical `(s, p, o)`.
1031 pub(crate) fn back(self, abc: Triple) -> Triple {
1032 let a = [abc.0, abc.1, abc.2];
1033 let r = self.roles();
1034 let mut out = [0u32; 3];
1035 out[r[0]] = a[0];
1036 out[r[1]] = a[1];
1037 out[r[2]] = a[2];
1038 (out[0], out[1], out[2])
1039 }
1040
1041 /// The pattern's bound components in this permutation's component order.
1042 pub(crate) fn order_pattern(self, p: Pattern) -> [Option<u32>; 3] {
1043 let c = [p.0, p.1, p.2];
1044 let r = self.roles();
1045 [c[r[0]], c[r[1]], c[r[2]]]
1046 }
1047
1048 /// Length of the leading run of bound components — higher is more selective.
1049 fn leading_bound(self, pat: Pattern) -> usize {
1050 self.order_pattern(pat)
1051 .iter()
1052 .take_while(|c| c.is_some())
1053 .count()
1054 }
1055}
1056
1057/// Build a [`GraphIndex`] from canonical `(s, p, o)` integer triples.
1058pub struct GraphIndexBuilder {
1059 triples: Vec<Triple>,
1060 tile_budget: usize,
1061 perms: PermSet,
1062 /// Report each finished permutation to [`crate::build_progress`].
1063 progress: bool,
1064}
1065
1066impl Default for GraphIndexBuilder {
1067 fn default() -> Self {
1068 Self {
1069 triples: Vec::new(),
1070 tile_budget: crate::write_opts::tile_budget(),
1071 perms: PermSet::ALL,
1072 progress: false,
1073 }
1074 }
1075}
1076
1077impl GraphIndexBuilder {
1078 pub fn new() -> Self {
1079 Self::default()
1080 }
1081
1082 /// Build directly from an owned triple vector — moves it in, avoiding the
1083 /// per-triple `push` copy when the caller already has the id-triples in a
1084 /// `Vec` (the assembly path, where they double as the pyramid input).
1085 pub fn from_triples(triples: Vec<Triple>) -> Self {
1086 Self {
1087 triples,
1088 ..Self::default()
1089 }
1090 }
1091
1092 /// Override the per-tile byte budget (tests use tiny budgets to force
1093 /// multi-tile sections on small data).
1094 pub fn with_tile_budget(mut self, bytes: usize) -> Self {
1095 self.tile_budget = bytes.max(1);
1096 self
1097 }
1098
1099 /// Build only the permutations in `perms` (default [`PermSet::ALL`]).
1100 /// Sections outside the set come out empty and are not written to the file.
1101 pub fn with_perms(mut self, perms: PermSet) -> Self {
1102 self.perms = perms;
1103 self
1104 }
1105
1106 /// Advance [`crate::build_progress`] by the triple count as each
1107 /// permutation is built (the in-RAM build's `index_permutations` phase).
1108 pub(crate) fn with_progress(mut self) -> Self {
1109 self.progress = true;
1110 self
1111 }
1112
1113 pub fn push(&mut self, t: Triple) {
1114 self.triples.push(t);
1115 }
1116
1117 /// Build the six permutations **one at a time** (each permutation's internal
1118 /// sort still uses every core), so only a single permuted copy of the triples
1119 /// is resident at once instead of all six concurrently. Trades the cross-
1120 /// permutation parallelism of [`Self::build`] for a much lower peak RAM — the
1121 /// large-graph low-memory path. Output is byte-identical to [`Self::build`].
1122 pub fn build_seq(self) -> GraphIndex {
1123 let triples = &self.triples;
1124 let budget = self.tile_budget;
1125 let progress = self.progress;
1126 let wanted: Vec<IndexPermutation> = self.perms.iter().collect();
1127 // Permute (parallel) + sort (parallel inside `build_tiles`) one permutation
1128 // at a time, so only a single permuted copy of the triples is resident — but
1129 // every core is still busy. Process permutations in batches of two to
1130 // overlap one permutation's tile-building with the next one's sort while
1131 // keeping the resident permuted copies to two. Byte-identical to `build`.
1132 let build_one = |perm: IndexPermutation| -> Vec<Tile> {
1133 #[cfg(feature = "parallel")]
1134 let permuted: Vec<Triple> = {
1135 use rayon::prelude::*;
1136 triples.par_iter().map(|&t| perm.forward(t)).collect()
1137 };
1138 #[cfg(not(feature = "parallel"))]
1139 let permuted: Vec<Triple> = triples.iter().map(|&t| perm.forward(t)).collect();
1140 let tiles = build_tiles(permuted, budget);
1141 if progress {
1142 crate::build_progress::advance(triples.len() as u64);
1143 }
1144 tiles
1145 };
1146 #[cfg(feature = "parallel")]
1147 let built: Vec<Vec<Tile>> = {
1148 use rayon::prelude::*;
1149 let mut built: Vec<Vec<Tile>> = Vec::with_capacity(wanted.len());
1150 for chunk in wanted.chunks(2) {
1151 built.extend(
1152 chunk
1153 .to_vec()
1154 .into_par_iter()
1155 .map(build_one)
1156 .collect::<Vec<_>>(),
1157 );
1158 }
1159 built
1160 };
1161 #[cfg(not(feature = "parallel"))]
1162 let built: Vec<Vec<Tile>> = wanted.iter().copied().map(build_one).collect();
1163 GraphIndex::from_sections(scatter(&wanted, built), self.perms)
1164 }
1165
1166 pub fn build(self) -> GraphIndex {
1167 let wanted: Vec<IndexPermutation> = self.perms.iter().collect();
1168 let triples = &self.triples;
1169 let budget = self.tile_budget;
1170 let progress = self.progress;
1171 // The permutations are independent — permute + sort + tile each.
1172 let build_one = move |perm: IndexPermutation| -> Vec<Tile> {
1173 let permuted: Vec<Triple> = triples.iter().map(|&t| perm.forward(t)).collect();
1174 let tiles = build_tiles(permuted, budget);
1175 if progress {
1176 crate::build_progress::advance(triples.len() as u64);
1177 }
1178 tiles
1179 };
1180 // Build the permutations concurrently when the `parallel` feature is on
1181 // (they share no state); the per-permutation sort inside is also
1182 // parallel. Output is byte-identical to the serial path.
1183 #[cfg(feature = "parallel")]
1184 let built: Vec<Vec<Tile>> = {
1185 use rayon::iter::{IntoParallelIterator, ParallelIterator};
1186 wanted.clone().into_par_iter().map(build_one).collect()
1187 };
1188 #[cfg(not(feature = "parallel"))]
1189 let built: Vec<Vec<Tile>> = wanted.iter().copied().map(build_one).collect();
1190 GraphIndex::from_sections(scatter(&wanted, built), self.perms)
1191 }
1192}
1193
1194/// Place each built section at its permutation's fixed [`IndexPermutation::section_index`]
1195/// slot, leaving unbuilt permutations empty. The array stays six wide however
1196/// few permutations a file carries, so every `sections[perm.section_index()]`
1197/// in the engine keeps working unchanged.
1198fn scatter(wanted: &[IndexPermutation], built: Vec<Vec<Tile>>) -> [Vec<Tile>; NUM_PERMS] {
1199 let mut sections: [Vec<Tile>; NUM_PERMS] = Default::default();
1200 for (perm, tiles) in wanted.iter().zip(built) {
1201 sections[perm.section_index()] = tiles;
1202 }
1203 sections
1204}
1205
1206/// The encoded varint length of `v` (LEB128).
1207fn varint_len(mut v: u64) -> usize {
1208 let mut n = 1;
1209 while v >= 0x80 {
1210 v >>= 7;
1211 n += 1;
1212 }
1213 n
1214}
1215
1216/// Incremental encoded-size accounting for one a-group of a tiled section —
1217/// the running-delta chain of [`build_tiles`], streamable triple by triple.
1218/// Shared by the in-memory tiler here and the external build's streaming
1219/// tiler ([`crate::extbuild`]) so both choose IDENTICAL tile boundaries and
1220/// their outputs stay byte-identical.
1221pub(crate) struct GroupSizer {
1222 /// finalized contributions: the a-delta plus every closed b-run
1223 size: usize,
1224 num_b: u64,
1225 cur_b: u32,
1226 /// the open b-run's c count / last c
1227 num_c: u64,
1228 prev_c: u32,
1229 empty: bool,
1230}
1231
1232impl GroupSizer {
1233 /// Start a group for leading id `a` whose delta base is `prev_a` (the
1234 /// previous group's leading id; `a` again for a mid-group continuation).
1235 pub(crate) fn start(a: u32, prev_a: u32) -> Self {
1236 GroupSizer {
1237 size: varint_len((a - prev_a) as u64),
1238 num_b: 0,
1239 cur_b: 0,
1240 num_c: 0,
1241 prev_c: 0,
1242 empty: true,
1243 }
1244 }
1245
1246 /// Account one `(b, c)` of this group; returns the group's total encoded
1247 /// size so far (open count varints included).
1248 pub(crate) fn push(&mut self, b: u32, c: u32) -> usize {
1249 if self.empty || b != self.cur_b {
1250 if self.empty {
1251 self.size += varint_len(b as u64); // first b-run: delta from 0
1252 } else {
1253 self.size += varint_len(self.num_c); // close the previous b-run
1254 self.size += varint_len((b - self.cur_b) as u64);
1255 }
1256 self.cur_b = b;
1257 self.num_c = 0;
1258 self.prev_c = 0;
1259 self.num_b += 1;
1260 self.empty = false;
1261 }
1262 self.size += varint_len((c - self.prev_c) as u64);
1263 self.prev_c = c;
1264 self.num_c += 1;
1265 self.total()
1266 }
1267
1268 /// The group's total encoded size so far.
1269 pub(crate) fn total(&self) -> usize {
1270 self.size
1271 + if self.empty {
1272 0
1273 } else {
1274 varint_len(self.num_c)
1275 }
1276 + varint_len(self.num_b)
1277 }
1278}
1279
1280/// Split sorted, deduped permuted triples into size-targeted tiles: append
1281/// whole a-groups until the (estimated, near-exact) encoded size would exceed
1282/// `budget`, then flush — and additionally cut WITHIN an a-group whose own
1283/// running size exceeds the budget (a mega-group: one predicate or class
1284/// carrying a large share of the graph, e.g. a 2B-triple `cites` predicate in
1285/// POS). Split slices become consecutive tiles sharing the leading id — a
1286/// bound leading id routes to the whole run of covering tiles (see
1287/// [`GraphIndex::tile_span`]) — which bounds both builder memory and the
1288/// bytes a remote reader faults for one lookup. Boundary accounting lives in
1289/// [`GroupSizer`], shared with the external build for byte-identity.
1290fn build_tiles(mut triples: Vec<Triple>, budget: usize) -> Vec<Tile> {
1291 #[cfg(feature = "parallel")]
1292 {
1293 use rayon::slice::ParallelSliceMut;
1294 triples.par_sort_unstable();
1295 }
1296 #[cfg(not(feature = "parallel"))]
1297 triples.sort_unstable();
1298 triples.dedup();
1299 if triples.is_empty() {
1300 return Vec::new();
1301 }
1302
1303 let make_tile = |run: &[Triple]| -> Tile {
1304 let mut b = TripleBlockBuilder::new();
1305 for &t in run {
1306 b.push(t);
1307 }
1308 Tile::local(run[0].0, run[run.len() - 1].0, b.build())
1309 };
1310
1311 let mut tiles = Vec::new();
1312 let mut tile_start = 0usize;
1313 let mut tile_size = 0usize; // completed groups in the current tile
1314 let mut prev_a = 0u32;
1315 let mut i = 0usize;
1316 while i < triples.len() {
1317 let a = triples[i].0;
1318 let mut slice_start = i; // group start, or the last mid-group cut
1319 let mut sizer = GroupSizer::start(a, prev_a);
1320 let mut gtotal = 0usize;
1321 while i < triples.len() && triples[i].0 == a {
1322 gtotal = sizer.push(triples[i].1, triples[i].2);
1323 i += 1;
1324 if gtotal > budget {
1325 // Mega-group cut: everything buffered — completed groups plus
1326 // the slice measured so far — becomes one tile; the group
1327 // continues in a fresh chain with `a` as its own delta base.
1328 tiles.push(make_tile(&triples[tile_start..i]));
1329 tile_start = i;
1330 tile_size = 0;
1331 prev_a = a;
1332 slice_start = i;
1333 sizer = GroupSizer::start(a, a);
1334 gtotal = 0;
1335 }
1336 }
1337 if slice_start == i {
1338 continue; // the cut landed exactly on the group's end
1339 }
1340 if slice_start > tile_start && tile_size + gtotal > budget {
1341 tiles.push(make_tile(&triples[tile_start..slice_start]));
1342 tile_start = slice_start;
1343 tile_size = 0;
1344 }
1345 tile_size += gtotal;
1346 prev_a = a;
1347 }
1348 if tile_start < triples.len() {
1349 tiles.push(make_tile(&triples[tile_start..]));
1350 }
1351 tiles
1352}
1353
1354/// The six tiled permutation sections, queryable by triple pattern.
1355pub struct GraphIndex {
1356 /// Tiles per permutation (SPO, SOP, PSO, POS, OSP, OPS — see [`ALL_PERMS`]),
1357 /// ascending in their leading-id ranges; consecutive tiles may share a
1358 /// leading id when a mega-group was split (see [`build_tiles`]).
1359 pub(crate) sections: [TileList; NUM_PERMS],
1360 /// Which of those six slots this file actually carries. Sections outside
1361 /// the set are empty and must never be consulted — an empty section is
1362 /// indistinguishable from an empty *graph*, which is why the planner asks
1363 /// this rather than the section lengths.
1364 perms: PermSet,
1365 /// Faults in remote tiles on first scan (`None` for local indexes).
1366 loader: Option<TileLoader>,
1367 /// Optional batched fetch for multi-tile scans (`None` falls back to
1368 /// one [`TileLoader`] call per tile).
1369 bulk: Option<TileBulkLoader>,
1370 /// The batched fetch for tiles known to lie far apart (a window of probes):
1371 /// exact ranges, not rounded out to cache blocks or merged across wide
1372 /// gaps. `None` falls back to [`bulk`](Self::bulk).
1373 scattered: Option<TileBulkLoader>,
1374 /// Set when the loader failed for some tile: results may be incomplete and
1375 /// the caller must surface an error rather than the partial answer.
1376 load_failed: std::sync::atomic::AtomicBool,
1377 /// The reader's concurrent-range fan-out (1 = strictly sequential) — see
1378 /// [`set_read_concurrency`](Self::set_read_concurrency).
1379 read_concurrency: usize,
1380 /// Decompressed tile bodies live here, keyed by `(section, tile)`, faulted
1381 /// in on first scan and evictable. Phase 0 caps it at unlimited, so a
1382 /// scanned permutation stays resident exactly as it did in the old per-tile
1383 /// `OnceLock`s.
1384 cache: Arc<ChunkCache>,
1385 /// The number of triples this index holds, when known without reading a
1386 /// tile (0 = unknown). A named graph's view takes it from the quad
1387 /// directory, so the planner sizes patterns inside `GRAPH <g>` against the
1388 /// graph rather than the whole file — see [`Self::size_hint`].
1389 size_hint: std::sync::atomic::AtomicU64,
1390}
1391
1392impl GraphIndex {
1393 fn from_sections(mut sections: [Vec<Tile>; NUM_PERMS], perms: PermSet) -> Self {
1394 // Move every locally-built/decoded tile body into the shared cache and
1395 // clear its staging slot, so the descriptors hold routing metadata only.
1396 let cache = ChunkCache::unlimited_arc();
1397 for (si, tiles) in sections.iter_mut().enumerate() {
1398 for (ti, tile) in tiles.iter_mut().enumerate() {
1399 if let Some(body) = tile.staged.take() {
1400 cache.insert((si as u8, ti as u32), Arc::from(body));
1401 }
1402 }
1403 }
1404 GraphIndex {
1405 sections: sections.map(TileList::Full),
1406 perms,
1407 loader: None,
1408 bulk: None,
1409 scattered: None,
1410 load_failed: std::sync::atomic::AtomicBool::new(false),
1411 read_concurrency: 1,
1412 cache,
1413 size_hint: std::sync::atomic::AtomicU64::new(0),
1414 }
1415 }
1416
1417 /// Rebuild from tiled sections: per permutation, `(min_a, max_a, block
1418 /// bytes)` per tile in ascending leading-id order — the v0.2 layout.
1419 /// Slots outside `perms` must be empty.
1420 pub fn from_tiles(sections: [Vec<(u32, u32, Vec<u8>)>; NUM_PERMS], perms: PermSet) -> Self {
1421 let sections = sections.map(|tiles| {
1422 tiles
1423 .into_iter()
1424 .map(|(min_a, max_a, bytes)| Tile::local(min_a, max_a, bytes))
1425 .collect()
1426 });
1427 Self::from_sections(sections, perms)
1428 }
1429
1430 /// The permutations this index carries.
1431 pub fn perms(&self) -> PermSet {
1432 self.perms
1433 }
1434
1435 /// Set the byte cap of this index's tile cache (bounded-export phase 2).
1436 /// `u64::MAX` (the open-time default) is unlimited — a scanned permutation
1437 /// stays resident exactly as before. A finite cap evicts least-recently-used
1438 /// tiles, so a full scan's tile residency is bounded.
1439 ///
1440 /// Only safe on an index whose tiles can be **re-faulted** (a remote/ranged
1441 /// index with a loader). A locally-built index has no loader — an evicted
1442 /// tile could never be reloaded — so the caller must not cap those; the
1443 /// bounded-export path caps only the ranged default-graph index and leaves
1444 /// resident named-graph indexes unlimited (released per slot instead).
1445 pub fn set_cache_cap(&self, cap: u64) {
1446 self.cache.set_cap(cap);
1447 }
1448
1449 /// The current tile-cache byte cap (`u64::MAX` = unlimited).
1450 pub fn cache_cap(&self) -> u64 {
1451 self.cache.cap()
1452 }
1453
1454 /// Does this index fault its tiles through a loader (a remote/ranged open)?
1455 /// Only such an index is safe to give a finite tile-cache cap: a locally
1456 /// decoded index has no loader, so an evicted tile could never be reloaded.
1457 pub fn is_lazy(&self) -> bool {
1458 self.loader.is_some()
1459 }
1460
1461 /// A snapshot of the tile cache's observability counters.
1462 pub fn cache_stats(&self) -> crate::chunk_cache::CacheStats {
1463 self.cache.stats()
1464 }
1465
1466 /// Heap bytes this index holds for as long as it is open, outside its tile
1467 /// cache: one descriptor per tile of every permutation the file carries
1468 /// (plus a tile's group directory, if a bound-leading lookup built one),
1469 /// and — on a ranged open — the two copies of the per-tile byte ranges the
1470 /// loaders capture. Grows with the tile count, never with what a scan
1471 /// touches.
1472 pub fn directory_bytes(&self) -> u64 {
1473 let tiles: u64 = self.sections.iter().map(|s| s.known_len() as u64).sum();
1474 let caps: u64 = self.sections.iter().map(TileList::directory_bytes).sum();
1475 let ranges = if self.loader.is_some() {
1476 2 * 24 * tiles
1477 } else {
1478 0
1479 };
1480 caps + ranges
1481 }
1482
1483 /// A **remote** index: only the tile directories (leading-id ranges per
1484 /// permutation, ascending) are known; tile payloads fault in through
1485 /// `loader` on first scan. Check [`load_incomplete`](Self::load_incomplete)
1486 /// after evaluating — a failed fetch must become an error, never a
1487 /// silently smaller result.
1488 #[allow(clippy::type_complexity)]
1489 pub fn from_remote_directories(
1490 directories: [Vec<(u32, u32, Option<(u32, u32, u32, u32)>)>; NUM_PERMS],
1491 perms: PermSet,
1492 loader: TileLoader,
1493 ) -> Self {
1494 let sections = directories.map(|dir| {
1495 TileList::Full(
1496 dir.into_iter()
1497 .map(|(min_a, max_a, syn)| Tile::remote(min_a, max_a, syn))
1498 .collect(),
1499 )
1500 });
1501 GraphIndex {
1502 sections,
1503 perms,
1504 loader: Some(loader),
1505 bulk: None,
1506 scattered: None,
1507 load_failed: std::sync::atomic::AtomicBool::new(false),
1508 read_concurrency: 1,
1509 cache: ChunkCache::unlimited_arc(),
1510 size_hint: std::sync::atomic::AtomicU64::new(0),
1511 }
1512 }
1513
1514 /// A remote index whose per-order tile lists are built on first use: an
1515 /// order the query never routes to costs nothing. `lists[si]` is `None` for
1516 /// an order the file does not carry. The loaders index the tiles the lists
1517 /// return.
1518 pub fn from_remote_deferred(
1519 lists: [Option<Arc<DeferredTiles>>; NUM_PERMS],
1520 perms: PermSet,
1521 loader: TileLoader,
1522 ) -> Self {
1523 GraphIndex {
1524 sections: lists.map(|l| match l {
1525 Some(l) => TileList::Deferred(l),
1526 None => TileList::Full(Vec::new()),
1527 }),
1528 perms,
1529 loader: Some(loader),
1530 bulk: None,
1531 scattered: None,
1532 load_failed: std::sync::atomic::AtomicBool::new(false),
1533 read_concurrency: 1,
1534 cache: ChunkCache::unlimited_arc(),
1535 size_hint: std::sync::atomic::AtomicU64::new(0),
1536 }
1537 }
1538
1539 /// Attach a batched tile fetcher (see [`TileBulkLoader`]): multi-tile
1540 /// scans prefetch their span through it instead of faulting tile by tile.
1541 pub fn with_bulk_loader(mut self, bulk: TileBulkLoader) -> Self {
1542 self.bulk = Some(bulk);
1543 self
1544 }
1545
1546 /// Attach the fetcher for scattered tiles ([`GraphIndex::probe_batch_fault`]):
1547 /// a probe window's tiles are far apart, so it asks for each one's exact
1548 /// range instead of the block-aligned, gap-merged spans a scan wants.
1549 pub fn with_scattered_loader(mut self, scattered: TileBulkLoader) -> Self {
1550 self.scattered = Some(scattered);
1551 self
1552 }
1553
1554 /// Record each remote tile's encoded (on-disk) byte length, per section —
1555 /// known to the ranged opener from the tile directory. Powers the join
1556 /// planner's fatness gates; never triggers a fetch.
1557 pub(crate) fn set_tile_lens(&mut self, lens: [Vec<u32>; NUM_PERMS]) {
1558 for (section, ls) in self.sections.iter_mut().zip(lens) {
1559 if let TileList::Full(section) = section {
1560 for (tile, l) in section.iter_mut().zip(ls) {
1561 tile.len = l;
1562 }
1563 }
1564 }
1565 }
1566
1567 /// Record the reader's concurrent-range fan-out (see
1568 /// [`RangeReader::concurrency`](crate::reader::RangeReader::concurrency)) —
1569 /// the join planner widens its remote probe budget when round trips
1570 /// overlap instead of serializing.
1571 pub(crate) fn set_read_concurrency(&mut self, c: usize) {
1572 self.read_concurrency = c.max(1);
1573 }
1574
1575 /// Record how many triples this index holds (see [`Self::size_hint`]).
1576 pub(crate) fn set_size_hint(&self, triples: u64) {
1577 self.size_hint
1578 .store(triples, std::sync::atomic::Ordering::Relaxed);
1579 }
1580
1581 /// The triples this index holds, when it was told without a tile read
1582 /// (a quad-index view's graph count, a resident per-graph index).
1583 pub(crate) fn size_hint(&self) -> Option<u64> {
1584 match self.size_hint.load(std::sync::atomic::Ordering::Relaxed) {
1585 0 => None,
1586 n => Some(n),
1587 }
1588 }
1589
1590 /// The reader's concurrent-range fan-out (1 = strictly sequential).
1591 pub(crate) fn read_concurrency(&self) -> usize {
1592 self.read_concurrency
1593 }
1594
1595 /// Did any tile fetch fail since this index was opened — or since the last
1596 /// [`reset_load_failure`](Self::reset_load_failure)?
1597 pub fn load_incomplete(&self) -> bool {
1598 self.load_failed.load(std::sync::atomic::Ordering::Relaxed)
1599 || self.sections.iter().any(TileList::load_failed)
1600 }
1601
1602 /// Forget recorded fetch failures — the start-of-evaluation reset for a
1603 /// RESIDENT session, making the incompleteness verdict per-query instead of
1604 /// per-open (one transient network blip used to fail every later query on
1605 /// the session). Safe because failed tiles are never cached: the next scan
1606 /// simply retries them.
1607 pub fn reset_load_failure(&self) {
1608 self.load_failed
1609 .store(false, std::sync::atomic::Ordering::Relaxed);
1610 for s in &self.sections {
1611 s.reset_load_failure();
1612 }
1613 }
1614
1615 /// True for a remote/lazy index (tiles fault in over a `RangeReader`). The
1616 /// join planner uses this to pick read-aware strategies: here a per-row index
1617 /// probe is a network round-trip, not a memory lookup, so a left-deep scan +
1618 /// hash join often beats probing a moderately-sized prefix.
1619 pub fn is_remote(&self) -> bool {
1620 self.loader.is_some()
1621 }
1622
1623 /// Batch-fault the unloaded tiles in `[start, end)` of `section` through
1624 /// the bulk loader, if one is attached and at least two tiles are missing
1625 /// (a single missing tile costs the same either way). A failed batch is
1626 /// not an error here: the tiles stay unloaded and the per-tile loader
1627 /// retries each one (recording failures) when the scan reaches it.
1628 /// Fault, in one batch per permutation, the tile-directory leaves the
1629 /// bound leading ids of `patterns` route to (format `0x06`; a no-op on a
1630 /// resident directory).
1631 pub(crate) fn prefault_leaves(&self, patterns: &[Pattern]) {
1632 let mut want: [std::collections::BTreeSet<usize>; NUM_PERMS] = Default::default();
1633 for &pat in patterns {
1634 let perm = Self::best_permutation_in(self.perms, pat);
1635 let [pa, _, _] = perm.order_pattern(pat);
1636 let si = perm.section_index();
1637 if let (TileList::Leafed(l), Some(a)) = (&self.sections[si], pa) {
1638 // Leaves ascend by leading id: the run whose range holds `a`.
1639 let from = l.leaves.partition_point(|lf| lf.max_a < a);
1640 for (li, lf) in l.leaves.iter().enumerate().skip(from) {
1641 if lf.min_a > a {
1642 break;
1643 }
1644 want[si].insert(li);
1645 }
1646 }
1647 }
1648 for (si, lis) in want.iter().enumerate() {
1649 if let TileList::Leafed(l) = &self.sections[si] {
1650 if !lis.is_empty() {
1651 l.ensure_leaves(&lis.iter().copied().collect::<Vec<_>>());
1652 }
1653 }
1654 }
1655 }
1656
1657 /// The absolute stored range of tile `(section, tile)` on a leafed
1658 /// (`0x06`) index, faulting its directory leaf.
1659 pub fn tile_range(&self, section: usize, tile: usize) -> Option<(u64, u64)> {
1660 self.sections.get(section)?.range(tile)
1661 }
1662
1663 /// A remote `0x06` index: per stored permutation a two-level tile
1664 /// directory, tiles faulted through `loader`/`bulk`.
1665 pub fn from_leafed(
1666 lists: [Option<Arc<LeafedTiles>>; NUM_PERMS],
1667 perms: PermSet,
1668 loader: TileLoader,
1669 ) -> Self {
1670 GraphIndex {
1671 sections: lists.map(|l| match l {
1672 Some(l) => TileList::Leafed(l),
1673 None => TileList::Full(Vec::new()),
1674 }),
1675 perms,
1676 loader: Some(loader),
1677 bulk: None,
1678 scattered: None,
1679 load_failed: std::sync::atomic::AtomicBool::new(false),
1680 read_concurrency: 1,
1681 cache: ChunkCache::unlimited_arc(),
1682 size_hint: std::sync::atomic::AtomicU64::new(0),
1683 }
1684 }
1685
1686 fn prefetch_span(&self, section: usize, start: usize, end: usize) {
1687 let missing: Vec<usize> = (start..end)
1688 .filter(|&ti| !self.tile_resident(section, ti))
1689 .collect();
1690 self.bulk_fault(section, &missing);
1691 }
1692
1693 /// Is tile `(section, tile)`'s body currently resident in the cache?
1694 fn tile_resident(&self, section: usize, tile: usize) -> bool {
1695 self.cache.contains((section as u8, tile as u32))
1696 }
1697
1698 /// Bulk-fault a set of (possibly scattered, ascending) missing tile indices in
1699 /// `section` through the bulk loader in one coalesced read, storing each image.
1700 /// No-op without a bulk loader or for fewer than two tiles (a single fault
1701 /// costs the same via the per-tile loader). A failed batch leaves the tiles
1702 /// unloaded for the per-tile loader to retry. Shared by the consecutive-span
1703 /// scan prefetch and the scattered batch-probe prefetch.
1704 fn bulk_fault(&self, section: usize, tiles: &[usize]) {
1705 self.fault_with(self.bulk.as_ref(), section, tiles);
1706 }
1707
1708 /// [`bulk_fault`](Self::bulk_fault) through `loader`.
1709 fn fault_with(&self, loader: Option<&TileBulkLoader>, section: usize, tiles: &[usize]) {
1710 if tiles.len() < 2 {
1711 return;
1712 }
1713 let Some(bulk) = loader else { return };
1714 if let Some(images) = bulk(section, tiles) {
1715 if images.len() == tiles.len() {
1716 for (&ti, img) in tiles.iter().zip(images) {
1717 self.cache
1718 .insert((section as u8, ti as u32), Arc::from(img));
1719 }
1720 }
1721 }
1722 }
1723
1724 /// Batch-fault the tiles a set of upcoming **probe** patterns will route to,
1725 /// before they are probed one at a time. Each probe (a bound leading id)
1726 /// routes to a single tile; gathered across the batch those tiles are
1727 /// scattered, so faulting them together turns N sequential remote round trips
1728 /// into a few coalesced parallel reads — the read-amplification win for a
1729 /// label-heavy join over a lazy reader. Honors the synopsis prune (never
1730 /// fetches a tile a bound secondary proves can't match). No-op for a local
1731 /// index or when fewer than two tiles per section need faulting.
1732 ///
1733 /// The batch is **bounded** ([`ProbeBatch`]): it covers the leading
1734 /// patterns whose tiles fit, and returns how many that is (at least one for
1735 /// a non-empty list; all of them for a local index). A caller with more
1736 /// patterns asks again for the rest after probing these.
1737 pub(crate) fn prefetch_probe_tiles(&self, patterns: &[Pattern]) -> usize {
1738 if self.bulk.is_none() {
1739 return patterns.len();
1740 }
1741 self.prefault_leaves(patterns);
1742 let mut batch = ProbeBatch::default();
1743 let mut covered = 0;
1744 for &pat in patterns {
1745 if !self.probe_batch_add(&mut batch, pat) {
1746 break;
1747 }
1748 covered += 1;
1749 }
1750 self.probe_batch_fault(batch);
1751 covered
1752 }
1753
1754 /// Add `pat`'s tiles to `batch` if the batch stays within its bounds
1755 /// (tiles, stored bytes); `false` leaves the batch as it was. The first
1756 /// pattern of a batch is always accepted: if it alone is wider than a
1757 /// batch (a bound predicate with nothing else routes to its whole span) it
1758 /// adds no tiles, and its own scan faults what it reads. The directory
1759 /// leaves of `pat` should be resident
1760 /// ([`prefault_leaves`](Self::prefault_leaves)), or this faults them one
1761 /// at a time.
1762 pub(crate) fn probe_batch_add(&self, batch: &mut ProbeBatch, pat: Pattern) -> bool {
1763 let perm = Self::best_permutation(pat);
1764 let [pa, pb, pc] = perm.order_pattern(pat);
1765 let si = perm.section_index();
1766 let (start, end) = self.tile_span(si, pa);
1767 // This pattern's tiles that are not already wanted or resident.
1768 let mut fresh: Vec<usize> = Vec::new();
1769 for ti in start..end {
1770 if self.sections[si][ti].syn_admits(pb, pc)
1771 && !self.tile_resident(si, ti)
1772 && !batch.want[si].contains(&ti)
1773 {
1774 fresh.push(ti);
1775 }
1776 }
1777 let fresh_bytes: u64 = fresh
1778 .iter()
1779 .map(|&ti| self.sections[si][ti].len as u64)
1780 .sum();
1781 if batch.tiles + fresh.len() > PROBE_WINDOW_TILES
1782 || batch.bytes + fresh_bytes > self.probe_window_bytes()
1783 {
1784 if batch.patterns > 0 {
1785 return false;
1786 }
1787 batch.patterns = 1;
1788 return true;
1789 }
1790 batch.tiles += fresh.len();
1791 batch.bytes += fresh_bytes;
1792 batch.want[si].extend(fresh);
1793 batch.patterns += 1;
1794 true
1795 }
1796
1797 /// Fault every tile of `batch`: per section one coalesced, parallel read.
1798 pub(crate) fn probe_batch_fault(&self, batch: ProbeBatch) {
1799 for (si, set) in batch.want.iter().enumerate() {
1800 let tiles: Vec<usize> = set.iter().copied().collect();
1801 self.fault_with(self.scattered.as_ref().or(self.bulk.as_ref()), si, &tiles);
1802 }
1803 }
1804
1805 /// The stored bytes one [`ProbeBatch`] may fault: an eighth of the tile
1806 /// cache's cap, within 256 KiB–64 MiB. A tile decodes to a few times its
1807 /// stored size and the window must stay resident until its probes have
1808 /// run, so a window that outgrew the cache would evict its own early
1809 /// tiles. With the CLI's default budget (a 128 MiB tile cache) that is
1810 /// 16 MiB: tens of coalesced requests, enough to keep the reader's
1811 /// fan-out busy.
1812 fn probe_window_bytes(&self) -> u64 {
1813 const KIB: u64 = 1 << 10;
1814 match self.cache.cap() {
1815 u64::MAX => 64 * KIB * KIB,
1816 cap => (cap / 8).clamp(256 * KIB, 64 * KIB * KIB),
1817 }
1818 }
1819
1820 /// The tile's cache entry, faulting its block image in through the loader if
1821 /// remote. Returns an owned [`CacheEntry`] handle (an `Arc`) the caller
1822 /// holds for the duration of one parse/scan — valid even if a later eviction
1823 /// drops the map slot (the handout invariant).
1824 ///
1825 /// A FAILED fetch records the failure and caches nothing (returns `None`),
1826 /// so a later evaluation retries the tile — a transient network error must
1827 /// not permanently poison a long-lived (resident) session. A local tile
1828 /// with no body returns an empty entry (constructed empty on purpose).
1829 fn tile_entry(&self, section: usize, tile: usize) -> Option<Arc<CacheEntry>> {
1830 // A query run under a memory ceiling ([`crate::guard`]) stops here.
1831 crate::guard::check();
1832 let key = (section as u8, tile as u32);
1833 if let Some(e) = self.cache.get(key) {
1834 return Some(e);
1835 }
1836 match &self.loader {
1837 Some(load) => match load(section, tile) {
1838 Some(bytes) => Some(self.cache.insert(key, Arc::from(bytes))),
1839 None => {
1840 self.load_failed
1841 .store(true, std::sync::atomic::Ordering::Relaxed);
1842 None
1843 }
1844 },
1845 None => Some(self.cache.insert(key, Arc::from(Vec::new()))),
1846 }
1847 }
1848
1849 /// Tile `(section, tile)`'s a-group directory, held in its cache entry and
1850 /// built from `block` (that entry's parsed body) on first use. Its bytes
1851 /// are charged to the tile cache when built, so a capped cache bounds the
1852 /// directories too, and they are evicted with their tile.
1853 fn tile_dir<'e>(
1854 &self,
1855 section: usize,
1856 tile: usize,
1857 entry: &'e Arc<CacheEntry>,
1858 block: &TripleBlock<'_>,
1859 ) -> &'e GroupDirectory {
1860 let (dir, built) = entry.group_dir(|| block.group_directory());
1861 if built {
1862 self.cache.charge_extra(
1863 (section as u8, tile as u32),
1864 entry,
1865 dir.heap_bytes() + std::mem::size_of::<GroupDirectory>() as u64,
1866 );
1867 }
1868 dir
1869 }
1870
1871 /// The tile's decompressed block image as a shared handle, for the file
1872 /// writer (local indexes, all tiles resident). Empty on a miss.
1873 pub(crate) fn tile_body(&self, section: usize, tile: usize) -> Arc<[u8]> {
1874 self.tile_entry(section, tile)
1875 .map(|e| e.body_arc())
1876 .unwrap_or_else(|| Arc::from(Vec::new()))
1877 }
1878
1879 /// Total triple count (sum of the SPO tiles' zone counts). For a remote
1880 /// index this faults in the SPO tiles — prefer the header's quad count.
1881 pub fn triple_count(&self) -> u32 {
1882 self.prefetch_span(0, 0, self.sections[0].len());
1883 (0..self.sections[0].len())
1884 .filter_map(|ti| {
1885 let entry = self.tile_entry(0, ti)?;
1886 TripleBlock::parse(entry.body())
1887 .ok()
1888 .map(|b| b.zone().count)
1889 })
1890 // A hostile tile can claim any count: saturate, never overflow.
1891 .fold(0u32, u32::saturating_add)
1892 }
1893
1894 /// Stream **every** triple of one stored permutation in that section's own
1895 /// order, as `(a, b, c)` in the permutation's column order (SPO gives
1896 /// `(s, p, o)`, OSP gives `(o, s, p)`, …).
1897 ///
1898 /// The access a whole-graph pass needs — `rete recard`'s streaming card
1899 /// derivation reads SPO for out-degrees and OSP for in-degrees and literal
1900 /// objects — and one [`scan_iter`](Self::scan_iter) cannot give, because a
1901 /// fully unbound pattern always routes to SPO. Tiles fault with the same
1902 /// geometric prefetch as a scan and are decoded one at a time; nothing is
1903 /// retained beyond the tile cache, so memory follows its cap, not the
1904 /// section. A failed fault is recorded in
1905 /// [`load_incomplete`](Self::load_incomplete) and the tile is skipped, as
1906 /// everywhere else. Returns `false` when the file does not store `perm`.
1907 pub fn for_each_in_order(
1908 &self,
1909 perm: IndexPermutation,
1910 f: &mut dyn FnMut(u32, u32, u32),
1911 ) -> bool {
1912 if !self.perms.contains(perm) {
1913 return false;
1914 }
1915 let si = perm.section_index();
1916 let end = self.sections[si].len();
1917 let mut window = PREFETCH_WINDOW_START;
1918 for ti in 0..end {
1919 if !self.tile_resident(si, ti) {
1920 self.prefetch_span(si, ti, (ti + window).min(end));
1921 window = window.saturating_mul(2).min(PREFETCH_WINDOW_MAX);
1922 }
1923 if let Some(entry) = self.tile_entry(si, ti) {
1924 if let Ok(block) = TripleBlock::parse(entry.body()) {
1925 for (a, b, c) in block.scan(None, None, None) {
1926 f(a, b, c);
1927 }
1928 }
1929 }
1930 }
1931 true
1932 }
1933
1934 /// The tiles of each permutation section (in `ALL_PERMS` order), for the
1935 /// file writer.
1936 pub fn tile_sections(&self) -> [&TileList; NUM_PERMS] {
1937 [
1938 &self.sections[0],
1939 &self.sections[1],
1940 &self.sections[2],
1941 &self.sections[3],
1942 &self.sections[4],
1943 &self.sections[5],
1944 ]
1945 }
1946
1947 /// The permutation selected for a pattern: the one with the longest bound
1948 /// prefix. Ties keep the canonical SPO order (then POS, then OSP), which
1949 /// makes provenance stable for unbound or equally selective shapes and routes
1950 /// a fully unbound pattern to the SPO block rather than fetching all three.
1951 ///
1952 /// Six-permutation shorthand for [`best_permutation_in`]. Its answer is
1953 /// always inside [`PermSet::CORE`] (see `perm_routing_never_leaves_core`),
1954 /// so a file with any legal [`PermSet`] routes identically — but call the
1955 /// `_in` form wherever the file's set is at hand, so that stays true by
1956 /// construction rather than by argument.
1957 ///
1958 /// [`best_permutation_in`]: GraphIndex::best_permutation_in
1959 pub fn best_permutation(pattern: Pattern) -> IndexPermutation {
1960 Self::best_permutation_in(PermSet::ALL, pattern)
1961 }
1962
1963 /// [`best_permutation`](GraphIndex::best_permutation) restricted to the
1964 /// permutations a file carries.
1965 pub fn best_permutation_in(perms: PermSet, pattern: Pattern) -> IndexPermutation {
1966 let mut best = IndexPermutation::Spo;
1967 let mut best_score = best.leading_bound(pattern);
1968 for perm in perms.iter() {
1969 let score = perm.leading_bound(pattern);
1970 if score > best_score {
1971 best = perm;
1972 best_score = score;
1973 }
1974 }
1975 best
1976 }
1977
1978 /// Choose a permutation that **routes** on `pattern`'s bound prefix *and*
1979 /// streams sorted on the canonical column `sort_col` (0=subject, 1=predicate,
1980 /// 2=object) — i.e. `sort_col` is the leading *free* component after the bound
1981 /// prefix. Returns `None` if `sort_col` is itself bound (then every row shares
1982 /// that value — any permutation is "sorted" on it) or no permutation qualifies.
1983 /// This is the precondition a merge join needs: both inputs sorted on the join
1984 /// key. Among qualifiers, the longest bound prefix (best routing) wins.
1985 ///
1986 /// Six-permutation shorthand for [`permutation_sorted_on_in`].
1987 ///
1988 /// [`permutation_sorted_on_in`]: GraphIndex::permutation_sorted_on_in
1989 pub fn permutation_sorted_on(pattern: Pattern, sort_col: usize) -> Option<IndexPermutation> {
1990 Self::permutation_sorted_on_in(PermSet::ALL, pattern, sort_col)
1991 }
1992
1993 /// [`permutation_sorted_on`](GraphIndex::permutation_sorted_on) restricted
1994 /// to the permutations a file carries — **and it will not trade routing for
1995 /// sort order**.
1996 ///
1997 /// A qualifier must match the bound prefix that
1998 /// [`best_permutation_in`](GraphIndex::best_permutation_in) achieves.
1999 /// With all six that costs nothing: every ordering of `(s, p, o)` exists, so
2000 /// "the bound components first, then `sort_col`" is always available at the
2001 /// maximal prefix. With [`PermSet::CORE`] it is what stops a merge join from
2002 /// buying its co-sorted stream with a **whole-section scan** — e.g.
2003 /// `?s :p ?o` sorted on `?s` has no `P*S` order left, and the only remaining
2004 /// sorted-on-subject stream is an unrouted SPO scan of the entire graph. The
2005 /// merge seed then declines and the hash/probe path answers, which is
2006 /// strictly the cheaper plan.
2007 pub fn permutation_sorted_on_in(
2008 perms: PermSet,
2009 pattern: Pattern,
2010 sort_col: usize,
2011 ) -> Option<IndexPermutation> {
2012 let bound = [
2013 pattern.0.is_some(),
2014 pattern.1.is_some(),
2015 pattern.2.is_some(),
2016 ];
2017 if bound[sort_col] {
2018 return None;
2019 }
2020 let routed = Self::best_permutation_in(perms, pattern).leading_bound(pattern);
2021 let mut best: Option<(IndexPermutation, usize)> = None;
2022 for perm in perms.iter() {
2023 let roles = perm.roles();
2024 let lead = perm.leading_bound(pattern);
2025 // The component at slot `lead` is the first free one; it must be the
2026 // sort column (so the stream is sorted on it after the bound prefix).
2027 if lead < 3 && roles[lead] == sort_col && best.map(|(_, s)| lead > s).unwrap_or(true) {
2028 best = Some((perm, lead));
2029 }
2030 }
2031 best.filter(|&(_, lead)| lead >= routed).map(|(p, _)| p)
2032 }
2033
2034 /// Match one already-decoded serialized permutation block (a single v0.1
2035 /// section or one v0.2 tile). This is the core primitive for range-routed
2036 /// readers: the caller fetches only the selected payload, then this scans
2037 /// it as if it came from a full [`GraphIndex`].
2038 pub fn match_serialized_block(
2039 bytes: &[u8],
2040 permutation: IndexPermutation,
2041 pattern: Pattern,
2042 ) -> Vec<Triple> {
2043 let [pa, pb, pc] = permutation.order_pattern(pattern);
2044 let mut out: Vec<Triple> = TripleBlock::parse(bytes)
2045 .ok()
2046 .filter(|b| b.zone().may_contain(pa, pb, pc))
2047 .map(|b| b.scan(pa, pb, pc))
2048 .into_iter()
2049 .flatten()
2050 .map(move |abc| permutation.back(abc))
2051 .collect();
2052 out.sort_unstable();
2053 out
2054 }
2055
2056 /// All triples matching `pattern`, returned in canonical `(s, p, o)` order.
2057 ///
2058 /// Thin eager wrapper over [`scan_iter`](Self::scan_iter): collect the lazy
2059 /// stream and restore the canonical sort (the stream is sorted in the chosen
2060 /// permutation's order, which differs from canonical once `perm.back`
2061 /// permutes the free components).
2062 pub fn match_pattern(&self, pattern: Pattern) -> Vec<Triple> {
2063 let mut out: Vec<Triple> = self.scan_iter(pattern).collect();
2064 out.sort_unstable();
2065 out
2066 }
2067
2068 /// Lazily stream the triples matching `pattern` in canonical `(s, p, o)`
2069 /// order *within the chosen permutation* — the streaming entry point for
2070 /// callers that can stop early (ASK, `LIMIT`, BGP probes) or that don't need
2071 /// the canonical re-sort. Decodes only the matching groups (see
2072 /// [`TripleBlock::scan`]); a malformed/absent block yields nothing rather
2073 /// than panicking. The permutation is chosen for the longest bound prefix.
2074 pub fn scan_iter(&self, pattern: Pattern) -> impl Iterator<Item = Triple> + '_ {
2075 self.scan_iter_with(pattern, Self::best_permutation(pattern))
2076 }
2077
2078 /// What [`scan_iter`](Self::scan_iter) *would* fetch for `pattern`, computed
2079 /// from the tile directories alone — **no tile is fetched, decoded, or even
2080 /// looked at**. This is the cost-preview half of the routed scan: it walks
2081 /// the same three decisions the scan makes (permutation, routed tile span,
2082 /// synopsis prune) and reports what survives, so a caller can say "this dump
2083 /// will pull 16 MB of index, not 376 MB" before starting it.
2084 ///
2085 /// It is a preview, not a promise: `tile_bytes` is an **upper bound** on the
2086 /// index bytes the scan reads (an admitted tile can still be rejected by its
2087 /// in-tile zone map once in hand, and a re-scan of an already-faulted tile
2088 /// reads nothing), and it says nothing about the dictionary, which is where
2089 /// a term-resolving dump's real cost usually lives.
2090 pub fn scan_plan(&self, pattern: Pattern) -> ScanPlan {
2091 let perm = Self::best_permutation_in(self.perms, pattern);
2092 let [_, pb, pc] = perm.order_pattern(pattern);
2093 let si = perm.section_index();
2094 let tiles = &self.sections[si];
2095 let (start, end) = self.tile_span(si, perm.order_pattern(pattern)[0]);
2096 let mut admitted = 0usize;
2097 let mut tile_bytes = 0u64;
2098 for ti in start..end {
2099 let t = &tiles[ti];
2100 if t.syn_admits(pb, pc) {
2101 admitted += 1;
2102 tile_bytes += t.encoded_len();
2103 }
2104 }
2105 ScanPlan {
2106 permutation: perm,
2107 tiles_total: tiles.len(),
2108 tiles_routed: end - start,
2109 tiles_admitted: admitted,
2110 tile_bytes,
2111 section_bytes: tiles.total_bytes(),
2112 }
2113 }
2114
2115 /// One bounded, **resumable** slice of `pattern`'s matches: the pull form of
2116 /// [`scan_iter`](Self::scan_iter), for a caller that cannot hold a borrow
2117 /// across calls.
2118 ///
2119 /// Returns `(triples, next_cursor, done)`; start at `cursor = 0` and feed the
2120 /// returned cursor back until `done`. The whole resume state is one opaque
2121 /// `u64` — `(tile index, next a-group id)` — so no iterator, no thread and no
2122 /// self-referential struct has to survive between calls. That is the shape a
2123 /// foreign-function boundary needs: a wasm module cannot suspend a Rust
2124 /// iterator mid-scan and hand control back to its host, but it can be handed
2125 /// a `u64` and told to carry on.
2126 ///
2127 /// `max_rows` is a **floor**, not a hard cut: a batch always ends on an
2128 /// a-group boundary of the chosen permutation, so no group is ever split
2129 /// across two calls and the overshoot is one group's fanout. Nothing is ever
2130 /// rescanned ([`TripleBlock::scan_resume`] jumps straight to the group
2131 /// through the tile's directory), so a full drain is O(n) overall — unlike
2132 /// `(offset, limit)`, which is O(n²/limit).
2133 ///
2134 /// Every call either yields at least one row or reports `done`, so a caller
2135 /// never has to guard against an empty non-final batch.
2136 ///
2137 /// **Order.** Rows arrive in the chosen permutation's order and each batch is
2138 /// re-sorted canonically, exactly as [`match_pattern`](Self::match_pattern)
2139 /// does for a whole result. For a fully unbound pattern (routed to SPO) that
2140 /// makes the concatenation of every batch identical to `match_pattern`'s
2141 /// output; for a bound pattern the *set* is identical but the canonical
2142 /// re-sort is per batch rather than global.
2143 ///
2144 /// **Memory** is O(`max_rows` + one a-group + the tiles faulted so far) —
2145 /// never O(matches). That is the whole point: an unbounded `(?s ?p ?o)` over
2146 /// a 26-million-quad graph used to materialize every match before the caller
2147 /// saw one row.
2148 pub fn scan_batch(
2149 &self,
2150 pattern: Pattern,
2151 cursor: u64,
2152 max_rows: usize,
2153 ) -> (Vec<Triple>, u64, bool) {
2154 let max_rows = max_rows.max(1);
2155 let perm = Self::best_permutation(pattern);
2156 let [pa, pb, pc] = perm.order_pattern(pattern);
2157 let si = perm.section_index();
2158 // Same routing as `scan_iter_with`: the tile directory alone decides the
2159 // span, so this is free on a lazy/remote open.
2160 let (start, end) = self.tile_span(si, pa);
2161 let mut ti = ((cursor >> 32) as usize).max(start);
2162 let mut from_a = cursor as u32;
2163 let mut out: Vec<Triple> = Vec::new();
2164 // Same geometric prefetch ramp as the streaming scan: a batch that stops
2165 // after one tile never fetches the ones behind it.
2166 let mut window = PREFETCH_WINDOW_START;
2167 while ti < end {
2168 // Synopsis pre-fault prune — no tile is fetched to reject it.
2169 if !self.sections[si][ti].syn_admits(pb, pc) {
2170 ti += 1;
2171 from_a = 0;
2172 continue;
2173 }
2174 if !self.tile_resident(si, ti) {
2175 self.prefetch_span(si, ti, (ti + window).min(end));
2176 window = window.saturating_mul(2).min(PREFETCH_WINDOW_MAX);
2177 }
2178 let entry = self.tile_entry(si, ti);
2179 let mut resume_at: Option<u32> = None;
2180 if let Some(block) = entry
2181 .as_deref()
2182 .and_then(|e| TripleBlock::parse(e.body()).ok())
2183 .filter(|b| b.zone().may_contain(pa, pb, pc))
2184 {
2185 let rows = match pa {
2186 // A bound leading id is one a-group; it cannot be split
2187 // inside a tile, only across tiles (a mega-group split), so
2188 // there is nothing finer to resume at than the tile.
2189 Some(a) => {
2190 let dir = self.tile_dir(si, ti, entry.as_ref().unwrap(), &block);
2191 block.scan_from(dir, a, pb, pc)
2192 }
2193 // Fresh tile: no directory needed, so a straight drain never
2194 // pays for one.
2195 None if from_a == 0 => block.scan(None, pb, pc),
2196 None => {
2197 let dir = self.tile_dir(si, ti, entry.as_ref().unwrap(), &block);
2198 block.scan_resume(dir, from_a, pb, pc)
2199 }
2200 };
2201 // Only a leading-unbound scan can be resumed mid-tile: with `pa`
2202 // bound there is one a-group and `scan_from` always restarts it
2203 // from the top, so cutting inside it would duplicate rows.
2204 let splittable = pa.is_none();
2205 let mut last_a: Option<u32> = None;
2206 for abc in rows {
2207 // Full, and this row opens a new a-group ⇒ cut here and
2208 // resume at this group. The row is dropped, not lost: the
2209 // next call re-enters at exactly this id.
2210 if splittable && out.len() >= max_rows && last_a != Some(abc.0) {
2211 resume_at = Some(abc.0);
2212 break;
2213 }
2214 last_a = Some(abc.0);
2215 out.push(perm.back(abc));
2216 }
2217 }
2218 if let Some(a) = resume_at {
2219 out.sort_unstable();
2220 return (out, ((ti as u64) << 32) | a as u64, false);
2221 }
2222 ti += 1;
2223 from_a = 0;
2224 if out.len() >= max_rows {
2225 break;
2226 }
2227 }
2228 let done = ti >= end;
2229 out.sort_unstable();
2230 (out, (ti as u64) << 32, done)
2231 }
2232
2233 /// Stream `pattern`'s matches **sorted on the canonical column `sort_col`**
2234 /// (0=subject, 1=predicate, 2=object): the stream's `sort_col` values are
2235 /// ascending. Chooses a permutation that routes on the bound prefix *and*
2236 /// orders by `sort_col` ([`permutation_sorted_on`](Self::permutation_sorted_on));
2237 /// `None` when none qualifies (e.g. `sort_col` is itself bound). The
2238 /// precondition for feeding a [merge join](crate::bgp): both inputs sorted on
2239 /// the shared join key.
2240 pub(crate) fn scan_iter_sorted_on(
2241 &self,
2242 pattern: Pattern,
2243 sort_col: usize,
2244 ) -> Option<impl Iterator<Item = Triple> + '_> {
2245 Some(self.scan_iter_with(
2246 pattern,
2247 Self::permutation_sorted_on_in(self.perms, pattern, sort_col)?,
2248 ))
2249 }
2250
2251 /// Stream `pattern`'s matches using a **specific** permutation `perm`; the
2252 /// stream is sorted in `perm`'s `(a, b, c)` order. Shared core of `scan_iter`
2253 /// and `scan_iter_sorted_on`.
2254 fn scan_iter_with(
2255 &self,
2256 pattern: Pattern,
2257 perm: IndexPermutation,
2258 ) -> impl Iterator<Item = Triple> + '_ {
2259 // Route: a bound leading id binary-searches the tile directory to exactly
2260 // one tile (groups are never split across tiles); an unbound one chains
2261 // every tile's cursor. Within a tile, a bound leading scan jumps to its
2262 // a-group via the tile's lazily-built group directory — built on first
2263 // use, costing one walk of that (budget-sized) tile.
2264 let [pa, pb, pc] = perm.order_pattern(pattern);
2265 let si = perm.section_index();
2266 let (start, end) = self.tile_span(si, pa);
2267 // Coalesce tile faults, but ramp the prefetch window geometrically as
2268 // the scan advances rather than fetching the whole span up front: a
2269 // small `LIMIT` (which stops pulling early) then never faults tiles
2270 // past the rows it needs, while a full scan still batches into a
2271 // handful of coalesced reads (4, 8, 16, … tiles). A bound leading scan
2272 // spans a single tile, so the prefetch no-ops and it faults just that
2273 // one tile, unchanged.
2274 let window = std::cell::Cell::new(PREFETCH_WINDOW_START);
2275 (start..end)
2276 // Synopsis pre-fault prune: drop a routed tile the directory proves
2277 // can't match a bound secondary component, **without fetching it**
2278 // (the remote win — a negative/sparse lookup costs zero tile reads).
2279 // A bound leading id routes to a single tile, so this is where it
2280 // bites; a fully-unbound leading scan leaves `pb`/`pc` unbound, so
2281 // `syn_admits` keeps every tile, unchanged.
2282 .filter(move |&ti| self.sections[si][ti].syn_admits(pb, pc))
2283 .flat_map(move |ti| {
2284 // Fault in (if remote), parse (untrusted bytes ⇒ `None` on
2285 // malformed), and zone-prune per tile, then stream the
2286 // matching groups.
2287 if !self.tile_resident(si, ti) {
2288 let w = window.get();
2289 self.prefetch_span(si, ti, (ti + w).min(end));
2290 window.set(w.saturating_mul(2).min(PREFETCH_WINDOW_MAX));
2291 }
2292 // A failed fetch is recorded by the index and scans as empty.
2293 let rows: Vec<Triple> = self.tile_rows(si, ti, pa, pb, pc).unwrap_or_default();
2294 rows.into_iter()
2295 })
2296 .map(move |abc| perm.back(abc))
2297 }
2298
2299 /// The rows of tile `ti` of section `si` matching `(pa, pb, pc)` (in that
2300 /// section's `a, b, c` order). The tile is decoded once and its matching
2301 /// rows materialized into an owned `Vec`, so the borrowed block/cursor never
2302 /// outlives the tile's cache handle (an `Arc` dropped on return); the buffer
2303 /// is bounded to one tile's rows. `None` = the tile could not be fetched
2304 /// (the failure is recorded, see [`Self::load_incomplete`]); a malformed or
2305 /// zone-pruned tile yields no rows.
2306 fn tile_rows(
2307 &self,
2308 si: usize,
2309 ti: usize,
2310 pa: Option<u32>,
2311 pb: Option<u32>,
2312 pc: Option<u32>,
2313 ) -> Option<Vec<Triple>> {
2314 let entry = self.tile_entry(si, ti)?;
2315 Some(
2316 TripleBlock::parse(entry.body())
2317 .ok()
2318 .filter(|b| b.zone().may_contain(pa, pb, pc))
2319 .map(|b| match pa {
2320 Some(a) => {
2321 let dir = self.tile_dir(si, ti, &entry, &b);
2322 b.scan_from(dir, a, pb, pc).collect()
2323 }
2324 None => b.scan(pa, pb, pc).collect(),
2325 })
2326 .unwrap_or_default(),
2327 )
2328 }
2329
2330 /// Is the object of every triple with predicate `p` inside an object-id
2331 /// range `covers` accepts? Read off the predicate-leading tiles (POS, or PSO):
2332 /// a tile whose synopsis object range `covers` accepts is proven without
2333 /// being read — the synopsis bounds every row of the tile, whatever its
2334 /// predicate, so a tile shared with other predicates only makes the test
2335 /// stricter. The tiles left (no synopsis, or a range it does not accept) are
2336 /// read and their rows of `p` checked one by one: on a lazy index only when
2337 /// there are at most `max_read` of them, in one batch, else the answer is
2338 /// `false`; on a resident one always, in memory. `false` also when a tile
2339 /// cannot be read. An index without `p` proves it trivially.
2340 pub(crate) fn predicate_objects_within(
2341 &self,
2342 p: u32,
2343 covers: impl Fn(u32, u32) -> bool,
2344 max_read: usize,
2345 ) -> bool {
2346 // (perm, which synopsis/row column holds the object)
2347 let (perm, object_is_b) = if self.perms.contains(IndexPermutation::Pos) {
2348 (IndexPermutation::Pos, true)
2349 } else if self.perms.contains(IndexPermutation::Pso) {
2350 (IndexPermutation::Pso, false)
2351 } else {
2352 return false;
2353 };
2354 let si = perm.section_index();
2355 let (start, end) = self.tile_span(si, Some(p));
2356 let unproven: Vec<usize> = (start..end)
2357 .filter(|&ti| match self.sections[si][ti].syn {
2358 Some((min_b, max_b, min_c, max_c)) => {
2359 let (lo, hi) = if object_is_b {
2360 (min_b, max_b)
2361 } else {
2362 (min_c, max_c)
2363 };
2364 !covers(lo, hi)
2365 }
2366 None => true,
2367 })
2368 .collect();
2369 if unproven.is_empty() {
2370 return true;
2371 }
2372 if self.is_lazy() {
2373 if unproven.len() > max_read {
2374 return false;
2375 }
2376 let missing: Vec<usize> = unproven
2377 .iter()
2378 .copied()
2379 .filter(|&ti| !self.tile_resident(si, ti))
2380 .collect();
2381 self.bulk_fault(si, &missing);
2382 }
2383 for ti in unproven {
2384 let Some(rows) = self.tile_rows(si, ti, Some(p), None, None) else {
2385 return false;
2386 };
2387 let all = rows.iter().all(|&(_, b, c)| {
2388 let o = if object_is_b { b } else { c };
2389 covers(o, o)
2390 });
2391 if !all {
2392 return false;
2393 }
2394 }
2395 true
2396 }
2397
2398 /// Fault, in one batched read across every permutation, the directory leaves
2399 /// [`estimate_matches`](Self::estimate_matches) reads for `patterns`: for each
2400 /// permutation that leads with a bound id, the leaves its run of tiles starts
2401 /// and ends in (a leaf wholly inside the run is counted from the root). The
2402 /// estimate then reads them from memory instead of one permutation at a time.
2403 /// A no-op on a resident directory.
2404 pub(crate) fn prefault_estimate_leaves(&self, patterns: &[Pattern]) {
2405 let mut want: [std::collections::BTreeSet<usize>; NUM_PERMS] = Default::default();
2406 for &pattern in patterns {
2407 for perm in self.perms.iter() {
2408 let [Some(a), _, _] = perm.order_pattern(pattern) else {
2409 continue;
2410 };
2411 let si = perm.section_index();
2412 if let TileList::Leafed(l) = &self.sections[si] {
2413 want[si].extend(l.run_end_leaves(a));
2414 }
2415 }
2416 }
2417 let wanted: Vec<(usize, Vec<usize>)> = want
2418 .into_iter()
2419 .enumerate()
2420 .filter(|(_, lis)| !lis.is_empty())
2421 .map(|(si, lis)| (si, lis.into_iter().collect()))
2422 .collect();
2423 let across: Vec<(&dyn LeafDir, &[usize])> = wanted
2424 .iter()
2425 .filter_map(|(si, lis)| match &self.sections[*si] {
2426 TileList::Leafed(l) => Some((&**l as &dyn LeafDir, lis.as_slice())),
2427 _ => None,
2428 })
2429 .collect();
2430 ensure_leaves_across(&across);
2431 }
2432
2433 /// A planner estimate of how many triples match `pattern`, read off the
2434 /// tile directories alone — no tile is fetched or decoded, so it costs the
2435 /// same on a lazy 50 GB file as on a resident one. `total` is the number
2436 /// of triples the index holds (an upper bound is fine).
2437 ///
2438 /// For every carried permutation that leads with a bound component, the
2439 /// routed tiles the synopsis admits are counted; a tile holds about
2440 /// `total / tiles` triples, and the routed run is charged half a tile less
2441 /// than its length, since a key rarely starts and ends on tile boundaries.
2442 /// The smallest figure wins, so `?s a :Rare` (a few tiles of OPS) beats
2443 /// `?s :p ?o` (every tile of `:p` in PSO). Its resolution is one tile:
2444 /// patterns inside one tile tie, which is all a join order needs. A
2445 /// pattern with no bound component estimates `total`; one the synopsis
2446 /// rules out estimates 0.
2447 pub(crate) fn estimate_matches(&self, pattern: Pattern, total: u64) -> f64 {
2448 let mut best = total as f64;
2449 for perm in self.perms.iter() {
2450 let [pa, pb, pc] = perm.order_pattern(pattern);
2451 if pa.is_none() {
2452 continue;
2453 }
2454 let si = perm.section_index();
2455 let tiles = &self.sections[si];
2456 if tiles.is_empty() {
2457 return 0.0;
2458 }
2459 let per_tile = total as f64 / tiles.len() as f64;
2460 // A leafed list whose leaf is not resident must still be estimated
2461 // from the root alone (a cost estimate must not cost a read), and
2462 // must give the figure the same directory gives when laid out flat
2463 // (format 0x05) to within the unavoidable: a predicate that fills
2464 // thousands of tiles is thousands of tiles whichever layout stores
2465 // them. Counting "one tile" here made the planner see a 2 M-row
2466 // predicate as 7 k rows and hash-join it (see [`LeafedTiles::admitted_estimate`]).
2467 let admitted = match (tiles, pa) {
2468 (TileList::Leafed(l), Some(a))
2469 if !l
2470 .leaves
2471 .iter()
2472 .enumerate()
2473 .filter(|(_, lf)| lf.min_a <= a && a <= lf.max_a)
2474 .all(|(li, _)| l.cells[li].get().is_some()) =>
2475 {
2476 l.admitted_estimate(a, pb, pc)
2477 }
2478 _ => {
2479 let (start, end) = self.tile_span(si, pa);
2480 (start..end)
2481 .filter(|&ti| tiles[ti].syn_admits(pb, pc))
2482 .count() as f64
2483 }
2484 };
2485 let est = if admitted <= 0.0 {
2486 0.0
2487 } else {
2488 (admitted - 0.5).max(0.5) * per_tile
2489 };
2490 best = best.min(est);
2491 }
2492 best
2493 }
2494
2495 /// The smallest leading id `>= a` present in permutation section `section`,
2496 /// or `None` past the last one. Routes by the tile directory, then reads the
2497 /// tile's own group directory — so it skips the ids a tile's *range* covers
2498 /// but its body does not hold. That matters where a range is conservative:
2499 /// a quad-index view's tile shared with other graphs is `0..=u32::MAX`, and
2500 /// stepping through it id by id would never finish. Faults the tiles it
2501 /// reads; a failed fault ends the search (`load_incomplete` says so).
2502 pub(crate) fn next_leading(&self, section: usize, a: u32) -> Option<u32> {
2503 let tiles = &self.sections[section];
2504 let mut ti = tiles.first_max_ge(a);
2505 while ti < tiles.len() {
2506 let from = a.max(tiles[ti].min_a);
2507 let entry = self.tile_entry(section, ti)?;
2508 if let Ok(block) = TripleBlock::parse(entry.body()) {
2509 let dir = self.tile_dir(section, ti, &entry, &block);
2510 if let Some(x) = dir.first_at_or_after(from) {
2511 return Some(x);
2512 }
2513 }
2514 ti += 1;
2515 }
2516 None
2517 }
2518
2519 /// The tile index span a scan must visit: every tile when the leading
2520 /// component is unbound, else the run of tiles whose leading-id ranges
2521 /// cover it. Tile ranges are ascending; consecutive tiles may SHARE a
2522 /// leading id when a mega-group was split across tiles (see
2523 /// [`build_tiles`]), so the span is a range scan, not a single hit —
2524 /// files without splits still yield a span of at most one tile.
2525 pub(crate) fn tile_span(&self, section: usize, pa: Option<u32>) -> (usize, usize) {
2526 let tiles = &self.sections[section];
2527 match pa {
2528 None => {
2529 // An unbound scan visits every tile: fault every leaf at once.
2530 tiles.ensure_span(0, tiles.len());
2531 (0, tiles.len())
2532 }
2533 Some(a) => {
2534 let i = tiles.first_max_ge(a);
2535 let j = tiles.first_min_gt(a).max(i);
2536 // The callers walk the run: fault its leaves in one batch, as
2537 // the tile-by-tile walk this replaced did one by one.
2538 tiles.ensure_span(i, j);
2539 (i, j)
2540 }
2541 }
2542 }
2543}
2544
2545#[cfg(test)]
2546mod tests {
2547 use super::*;
2548
2549 fn graph() -> (GraphIndex, Vec<Triple>) {
2550 let data = vec![
2551 (1, 10, 100),
2552 (1, 10, 101),
2553 (1, 11, 100),
2554 (2, 10, 100),
2555 (2, 12, 200),
2556 (3, 11, 300),
2557 ];
2558 let mut b = GraphIndexBuilder::new();
2559 for &t in &data {
2560 b.push(t);
2561 }
2562 (b.build(), data)
2563 }
2564
2565 /// Brute-force reference for a pattern.
2566 fn reference(data: &[Triple], (s, p, o): Pattern) -> Vec<Triple> {
2567 let mut v: Vec<Triple> = data
2568 .iter()
2569 .copied()
2570 .filter(|&(a, b, c)| {
2571 s.is_none_or(|x| x == a) && p.is_none_or(|x| x == b) && o.is_none_or(|x| x == c)
2572 })
2573 .collect();
2574 v.sort_unstable();
2575 v
2576 }
2577
2578 /// The tile-directory estimate orders patterns the way their true sizes
2579 /// do, never fetches anything, and is exact at its extremes: unbound is the
2580 /// total, a key no tile can hold is zero.
2581 #[test]
2582 fn tile_estimates_rank_patterns_by_size() {
2583 // Predicate 1 on every subject, predicate 2 on a few; object 50 is a
2584 // "class" with 4000 instances, object 60 one with 3.
2585 let mut b = GraphIndexBuilder::new().with_tile_budget(256);
2586 let mut n = 0u64;
2587 for s in 0..4000u32 {
2588 b.push((s, 1, 50));
2589 b.push((s, 3, 1000 + s));
2590 n += 2;
2591 if s % 1000 == 0 {
2592 b.push((s, 2, 60));
2593 n += 1;
2594 }
2595 }
2596 let idx = b.build();
2597 let est = |p: Pattern| idx.estimate_matches(p, n);
2598 assert_eq!(est((None, None, None)), n as f64);
2599 assert!(est((None, Some(2), None)) < est((None, Some(1), None)));
2600 assert!(est((None, None, Some(60))) < est((None, None, Some(50))));
2601 assert!(est((Some(7), None, None)) < est((None, Some(3), None)));
2602 // A big predicate spans many tiles; the rare one fits in at most one.
2603 let per_tile = n as f64 / idx.tile_sections()[0].len() as f64;
2604 assert!(est((None, Some(2), None)) <= 1.5 * per_tile);
2605 assert!(est((None, Some(1), None)) > 10.0 * per_tile);
2606 // A leading id beyond every tile's range is provably absent.
2607 assert_eq!(est((Some(99_999), None, None)), 0.0);
2608 // An empty index estimates nothing for a bound pattern.
2609 let empty = GraphIndexBuilder::new().build();
2610 assert_eq!(empty.estimate_matches((Some(1), None, None), 0), 0.0);
2611 }
2612
2613 #[test]
2614 fn every_pattern_shape_matches_reference() {
2615 let (idx, data) = graph();
2616 let vals = |opts: &[u32]| {
2617 let mut v: Vec<Option<u32>> = opts.iter().map(|&x| Some(x)).collect();
2618 v.push(None);
2619 v
2620 };
2621 // Exercise all 8 bound/unbound shapes across representative values.
2622 for s in vals(&[1, 2, 9]) {
2623 for p in vals(&[10, 11, 99]) {
2624 for o in vals(&[100, 300, 999]) {
2625 let pat = (s, p, o);
2626 assert_eq!(
2627 idx.match_pattern(pat),
2628 reference(&data, pat),
2629 "pattern {pat:?}"
2630 );
2631 // The lazy stream must match the eager result once sorted —
2632 // same triples, just without the up-front canonical re-sort.
2633 let mut streamed: Vec<Triple> = idx.scan_iter(pat).collect();
2634 streamed.sort_unstable();
2635 assert_eq!(streamed, reference(&data, pat), "scan_iter {pat:?}");
2636 }
2637 }
2638 }
2639 }
2640
2641 #[test]
2642 fn unbound_returns_everything_sorted() {
2643 let (idx, data) = graph();
2644 let mut sorted = data.clone();
2645 sorted.sort_unstable();
2646 assert_eq!(idx.match_pattern((None, None, None)), sorted);
2647 }
2648
2649 /// A tiny tile budget must split sections into many tiles — including
2650 /// MID-GROUP cuts (budget 1/16 makes every a-group oversized) — and every
2651 /// pattern shape must still match the brute-force reference: bound
2652 /// leading ids route to the run of covering tiles, unbound scans chain
2653 /// all tiles.
2654 #[test]
2655 fn multi_tile_sections_match_reference_every_shape() {
2656 // Enough distinct leading ids to split under a tiny budget.
2657 let mut data: Vec<Triple> = Vec::new();
2658 for s in 1..=40u32 {
2659 for p in [10u32, 11] {
2660 for o in [100u32, 100 + s] {
2661 data.push((s, p, o));
2662 }
2663 }
2664 }
2665 for budget in [1usize, 16, 64, 1 << 20] {
2666 let mut b = GraphIndexBuilder::new().with_tile_budget(budget);
2667 for &t in &data {
2668 b.push(t);
2669 }
2670 let idx = b.build();
2671 let spo_tiles = idx.tile_sections()[0].len();
2672 if budget <= 16 {
2673 assert!(spo_tiles > 1, "budget {budget} should force tiling");
2674 }
2675 // Tile ranges must be ascending; a split mega-group may leave
2676 // consecutive tiles SHARING a leading id (never overlapping past it).
2677 for w in idx.tile_sections()[0].as_full().unwrap().windows(2) {
2678 assert!(w[0].leading_range().1 <= w[1].leading_range().0);
2679 }
2680 assert_eq!(idx.triple_count() as usize, data.len(), "budget {budget}");
2681
2682 let vals = |opts: &[u32]| {
2683 let mut v: Vec<Option<u32>> = opts.iter().map(|&x| Some(x)).collect();
2684 v.push(None);
2685 v
2686 };
2687 for _round in 0..2 {
2688 for s in vals(&[1, 20, 40, 99]) {
2689 for p in vals(&[10, 11, 99]) {
2690 for o in vals(&[100, 120, 999]) {
2691 let pat = (s, p, o);
2692 assert_eq!(
2693 idx.match_pattern(pat),
2694 reference(&data, pat),
2695 "budget {budget} pattern {pat:?}"
2696 );
2697 }
2698 }
2699 }
2700 }
2701 }
2702 }
2703
2704 /// The Crossref shape: ONE mega-predicate carrying most of the graph. At
2705 /// the real 64 KiB budget its P-leading a-group must be split across
2706 /// multiple tiles (bounding builder memory), and a bound leading id must
2707 /// route to the whole run of covering tiles with complete results.
2708 #[test]
2709 fn mega_group_splits_across_tiles_and_lookups_stay_complete() {
2710 let hot_p = 7u32;
2711 let mut data: Vec<Triple> = Vec::new();
2712 for i in 0..40_000u32 {
2713 data.push((1_000 + i % 200, hot_p, 50_000 + i));
2714 }
2715 data.push((1, 1, 1));
2716 data.push((2, 2, 2));
2717 let mut b = GraphIndexBuilder::new(); // default INDEX_TILE_BUDGET
2718 for &t in &data {
2719 b.push(t);
2720 }
2721 let idx = b.build();
2722
2723 // PSO (a = predicate): the hot_p group must span several tiles.
2724 let pso = ALL_PERMS
2725 .iter()
2726 .position(|p| matches!(p, IndexPermutation::Pso))
2727 .unwrap();
2728 let covering = idx.tile_sections()[pso]
2729 .iter()
2730 .filter(|t| {
2731 let (lo, hi) = t.leading_range();
2732 lo <= hot_p && hot_p <= hi
2733 })
2734 .count();
2735 assert!(
2736 covering > 1,
2737 "expected the hot predicate split across tiles, got {covering}"
2738 );
2739
2740 // Bound-p lookup must still return every triple of the mega-group —
2741 // probing bound objects in the FIRST, MIDDLE and LAST slices of the
2742 // split run (a mid-run object once returned 0 on the first split file).
2743 for pat in [
2744 (None, Some(hot_p), None),
2745 (Some(1_050), Some(hot_p), None),
2746 (None, Some(hot_p), Some(50_123)), // first slice
2747 (None, Some(hot_p), Some(70_000)), // middle of the run
2748 (None, Some(hot_p), Some(89_999)), // last slice
2749 (None, Some(hot_p), Some(49_000)), // below the range → empty
2750 (None, Some(hot_p), Some(95_000)), // above the range → empty
2751 ] {
2752 assert_eq!(idx.match_pattern(pat), reference(&data, pat), "{pat:?}");
2753 }
2754 }
2755
2756 /// Scans must yield identical results across the directory lifecycle: the
2757 /// first leading-bound scan walks linearly, the second builds the
2758 /// directory, and later ones jump through it — including absent ids
2759 /// (between, below, and above the stored groups).
2760 #[test]
2761 fn directory_backed_scans_match_reference_every_shape() {
2762 let (idx, data) = graph();
2763 let vals = |opts: &[u32]| {
2764 let mut v: Vec<Option<u32>> = opts.iter().map(|&x| Some(x)).collect();
2765 v.push(None);
2766 v
2767 };
2768 for round in 0..3 {
2769 for s in vals(&[1, 2, 3, 0, 9]) {
2770 for p in vals(&[10, 11, 12, 99]) {
2771 for o in vals(&[100, 200, 300, 999]) {
2772 let pat = (s, p, o);
2773 let mut got: Vec<Triple> = idx.scan_iter(pat).collect();
2774 got.sort_unstable();
2775 assert_eq!(got, reference(&data, pat), "round {round} scan {pat:?}");
2776 }
2777 }
2778 }
2779 }
2780 }
2781
2782 /// A tile synopsis must let a routed scan **skip the fetch** of a tile its
2783 /// secondary-column range rules out — and never skip one that could match.
2784 #[test]
2785 fn synopsis_prunes_the_routed_tile_before_fetch() {
2786 use std::sync::atomic::{AtomicUsize, Ordering::SeqCst};
2787 use std::sync::Arc;
2788
2789 // One real SPO tile: subject a=5, predicates b∈{10,11}, object c=100.
2790 let block = {
2791 let mut b = TripleBlockBuilder::new();
2792 b.push((5, 10, 100));
2793 b.push((5, 11, 100));
2794 b.build()
2795 };
2796 let fetches = Arc::new(AtomicUsize::new(0));
2797 let (blk, fc) = (block.clone(), fetches.clone());
2798 // The loader returns the (already-decompressed) tile image and counts calls.
2799 let loader: TileLoader = Box::new(move |_si, _ti| {
2800 fc.fetch_add(1, SeqCst);
2801 Some(blk.clone())
2802 });
2803 // Remote SPO directory: one tile, leading a∈[5,5], synopsis b∈[10,11], c∈[100,100].
2804 let dirs = [
2805 vec![(5u32, 5u32, Some((10u32, 11u32, 100u32, 100u32)))],
2806 Vec::new(),
2807 Vec::new(),
2808 Vec::new(),
2809 Vec::new(),
2810 Vec::new(),
2811 ];
2812 let idx = GraphIndex::from_remote_directories(dirs, PermSet::ALL, loader);
2813
2814 // Predicate 99 is outside the tile's b-range → prune, zero fetches, empty.
2815 assert!(idx.match_pattern((Some(5), Some(99), None)).is_empty());
2816 assert_eq!(fetches.load(SeqCst), 0, "synopsis must skip the fetch");
2817
2818 // Object 999 outside the tile's c-range → prune, still zero fetches.
2819 assert!(idx.match_pattern((Some(5), None, Some(999))).is_empty());
2820 assert_eq!(
2821 fetches.load(SeqCst),
2822 0,
2823 "secondary-c prune also skips the fetch"
2824 );
2825
2826 // Predicate 10 is in range → the tile is fetched and the match returned.
2827 assert_eq!(
2828 idx.match_pattern((Some(5), Some(10), None)),
2829 vec![(5, 10, 100)]
2830 );
2831 assert_eq!(
2832 fetches.load(SeqCst),
2833 1,
2834 "an admissible secondary still fetches"
2835 );
2836 }
2837
2838 /// Without a synopsis (`None`), nothing is pruned early — the tile is always
2839 /// fetched and the in-tile zone map does the (correct) pruning, as before.
2840 #[test]
2841 fn absent_synopsis_never_prunes() {
2842 use std::sync::atomic::{AtomicUsize, Ordering::SeqCst};
2843 use std::sync::Arc;
2844 let block = {
2845 let mut b = TripleBlockBuilder::new();
2846 b.push((5, 10, 100));
2847 b.build()
2848 };
2849 let fetches = Arc::new(AtomicUsize::new(0));
2850 let (blk, fc) = (block.clone(), fetches.clone());
2851 let loader: TileLoader = Box::new(move |_si, _ti| {
2852 fc.fetch_add(1, SeqCst);
2853 Some(blk.clone())
2854 });
2855 let dirs = [
2856 vec![(5u32, 5u32, None)],
2857 Vec::new(),
2858 Vec::new(),
2859 Vec::new(),
2860 Vec::new(),
2861 Vec::new(),
2862 ];
2863 let idx = GraphIndex::from_remote_directories(dirs, PermSet::ALL, loader);
2864 // Predicate 99 absent, but with no synopsis the tile is fetched (then the
2865 // zone map yields no match) — correctness preserved, just no fetch saved.
2866 assert!(idx.match_pattern((Some(5), Some(99), None)).is_empty());
2867 assert_eq!(fetches.load(SeqCst), 1, "no synopsis ⇒ no early prune");
2868 }
2869
2870 /// The eight triple-pattern shapes, as `(s, p, o)` boundness.
2871 fn all_patterns() -> Vec<Pattern> {
2872 (0..8u8)
2873 .map(|mask| {
2874 (
2875 (mask & 1 != 0).then_some(1u32),
2876 (mask & 2 != 0).then_some(2u32),
2877 (mask & 4 != 0).then_some(3u32),
2878 )
2879 })
2880 .collect()
2881 }
2882
2883 /// The claim the whole option rests on: **{SPO, POS, OSP} ties the longest
2884 /// bound prefix on all eight pattern shapes**, so the other three never win
2885 /// routing — only sort order. Enumerated and printed rather than asserted in
2886 /// prose, because if it were ever false a lean file would answer a pattern
2887 /// from a worse index without anything failing.
2888 #[test]
2889 fn perm_routing_never_leaves_core() {
2890 for pat in all_patterns() {
2891 let best_all = ALL_PERMS
2892 .iter()
2893 .map(|p| p.leading_bound(pat))
2894 .max()
2895 .unwrap();
2896 let best_core = PermSet::CORE
2897 .iter()
2898 .map(|p| p.leading_bound(pat))
2899 .max()
2900 .unwrap();
2901 let scores: Vec<String> = ALL_PERMS
2902 .iter()
2903 .map(|p| format!("{}={}", p.name(), p.leading_bound(pat)))
2904 .collect();
2905 eprintln!(
2906 "s={} p={} o={} | {} | best6={best_all} best3={best_core} chosen={}",
2907 pat.0.is_some() as u8,
2908 pat.1.is_some() as u8,
2909 pat.2.is_some() as u8,
2910 scores.join(" "),
2911 GraphIndex::best_permutation(pat).name(),
2912 );
2913 assert_eq!(
2914 best_core, best_all,
2915 "the three routing permutations must tie the best of six on {pat:?}"
2916 );
2917 let chosen = GraphIndex::best_permutation(pat);
2918 assert!(
2919 PermSet::CORE.contains(chosen),
2920 "best_permutation chose {} for {pat:?}, outside the routing three",
2921 chosen.name()
2922 );
2923 assert_eq!(
2924 chosen,
2925 GraphIndex::best_permutation_in(PermSet::CORE, pat),
2926 "routing must be identical with three permutations and with six"
2927 );
2928 }
2929 }
2930
2931 /// Restricting the set may never make a merge join pick a *worse-routed*
2932 /// stream than the plain lookup would use: with six the co-sorted
2933 /// permutation always ties the best prefix, and with three the guard
2934 /// declines rather than buying sort order with a whole-section scan.
2935 #[test]
2936 fn sorted_permutation_never_sacrifices_routing() {
2937 let mut declined = Vec::new();
2938 for pat in all_patterns() {
2939 for col in 0..3usize {
2940 let routed_all = GraphIndex::best_permutation(pat).leading_bound(pat);
2941 if let Some(p) = GraphIndex::permutation_sorted_on(pat, col) {
2942 assert_eq!(
2943 p.leading_bound(pat),
2944 routed_all,
2945 "six permutations must always co-sort at the best prefix"
2946 );
2947 assert_eq!(p.roles()[p.leading_bound(pat)], col);
2948 }
2949 match GraphIndex::permutation_sorted_on_in(PermSet::CORE, pat, col) {
2950 Some(p) => {
2951 assert!(PermSet::CORE.contains(p));
2952 assert_eq!(
2953 p.leading_bound(pat),
2954 GraphIndex::best_permutation_in(PermSet::CORE, pat).leading_bound(pat)
2955 );
2956 }
2957 None => {
2958 if GraphIndex::permutation_sorted_on(pat, col).is_some() {
2959 declined.push((pat, col));
2960 }
2961 }
2962 }
2963 }
2964 }
2965 // Exactly the three shapes SOP / PSO / OPS exist for: one bound
2966 // component, sorted on a column no remaining order can lead with after
2967 // it — subject-bound sorted on object, predicate-bound sorted on
2968 // subject, object-bound sorted on predicate.
2969 for (pat, col) in &declined {
2970 let bound: Vec<usize> = (0..3)
2971 .filter(|&i| [pat.0, pat.1, pat.2][i].is_some())
2972 .collect();
2973 assert_eq!(
2974 bound.len(),
2975 1,
2976 "only single-bound shapes lose a merge order"
2977 );
2978 assert_ne!(bound[0], *col);
2979 }
2980 eprintln!("three-permutation merge-join declines: {declined:?}");
2981 assert_eq!(declined.len(), 3, "exactly three (bound, sort) shapes lost");
2982 }
2983
2984 #[test]
2985 fn perm_set_masks_and_positions() {
2986 assert_eq!(PermSet::ALL.len(), 6);
2987 assert_eq!(PermSet::CORE.len(), 3);
2988 assert_eq!(PermSet::CORE.bits(), 0b0000_0111);
2989 assert_eq!(PermSet::CORE.names(), vec!["SPO", "POS", "OSP"]);
2990 assert_eq!(
2991 PermSet::ALL.names(),
2992 vec!["SPO", "POS", "OSP", "SOP", "PSO", "OPS"]
2993 );
2994 // Container position is rank among the STORED permutations, which is
2995 // why the routing three had to lead ALL_PERMS.
2996 for (i, p) in PermSet::CORE.iter().enumerate() {
2997 assert_eq!(PermSet::CORE.position(p), Some(i));
2998 assert_eq!(p.section_index(), i);
2999 }
3000 assert_eq!(PermSet::CORE.position(IndexPermutation::Sop), None);
3001 assert_eq!(PermSet::ALL.position(IndexPermutation::Ops), Some(5));
3002 // Only supersets of the routing three are legal.
3003 assert!(PermSet::from_bits(0b0000_0111).is_ok());
3004 assert!(PermSet::from_bits(0b0011_1111).is_ok());
3005 assert!(PermSet::from_bits(0b0000_1111).is_ok());
3006 assert!(PermSet::from_bits(0b0000_0011).is_err());
3007 assert!(PermSet::from_bits(0b0011_1000).is_err());
3008 assert!(PermSet::from_bits(0b1000_0111).is_err());
3009 }
3010
3011 /// A lean index answers every pattern shape with exactly the rows a full
3012 /// one does — the property that makes the option safe at all.
3013 #[test]
3014 fn three_and_six_agree_on_every_pattern() {
3015 let triples: Vec<Triple> = (0..200u32)
3016 .map(|i| (i % 17 + 1, i % 5 + 1, i % 23 + 1))
3017 .collect();
3018 let six = GraphIndexBuilder::from_triples(triples.clone())
3019 .with_tile_budget(128)
3020 .build();
3021 let three = GraphIndexBuilder::from_triples(triples)
3022 .with_tile_budget(128)
3023 .with_perms(PermSet::CORE)
3024 .build();
3025 assert_eq!(six.perms(), PermSet::ALL);
3026 assert_eq!(three.perms(), PermSet::CORE);
3027 for perm in [
3028 IndexPermutation::Sop,
3029 IndexPermutation::Pso,
3030 IndexPermutation::Ops,
3031 ] {
3032 assert!(three.tile_sections()[perm.section_index()].is_empty());
3033 }
3034 for a in [None, Some(3u32)] {
3035 for b in [None, Some(2u32)] {
3036 for c in [None, Some(9u32)] {
3037 let mut x = six.match_pattern((a, b, c));
3038 let mut y = three.match_pattern((a, b, c));
3039 x.sort_unstable();
3040 y.sort_unstable();
3041 assert_eq!(x, y, "pattern {:?} disagreed", (a, b, c));
3042 }
3043 }
3044 }
3045 }
3046}