spg_engine/maintenance.rs
1//! Table-maintenance executors: `ANALYZE` (re-stat) and
2//! `COMPACT COLD SEGMENTS` (cold-segment merge). Split out of
3//! `lib.rs` (cut 19); verbatim move, the only edit beyond
4//! visibility is reuniting the `exec_compact_cold_segments` doc,
5//! whose first half had drifted above `set_session_param`.
6
7use alloc::string::{String, ToString};
8use alloc::vec::Vec;
9
10use spg_storage::{ColumnSchema, CompactReport, DataType, IndexKind, Row, StorageError, Value};
11
12use crate::{
13 COMPACTION_TARGET_DEFAULT_BYTES, Engine, EngineError, QueryResult, canonical_value_repr,
14 is_internal_table_name, sort_values_for_histogram, statistics,
15};
16
17impl Engine {
18 /// v6.2.0 — `ANALYZE [<table>]` runtime. Bare `ANALYZE` walks
19 /// every user table; `ANALYZE <name>` re-stats one. For each
20 /// target table, single-pass scan + per-column histogram +
21 /// `null_frac` + `n_distinct`. Replaces the table's prior
22 /// stats; resets the modified-row counter.
23 ///
24 /// v6.2.0 doesn't sample — it scans the full table. v6.2.x
25 /// can add reservoir sampling at the > 100 K-row mark; not a
26 /// scope blocker for the current commit since rows ≤ 100 K
27 /// analyse in milliseconds.
28 /// v7.39 (round 320, V53) — `DISCARD { ALL | PLANS | SEQUENCES | TEMP }`.
29 /// pgbouncer sends `DISCARD ALL` between pooled client sessions so the
30 /// next client sees a clean connection. It used to be swallowed as
31 /// "dump noise" on the theory that SPG had no per-connection state
32 /// worth discarding — untrue since round 279 gave every connection its
33 /// own session bag (GUC overrides, prepared statements, large-object
34 /// descriptors). A no-op meant one pooled client's state leaked to the
35 /// next one.
36 ///
37 /// PG 18.4 measured: `SET application_name='x'; PREPARE p …;
38 /// DISCARD ALL` leaves application_name back at its startup value and
39 /// `EXECUTE p` failing with "prepared statement \"p\" does not
40 /// exist"; inside a transaction block it is
41 /// `ERROR: DISCARD ALL cannot run inside a transaction block`.
42 ///
43 /// v7.39 (round 321, V54) — cursors are closed here too, as PG's
44 /// DISCARD ALL does. Round 320 had to leave them alone because the
45 /// cursor table was process-wide: closing "all" from one connection
46 /// would have closed another's. They live in the session bag now.
47 pub(crate) fn exec_discard(
48 &mut self,
49 target: spg_sql::ast::DiscardTarget,
50 ) -> Result<QueryResult, EngineError> {
51 use spg_sql::ast::DiscardTarget;
52 if matches!(target, DiscardTarget::All | DiscardTarget::Temp)
53 && self.current_tx.is_some_and(|tx| self.is_tx_open(tx))
54 {
55 return Err(EngineError::Unsupported(alloc::format!(
56 "DISCARD {target} cannot run inside a transaction block"
57 )));
58 }
59 match target {
60 DiscardTarget::All => {
61 self.reset_all_gucs();
62 self.prepared_statements.clear();
63 self.lo_descriptors.clear();
64 self.lo_next_fd = 0;
65 self.cursors.clear();
66 self.listen_channels.clear();
67 self.plan_cache.clear();
68 self.refresh_render_style();
69 }
70 DiscardTarget::Plans => self.plan_cache.clear(),
71 // SPG has neither temp tables nor per-session sequence state
72 // (`currval` reads the catalog), so there is nothing of either
73 // kind to throw away. Accepted so a client's sequence of
74 // DISCARDs runs, and tagged with the target it named.
75 DiscardTarget::Sequences | DiscardTarget::Temp => {}
76 }
77 Ok(QueryResult::CommandOk {
78 affected: 0,
79 modified_catalog: false,
80 })
81 }
82
83 pub(crate) fn exec_analyze(
84 &mut self,
85 target: Option<&str>,
86 ) -> Result<QueryResult, EngineError> {
87 // v7.38 元机制 D acceptor — `SPG_TEST_STATS_FROZEN=1` turns
88 // ANALYZE into a no-op so a test's statistic-version snapshot
89 // is stable across sessions. Pairs with the plan-cache
90 // version-aware invalidation: freezing stats also freezes
91 // plan reuse, which is what regression tests want when
92 // proving "same SQL, same plan".
93 if self.env_cfg().stats_frozen {
94 return Ok(QueryResult::CommandOk {
95 affected: 0,
96 modified_catalog: false,
97 });
98 }
99 let names: Vec<String> = if let Some(name) = target {
100 // Verify the table exists; surface a clear error if not.
101 if self.catalog.get(name).is_none() {
102 return Err(EngineError::Storage(StorageError::TableNotFound {
103 name: name.to_string(),
104 }));
105 }
106 alloc::vec![name.to_string()]
107 } else {
108 self.catalog
109 .table_names()
110 .into_iter()
111 .filter(|n| !is_internal_table_name(n))
112 .collect()
113 };
114 let mut analysed = 0usize;
115 let now_us = self.clock.map(|f| f());
116 for table_name in &names {
117 self.analyze_one_table(table_name)?;
118 // v7.39 (pg_stat knife C) — stamp last_analyze.
119 if let Some(us) = now_us
120 && let Some(t) = self.catalog.get_mut(table_name)
121 {
122 t.stamp_analyze(us);
123 }
124 analysed += 1;
125 }
126 // v6.3.1 — plan cache invalidation. Bump stats version so
127 // future lookups see the new generation, and selectively
128 // evict every plan whose `source_tables` overlap with the
129 // ANALYZE target set. Bare ANALYZE (all tables) clears the
130 // whole cache.
131 if analysed > 0 {
132 self.statistics.bump_version();
133 if target.is_some() {
134 for t in &names {
135 self.plan_cache.evict_referencing(t);
136 }
137 } else {
138 self.plan_cache.clear();
139 }
140 }
141 Ok(QueryResult::CommandOk {
142 affected: analysed,
143 modified_catalog: true,
144 })
145 }
146
147 /// Walk a single table's rows once and (re-)populate per-column
148 /// stats. Drops the existing stats for `table` first so columns
149 /// that have been DROP-ed between ANALYZEs don't leave stale
150 /// rows.
151 fn analyze_one_table(&mut self, table_name: &str) -> Result<(), EngineError> {
152 let table = self.catalog.get(table_name).ok_or_else(|| {
153 EngineError::Storage(StorageError::TableNotFound {
154 name: table_name.to_string(),
155 })
156 })?;
157 let schema = table.schema().clone();
158 let row_count = table.rows().len();
159 // For each column, collect (sorted) non-NULL textual values
160 // + count NULLs; then ask `statistics::build_histogram` to
161 // produce the 101 bounds and `estimate_n_distinct` the
162 // distinct count.
163 self.statistics.clear_table(table_name);
164 for (col_pos, col_schema) in schema.columns.iter().enumerate() {
165 // v6.2.0 skip: vector columns have their own stats
166 // shape (HNSW graph topology). v6.2 deliberation #1.
167 if matches!(col_schema.ty, DataType::Vector { .. }) {
168 continue;
169 }
170 let mut non_null_values: Vec<Value<'static>> = Vec::with_capacity(row_count);
171 let mut nulls: u64 = 0;
172 for row in table.rows() {
173 match row.values.get(col_pos) {
174 Some(Value::Null) | None => nulls += 1,
175 Some(v) => non_null_values.push(v.clone()),
176 }
177 }
178 // Sort by type-aware ordering (Int as int, Text as
179 // lex, etc.) so histogram bounds reflect the column's
180 // natural order — not lexicographic on the string
181 // representation, which would put "9" after "49".
182 non_null_values.sort_by(|a, b| sort_values_for_histogram(a, b));
183 let non_null: Vec<String> = non_null_values.iter().map(canonical_value_repr).collect();
184 let null_frac = if row_count == 0 {
185 0.0
186 } else {
187 #[allow(clippy::cast_precision_loss)]
188 let f = nulls as f32 / row_count as f32;
189 f
190 };
191 let n_distinct = statistics::estimate_n_distinct(&non_null);
192 let histogram_bounds = statistics::build_histogram(&non_null);
193 self.statistics.set(
194 table_name.to_string(),
195 col_schema.name.clone(),
196 statistics::ColumnStats {
197 null_frac,
198 n_distinct,
199 histogram_bounds,
200 },
201 );
202 }
203 self.statistics.reset_modified(table_name);
204 // v6.7.0 — refresh the per-table cold_rows cache. Walk the
205 // BTree indices and count Cold locators (MAX across
206 // indices); store the result on the table. Surfaced via
207 // `spg_statistic.cold_row_count` (new column) and
208 // `spg_stat_segment.table_name` (new column).
209 let cold_count = {
210 let table = self
211 .active_catalog()
212 .get(table_name)
213 .expect("table still present");
214 table.count_cold_locators()
215 };
216 let table_mut = self
217 .active_catalog_mut()
218 .get_mut(table_name)
219 .expect("table still present");
220 table_mut.set_cold_row_count(cold_count);
221 Ok(())
222 }
223
224 /// v6.7.3 — `COMPACT COLD SEGMENTS` runtime path. Drives the
225 /// engine-layer compaction shim with the default
226 /// 4 MiB segment-size threshold. spg-server intercepts the
227 /// SQL before it reaches the engine on a server build —
228 /// it reads `SPG_COMPACTION_TARGET_SEGMENT_BYTES`, calls
229 /// `Engine::compact_cold_segments_with_target` directly with
230 /// the env value, and persists every merged segment to
231 /// `<db>.spg/segments/`. This arm only fires for engine-only
232 /// callers (spg-embedded, lib tests); in that mode merged
233 /// segments live in memory and are dropped at process exit.
234 /// v7.37.17 (17.6 sibling) — `TRUNCATE [TABLE] <t>[, ...]
235 /// [RESTART IDENTITY]`. Clears every row from each named
236 /// table by dispatching to `Table::truncate()` (already exists
237 /// for internal callers). RESTART IDENTITY additionally resets
238 /// the table's associated sequence back to its start value.
239 pub(crate) fn exec_truncate(
240 &mut self,
241 tables: &[String],
242 _restart_identity: bool,
243 only: bool,
244 ) -> Result<QueryResult, EngineError> {
245 // RESTART IDENTITY is parsed but not honored yet — the
246 // SequenceDef doesn't expose a restart primitive on the
247 // storage side today. Accepted-and-no-op for pg_dump compat;
248 // real sequence reset lands with the sequence-lifecycle epic.
249 //
250 // v7.39 (round 647) — `ONLY` and the descent it turns off.
251 // Measured on PG18: a plain TRUNCATE of a parent empties its
252 // children too; `TRUNCATE ONLY <inheritance parent>` empties the
253 // parent alone; and `TRUNCATE ONLY <partitioned parent>` is not
254 // a no-op but an error, because a partitioned parent holds
255 // nothing and the spelling can only mean a mistake.
256 //
257 // The whole expansion sits inside this branch rather than beside
258 // it — rounds 641/643/644/646 each measured what an extra test
259 // in a function's body costs when the row loop is downstream.
260 let mut targets: Vec<String> = tables.to_vec();
261 if !tables.is_empty() {
262 for name in tables {
263 if only {
264 if crate::partition::is_partition_parent(self.active_catalog(), name) {
265 return Err(EngineError::Unsupported(alloc::format!(
266 "cannot truncate only a partitioned table"
267 )));
268 }
269 continue;
270 }
271 let mut frontier = alloc::vec![name.clone()];
272 while let Some(cur) = frontier.pop() {
273 for kid in crate::partition::children_of_parent(self.active_catalog(), &cur) {
274 frontier.push(kid.clone());
275 targets.push(kid);
276 }
277 }
278 }
279 }
280 let mut affected: usize = 0;
281 for name in &targets {
282 let cat = self.active_catalog_mut();
283 let Some(t) = cat.get_mut(name) else {
284 // v7.39 (read01 round 50) — PG says "relation" for TRUNCATE
285 // (only DROP TABLE says "table"). 42P01 at the wire.
286 return Err(EngineError::Storage(StorageError::Corrupt(alloc::format!(
287 "relation {name:?} does not exist"
288 ))));
289 };
290 affected = affected.saturating_add(t.row_count());
291 t.truncate();
292 }
293 Ok(QueryResult::CommandOk {
294 affected,
295 modified_catalog: false,
296 })
297 }
298
299 pub(crate) fn exec_compact_cold_segments(&mut self) -> Result<QueryResult, EngineError> {
300 let target = COMPACTION_TARGET_DEFAULT_BYTES;
301 let reports = self.compact_cold_segments_with_target(target)?;
302 let columns = alloc::vec![
303 ColumnSchema::new("table_name", DataType::Text, false),
304 ColumnSchema::new("index_name", DataType::Text, false),
305 ColumnSchema::new("sources_merged", DataType::BigInt, false),
306 ColumnSchema::new("merged_segment_id", DataType::BigInt, false),
307 ColumnSchema::new("merged_rows", DataType::BigInt, false),
308 ColumnSchema::new("deleted_rows_pruned", DataType::BigInt, false),
309 ColumnSchema::new("bytes_reclaimed_estimate", DataType::BigInt, false),
310 ];
311 let rows: Vec<Row<'static>> = reports
312 .into_iter()
313 .map(|(tname, iname, report)| {
314 Row::new(alloc::vec![
315 Value::text(tname),
316 Value::text(iname),
317 Value::BigInt(i64::try_from(report.sources.len()).unwrap_or(i64::MAX)),
318 Value::BigInt(i64::from(report.merged_segment_id.unwrap_or(0))),
319 Value::BigInt(i64::try_from(report.merged_rows).unwrap_or(i64::MAX)),
320 Value::BigInt(i64::try_from(report.deleted_rows_pruned).unwrap_or(i64::MAX),),
321 Value::BigInt(
322 i64::try_from(report.bytes_reclaimed_estimate).unwrap_or(i64::MAX),
323 ),
324 ])
325 })
326 .collect();
327 Ok(QueryResult::Rows { columns, rows })
328 }
329
330 /// v6.7.3 — public shim around `Catalog::compact_cold_segments`
331 /// driving every BTree index on every user table. Returns one
332 /// `(table, index, report)` triple for each merge that
333 /// actually happened (no-op (table, index) pairs are filtered
334 /// out so callers can size persist-side work to the live
335 /// merges). Caller is responsible for persisting each
336 /// `report.merged_segment_bytes` and updating the on-disk
337 /// segment registry; engine layer is no_std and never
338 /// touches disk.
339 ///
340 /// Marks every touched table's cached `cold_row_count` stale
341 /// — compaction GC'd some shadowed rows, so the count must be
342 /// re-derived on the next ANALYZE.
343 pub fn compact_cold_segments_with_target(
344 &mut self,
345 target_segment_bytes: u64,
346 ) -> Result<Vec<(String, String, CompactReport)>, EngineError> {
347 let table_names = self.active_catalog().table_names();
348 let mut reports: Vec<(String, String, CompactReport)> = Vec::new();
349 for tname in table_names {
350 if is_internal_table_name(&tname) {
351 continue;
352 }
353 let idx_names: Vec<String> = {
354 let Some(t) = self.active_catalog().get(&tname) else {
355 continue;
356 };
357 t.indices()
358 .iter()
359 .filter(|i| matches!(i.kind, IndexKind::BTree(_)))
360 .map(|i| i.name.clone())
361 .collect()
362 };
363 for iname in idx_names {
364 let report = self
365 .active_catalog_mut()
366 .compact_cold_segments(&tname, &iname, target_segment_bytes)
367 .map_err(EngineError::Storage)?;
368 if report.merged_segment_id.is_some() {
369 if let Some(t) = self.active_catalog_mut().get_mut(&tname) {
370 t.mark_cold_row_count_stale();
371 }
372 reports.push((tname.clone(), iname, report));
373 }
374 }
375 }
376 Ok(reports)
377 }
378}