1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
use std::path::Path;
use anyhow::{Context, Result, anyhow};
use tantivy::collector::TopDocs;
use tantivy::directory::MmapDirectory;
use tantivy::query::QueryParser;
use tantivy::schema::{Field, STORED, STRING, Schema, TEXT, Value};
use tantivy::{Index, IndexReader, IndexWriter, ReloadPolicy, TantivyDocument, Term, doc};
use uuid::Uuid;
/// `ReadWrite` indexes may construct an `IndexWriter`, which takes a
/// cross-process exclusive lock file. `Reader` indexes never do: they exist
/// so a multi-tenant service can search beside a writing daemon, or beside
/// other readers, without contending for that lock and without risking the
/// zero-byte `.tantivy-writer.lock` a SIGKILLed writer can leave behind.
#[derive(Clone, Copy, PartialEq, Eq)]
enum LexicalMode {
ReadWrite,
Reader,
}
pub struct LexicalIndex {
index: Index,
// Our own handle on the same directory the index was opened from, kept
// solely so `evict_dead_mmap_entries` can reach the mmap cache: `Index`
// hands back a `ManagedDirectory`, which does not expose the concrete
// `MmapDirectory` underneath. Cloning shares one `Arc`, so this is the
// same cache the index uses, not a second one.
directory: MmapDirectory,
// Created on first mutation, not at open: the tantivy writer holds an
// exclusive cross-process lock, and read-only consumers (recall, eval)
// must be able to open the index beside a writing daemon.
writer: Option<IndexWriter>,
reader: IndexReader,
id_field: Field,
body_field: Field,
mode: LexicalMode,
}
impl LexicalIndex {
pub fn open(dir: &Path) -> Result<Self> {
std::fs::create_dir_all(dir)
.with_context(|| format!("creating tantivy dir {}", dir.display()))?;
let mut builder = Schema::builder();
let id_field = builder.add_text_field("memory_id", STRING | STORED);
let body_field = builder.add_text_field("body", TEXT);
let schema = builder.build();
let directory = MmapDirectory::open(dir)?;
let index =
Index::open_or_create(directory.clone(), schema).context("opening tantivy index")?;
let reader = index.reader().context("creating tantivy reader")?;
Ok(Self {
index,
directory,
writer: None,
reader,
id_field,
body_field,
mode: LexicalMode::ReadWrite,
})
}
/// Opens an existing index purely for search: no `IndexWriter` is ever
/// constructed here, so this never touches the writer lock file and can
/// sit beside a writing daemon, or other readers, indefinitely. Errors
/// instead of creating an index, because a reader that can create a
/// tenant's index is a reader that can corrupt one.
///
/// The reader uses `ReloadPolicy::Manual`: a service holding thousands
/// of these open cannot afford a background filesystem-watch thread per
/// tenant, so visibility of writer changes is pulled explicitly via
/// `reload_reader`, not pushed on a delay.
pub fn open_reader(dir: &Path) -> Result<Self> {
let mmap_dir = MmapDirectory::open(dir)
.with_context(|| format!("opening tantivy directory {}", dir.display()))?;
let index = Index::open(mmap_dir.clone()).context("opening tantivy index")?;
let schema = index.schema();
let id_field = schema
.get_field("memory_id")
.context("tantivy schema missing memory_id field")?;
let body_field = schema
.get_field("body")
.context("tantivy schema missing body field")?;
let reader = index
.reader_builder()
.reload_policy(ReloadPolicy::Manual)
.try_into()
.context("creating tantivy reader")?;
Ok(Self {
index,
directory: mmap_dir,
writer: None,
reader,
id_field,
body_field,
mode: LexicalMode::Reader,
})
}
/// Drops the cache entries tantivy keeps for files it has already unmapped.
///
/// `MmapDirectory` memoizes one `Weak` per file it has ever mmapped, keyed
/// by path. `MmapDirectory::delete` documents that it evicts a file's entry
/// before deleting the file, but it does not: it only unlinks. Nothing else
/// evicts either, because the eviction pass (`MmapCache::remove_weak_ref`)
/// is reachable only through `get_cache_info`, which tantivy never calls
/// itself. Every merged-away segment therefore leaves its path, its `Weak`,
/// and the retained allocation of its dead `Arc` behind for the life of the
/// process.
///
/// That is unbounded here, because a daemon draining the index outbox
/// commits every few seconds forever and each commit writes a fresh
/// segment: measured at roughly 7 KB retained per commit against a
/// fixed-size corpus, which is what OOM-killed the API server. Sweeping
/// after every commit and every reader reload bounds the cache to the
/// segments actually mapped.
///
/// `get_cache_info` is called for that eviction, not for what it returns.
fn evict_dead_mmap_entries(&self) {
let _ = self.directory.get_cache_info();
}
fn writer_mut(&mut self) -> Result<&mut IndexWriter> {
if self.mode == LexicalMode::Reader {
return Err(anyhow!(
"tantivy index is a read-only reader; mutation is not permitted"
));
}
if self.writer.is_none() {
self.writer = Some(
self.index
.writer(50_000_000)
.context("creating tantivy writer")?,
);
}
Ok(self.writer.as_mut().unwrap())
}
/// Delete-then-add so a replayed outbox row cannot duplicate a document.
pub fn upsert(&mut self, id: Uuid, body: &str) -> Result<()> {
let id_field = self.id_field;
let body_field = self.body_field;
let id_str = id.to_string();
let writer = self.writer_mut()?;
writer.delete_term(Term::from_field_text(id_field, &id_str));
writer.add_document(doc!(id_field => id_str, body_field => body))?;
Ok(())
}
pub fn remove(&mut self, id: Uuid) -> Result<()> {
let id_field = self.id_field;
let id_str = id.to_string();
self.writer_mut()?
.delete_term(Term::from_field_text(id_field, &id_str));
Ok(())
}
pub fn commit(&mut self) -> Result<()> {
if self.mode == LexicalMode::Reader {
return Err(anyhow!(
"tantivy index is a read-only reader; commit is not permitted"
));
}
if let Some(writer) = self.writer.as_mut() {
writer.commit().context("tantivy commit")?;
self.reader.reload().context("tantivy reader reload")?;
self.evict_dead_mmap_entries();
}
Ok(())
}
pub(crate) fn close_writer(&mut self) -> Result<()> {
if let Some(writer) = self.writer.take() {
writer
.wait_merging_threads()
.context("waiting for tantivy merges")?;
}
Ok(())
}
#[cfg(test)]
pub(crate) fn segment_count(&self) -> usize {
self.index
.searchable_segment_ids()
.map_or(0, |ids| ids.len())
}
/// Commits, then merges every segment into one and deletes the files the
/// merge replaced. A deleted document's terms stay in its segment's files
/// until that segment is merged away, and the merge policy may never pick
/// a large old segment, so this is how a removed text leaves the disk.
pub fn compact(&mut self) -> Result<()> {
self.commit()?;
if let Some(writer) = self.writer.take() {
writer
.wait_merging_threads()
.context("waiting for tantivy merges")?;
}
let segments = self
.index
.searchable_segment_ids()
.context("listing tantivy segments")?;
let writer = self.writer_mut()?;
if !segments.is_empty() {
writer.merge(&segments).wait().context("tantivy merge")?;
}
writer
.garbage_collect_files()
.wait()
.context("tantivy garbage collection")?;
self.reader.reload().context("tantivy reader reload")?;
self.evict_dead_mmap_entries();
Ok(())
}
/// Returns (memory id, BM25 score) pairs, best first.
pub fn search(&self, query_text: &str, k: usize) -> Result<Vec<(Uuid, f32)>> {
let searcher = self.reader.searcher();
let parser = QueryParser::for_index(&self.index, vec![self.body_field]);
// Lenient parse: user queries are arbitrary text, not tantivy syntax,
// and a stray `(` must degrade to a best-effort query, not an error.
let (query, _errors) = parser.parse_query_lenient(query_text);
let top = searcher.search(&query, &TopDocs::with_limit(k.max(1)).order_by_score())?;
let mut out = Vec::with_capacity(top.len());
for (score, address) in top {
let document: TantivyDocument = searcher.doc(address)?;
let id_str = document
.get_first(self.id_field)
.and_then(|v| v.as_str())
.unwrap_or_default();
if let Ok(id) = Uuid::parse_str(id_str) {
out.push((id, score));
}
}
Ok(out)
}
pub fn num_docs(&self) -> u64 {
self.reader.searcher().num_docs()
}
/// Decompressed doc-store blocks the current searcher has cached, each
/// counted at the configured block size, which a block passes by at most
/// one document. Everything else a reader searches is the mapped segment
/// files, and the writer's indexing arenas have no public size.
pub fn heap_bytes(&self) -> u64 {
let blocks = self.reader.searcher().doc_store_cache_stats().num_entries;
(blocks * self.index.settings().docstore_blocksize) as u64
}
/// Reloads the reader so it observes segments committed since it was
/// opened. Harmless on a `ReadWrite` instance: `commit` already reloads,
/// so this just repeats an already-current reload.
pub fn reload_reader(&mut self) -> Result<()> {
self.reader.reload().context("tantivy reader reload")?;
self.evict_dead_mmap_entries();
Ok(())
}
/// Drop every document; used by full rebuild.
pub fn reset(&mut self) -> Result<()> {
self.writer_mut()?
.delete_all_documents()
.context("tantivy delete_all_documents")?;
self.commit()
}
}