Skip to main content

ic_db/store/query/aggregate/
mod.rs

1//! Bounded aggregates over the shared scanner and exact reducers, without row history.
2pub(super) mod jobs;
3mod numeric;
4
5use super::{
6    Comparison, MAX_COMPARISON_KEY, MAX_DISTINCT_VALUES, MAX_STATE_BYTES, MAX_TERMS, Plan, Query,
7    QueryPage, admit_ordered_fields, admit_output, admit_projection, comparison_key,
8    copy_projection, execute, projection_size,
9};
10use crate::{
11    Entity, EntitySchema, Error, FieldType, IndexValue, TypeSchema, Value, ValueSchema,
12    codec::Decoder, key,
13};
14use ic_storage::Memory;
15use std::collections::{BTreeMap, BTreeSet};
16
17const MAX_GROUPS: usize = 32;
18
19/// Completed-group ordering. Aggregate ties always use ascending canonical keys.
20#[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
21pub enum GroupOrder {
22    /// Ascending group-key tuple, in canonical field-name order (the default).
23    #[default]
24    Key,
25    /// Ascending aggregate value; absent values precede present values.
26    AggregateAscending,
27    /// Descending aggregate value; absent values follow present values.
28    AggregateDescending,
29    /// Descending group-key tuple, in canonical field-name order; absence sorts last.
30    KeyDescending,
31}
32
33#[derive(Default)]
34pub(super) struct GroupRequest<'a> {
35    // The final bool joins this term to the next alternative in one OR clause.
36    predicates: Vec<(Comparison, IndexValue<'a>, bool)>,
37    order: Option<GroupOrder>,
38    page: Option<jobs::pages::PageRequest<'a>>,
39}
40impl GroupRequest<'_> {
41    pub(super) fn is_empty(&self) -> bool {
42        self.predicates.is_empty() && self.order.is_none() && self.page.is_none()
43    }
44
45    fn admit(&self, entity: &EntitySchema, reducer: &Reducer) -> Result<Selection, Error> {
46        if self
47            .page
48            .is_some_and(|page| !(1..=MAX_GROUPS).contains(&page.size))
49        {
50            return Err(Error::Limit);
51        }
52        let field = match reducer {
53            Reducer::Count { .. } | Reducer::Distinct { .. } => ValueSchema {
54                value: FieldType::Nat64,
55                optional: false,
56            },
57            Reducer::Extremum { position, .. } => {
58                let field = &entity.fields[*position];
59                ValueSchema {
60                    value: field.value,
61                    optional: field.optional,
62                }
63            }
64            Reducer::Numeric(state) => state.output(),
65        };
66        // Extrema/numeric results may be absent even with required input fields.
67        let absent_result = !matches!(reducer, Reducer::Count { .. } | Reducer::Distinct { .. });
68        let predicates = self
69            .predicates
70            .iter()
71            .map(|&(comparison, value, or_next)| {
72                if match value {
73                    IndexValue::Text(text) => text.len() > crate::codec::MAX_ROW_BYTES,
74                    IndexValue::Blob(bytes) => bytes.len() > crate::codec::MAX_ROW_BYTES,
75                    _ => false,
76                } {
77                    return Err(Error::Limit);
78                }
79                comparison.admit(field.value, absent_result, value)?;
80                let key = if value == IndexValue::Absent {
81                    None
82                } else {
83                    let mut key = Vec::new();
84                    value.append_component(field.optional, &mut key);
85                    if key.len() > MAX_COMPARISON_KEY {
86                        return Err(Error::Limit);
87                    }
88                    Some(key)
89                };
90                Ok((comparison, key, or_next))
91            })
92            .collect::<Result<_, Error>>()?;
93        Ok(Selection {
94            predicates,
95            order: self.order.unwrap_or_default(),
96            optional: field.optional,
97        })
98    }
99}
100
101struct Selection {
102    predicates: Vec<(Comparison, Option<Vec<u8>>, bool)>,
103    order: GroupOrder,
104    // Physical comparison-key framing, distinct from an absent aggregate result.
105    optional: bool,
106}
107
108/// A completed aggregate and its execution costs. No partial result or cursor.
109#[derive(Debug)]
110pub struct Aggregate<T> {
111    /// Completed value over all matching input rows.
112    pub value: T,
113    /// Examined primary/index candidates, including residual nonmatches.
114    pub examined: usize,
115    /// Fetched and validated source rows.
116    pub fetched: usize,
117    /// Encoded result bytes, excluding transport framing.
118    pub output_bytes: usize,
119}
120
121/// One complete group. Absent optional key values belong to a group too.
122#[derive(Debug)]
123pub struct GroupCount<K> {
124    /// Projected grouping tuple in an ordinary named record.
125    pub key: K,
126    /// Matching rows, or present operand values for `group_count_field`.
127    pub count: u64,
128}
129
130/// One complete group's aggregate value, preserving all-absent groups.
131#[derive(Debug)]
132pub struct GroupValue<K, R> {
133    /// Projected grouping tuple, including optional absent keys.
134    pub key: K,
135    /// Single-field result record, or `None` when the group has no present operand.
136    pub value: Option<R>,
137}
138
139impl<'a, E: Entity, M: Memory> Query<'a, E, M> {
140    /// Page completed retained groups, returning at most 1..=32 rows per response.
141    /// `limit` still caps the complete selected result and binds the job. Page size
142    /// and byte allowance may change. Supply None for the first page; continuation
143    /// is valid only for a completed job. Only grouped resume terminals accept this.
144    #[must_use]
145    pub fn group_page(mut self, size: usize, after: Option<&'a jobs::GroupCursor>) -> Self {
146        self.request.groups.page = Some(jobs::pages::PageRequest { size, after });
147        self
148    }
149
150    /// Add an AND predicate on the completed aggregate value, before group ordering
151    /// and output limits. At most eight terms. Counts require Nat64 operands;
152    /// extrema require the source type; sums require the declared result type.
153    /// Absent values support Eq/Ne only. Text and Blob extrema also accept StartsWith,
154    /// EndsWith and Contains.
155    /// Only grouped terminals accept this modifier, even for empty input.
156    #[must_use]
157    pub fn filter_groups(
158        mut self,
159        comparison: Comparison,
160        value: impl Into<IndexValue<'a>>,
161    ) -> Self {
162        if self.request.groups.predicates.len() == MAX_TERMS {
163            self.request.over_limit = true;
164        } else {
165            self.request
166                .groups
167                .predicates
168                .push((comparison, value.into(), false));
169        }
170        self
171    }
172
173    /// Add an OR clause on completed aggregate values, ANDed with other group
174    /// filters. The clause must be nonempty; all group filters together permit
175    /// eight comparisons. Every operand is admitted before reading any rows.
176    /// Uses `filter_groups`' exact type and absence rules. All groups must finish
177    /// and validate before selection, ordering and output limits. Clause boundaries
178    /// bind retained jobs; a singleton clause is equivalent to `filter_groups`.
179    #[must_use]
180    pub fn filter_groups_any(mut self, alternatives: &[(Comparison, IndexValue<'a>)]) -> Self {
181        let count = alternatives.len();
182        if count == 0 || count > MAX_TERMS - self.request.groups.predicates.len() {
183            self.request.over_limit = true;
184        } else {
185            self.request.groups.predicates.extend(
186                alternatives
187                    .iter()
188                    .enumerate()
189                    .map(|(i, &(comparison, value))| (comparison, value, i + 1 != count)),
190            );
191        }
192        self
193    }
194
195    /// Order completed groups before applying output limits. Only grouped
196    /// terminals accept this modifier. The last call selects the order.
197    #[must_use]
198    pub fn order_groups(mut self, order: GroupOrder) -> Self {
199        self.request.groups.order = Some(order);
200        self
201    }
202
203    /// Minimum present value of an exact single-field record projection.
204    /// Supports ordered scalars/unit enums without requiring an index. The result
205    /// preserves the declared type and optionality; empty/all-absent input returns
206    /// `None`. `limit` never truncates input. Only the winning record's bytes count
207    /// toward the engine output bound; an absent result occupies zero payload bytes.
208    /// # Errors
209    /// Refuses multi-field projections, unsupported types, cursors, DISTINCT and
210    /// explicit ordering. Work exhaustion or a comparison key beyond 4096 bytes
211    /// returns `Limit`; an oversized winning result returns `RowTooLarge`.
212    /// Ordinary projection/query errors apply. Completion must precede `work_limit`.
213    /// # Panics
214    /// Supplier corruption traps without repair.
215    pub fn min<R: Value>(self) -> Result<Aggregate<Option<R>>, Error> {
216        self.extremum(false)
217    }
218
219    /// Maximum present value, with the same types, bounds and empty-input behavior
220    /// as [`Self::min`]. Equal operands retain the first encountered value.
221    /// # Errors
222    /// The same projection, resource and query errors as [`Self::min`].
223    /// # Panics
224    /// Supplier corruption traps without repair.
225    pub fn max<R: Value>(self) -> Result<Aggregate<Option<R>>, Error> {
226        self.extremum(true)
227    }
228
229    fn extremum<R: Value>(self, maximum: bool) -> Result<Aggregate<Option<R>>, Error> {
230        let (plan, reducer) = self.extremum_plan::<R>(maximum)?;
231        self.scalar_value(&plan, reducer)
232    }
233
234    fn extremum_plan<R: Value>(&self, maximum: bool) -> Result<(Plan<'a>, Reducer), Error> {
235        let plan = self.aggregate_plan(Some(R::definition()), false)?;
236        let reducer = Reducer::extremum(
237            E::SCHEMA,
238            plan.projection.ok_or(Error::InvalidQuery)?,
239            maximum,
240        )?;
241        Ok((plan, reducer))
242    }
243
244    fn scalar_value<R: Value>(
245        self,
246        plan: &Plan<'_>,
247        mut reducer: Reducer,
248    ) -> Result<Aggregate<Option<R>>, Error> {
249        let page = execute(
250            &self.db.map,
251            E::SCHEMA,
252            &self.request,
253            plan,
254            Some(Accumulator::Scalar(&mut reducer)),
255        )?;
256        let mut output_bytes = 0;
257        let value = reducer
258            .finish()?
259            .into_value()?
260            .map(|bytes| {
261                admit_output(bytes.len(), self.request.bytes, &mut output_bytes)?;
262                decode_record(&bytes)
263            })
264            .transpose()?;
265        Ok(Aggregate {
266            value,
267            examined: page.examined,
268            fetched: page.fetched,
269            output_bytes,
270        })
271    }
272
273    /// Minimum present operand per group. `K` is an exact grouping record; `R`
274    /// is an exact single-field operand record. They may overlap. All-absent
275    /// groups retain `None`; empty input yields no groups. No index is required.
276    /// Groups default to ascending canonical field-name tuple order;
277    /// `filter_groups` and `order_groups` select completed results. `limit` selects
278    /// completed groups, never input rows. Output bytes are the key record plus
279    /// the present winning record; transport option framing is excluded.
280    /// At most 32 groups and 128 KiB of retained keys, records and reducer payloads
281    /// are admitted. Every comparison key is bounded to 4096 bytes. Winner growth
282    /// is charged, and replacement releases the previous winner's charge.
283    /// # Errors
284    /// The projection, work and shape refusals of [`Self::min`] apply. Group or
285    /// retained-state exhaustion returns `Limit`. An oversized output group
286    /// returns `RowTooLarge`; combined output exhaustion returns `Limit`.
287    /// # Panics
288    /// Supplier corruption traps without repair.
289    pub fn group_min<K: Value, R: Value>(self) -> Result<Aggregate<Vec<GroupValue<K, R>>>, Error> {
290        self.group_extrema(false)
291    }
292
293    /// Maximum present operand per group, with the contract of [`Self::group_min`].
294    /// # Errors
295    /// The same projection, work, state, shape and output refusals as grouped minimum.
296    /// # Panics
297    /// Supplier corruption traps without repair.
298    pub fn group_max<K: Value, R: Value>(self) -> Result<Aggregate<Vec<GroupValue<K, R>>>, Error> {
299        self.group_extrema(true)
300    }
301
302    fn group_extrema<K: Value, R: Value>(
303        self,
304        maximum: bool,
305    ) -> Result<Aggregate<Vec<GroupValue<K, R>>>, Error> {
306        let (plan, initial) = self.group_extremum_plan::<K, R>(maximum)?;
307        self.group_values(&plan, initial)
308    }
309
310    fn group_extremum_plan<K: Value, R: Value>(
311        &self,
312        maximum: bool,
313    ) -> Result<(Plan<'a>, Reducer), Error> {
314        let mut plan = self.aggregate_plan(Some(K::definition()), true)?;
315        let operand = admit_projection(E::SCHEMA, R::definition())?;
316        let initial = Reducer::extremum(E::SCHEMA, operand, maximum)?;
317        plan.selected |= operand;
318        Ok((plan, initial))
319    }
320
321    fn group_values<K: Value, R: Value>(
322        self,
323        plan: &Plan<'_>,
324        initial: Reducer,
325    ) -> Result<Aggregate<Vec<GroupValue<K, R>>>, Error> {
326        let (groups, page) = self.grouped(plan, initial)?;
327        let (value, output_bytes) = group_output(
328            groups,
329            self.request.rows,
330            self.request.bytes,
331            decode_group_value,
332        )?;
333        Ok(Aggregate {
334            value,
335            examined: page.examined,
336            fetched: page.fetched,
337            output_bytes,
338        })
339    }
340
341    /// Whether at least one row matches, without materializing an entity.
342    /// Can seek a mandatory equality prefix and evaluate remaining filters on its owner rows.
343    /// Stops after validating the first matching row; later rows are not inspected.
344    /// Returns false only after observing the end of the selected input.
345    /// `limit` does not truncate input. The boolean result occupies one byte.
346    /// # Errors
347    /// Refuses cursors, DISTINCT, explicit ordering and group modifiers, as for
348    /// `count`. Exhausting `work_limit` without a match returns `Limit`, never
349    /// false; an exact-boundary no-match scan can conservatively refuse.
350    /// Ordinary admission and examined-row integrity errors still apply.
351    /// # Panics
352    /// Supplier corruption traps without repair, as for ordinary query execution.
353    pub fn exists(self) -> Result<Aggregate<bool>, Error> {
354        let mut plan = self.aggregate_plan(None, false)?;
355        self.select_aggregate_row_path(&mut plan)?;
356        let mut found = false;
357        let page = execute(
358            &self.db.map,
359            E::SCHEMA,
360            &self.request,
361            &plan,
362            Some(Accumulator::Exists(&mut found)),
363        )?;
364        Ok(Aggregate {
365            value: found,
366            examined: page.examined,
367            fetched: page.fetched,
368            output_bytes: 1,
369        })
370    }
371
372    /// Count all matching rows in one synchronous execution. Empty input yields zero.
373    /// Uses maintained cardinality for unfiltered input and physical index keys for
374    /// fully covered index ranges or primary-key comparison ranges. A partially covered
375    /// equality prefix narrows the owner rows checked by the remaining predicates. Use
376    /// [`Self::count_validated`] to check each visited row and selected index owner.
377    /// `limit` does not limit input; the scalar output occupies eight bytes.
378    /// Completion must be established within `work_limit`; an exact-boundary
379    /// scan can conservatively refuse. Use `start_count`/`resume_count` for retained work.
380    /// # Errors
381    /// Refuses cursors, DISTINCT, explicit ordering and the ordinary query failures.
382    /// Work exhaustion returns `Limit`, never a partial count. Too few output bytes
383    /// returns `RowTooLarge`.
384    /// # Panics
385    /// Supplier corruption traps without repair, as for ordinary query execution.
386    pub fn count(self) -> Result<Aggregate<u64>, Error> {
387        if self.request.bytes < 8 {
388            return Err(Error::RowTooLarge);
389        }
390        let mut plan = self.aggregate_plan(None, false)?;
391        if self.request.predicates.is_empty() && self.request.index.is_none() {
392            return Ok(Aggregate {
393                value: self.db.cardinality::<E>()?,
394                examined: 0,
395                fetched: 0,
396                output_bytes: 8,
397            });
398        }
399        if super::range::covers_primary(&plan)
400            || super::order::select_aggregate_prefix(E::SCHEMA, &self.request, &mut plan)?
401        {
402            return self.count_keys(&plan);
403        }
404        let reducer = self.count_reducer(None)?;
405        self.scalar_count_value(&plan, reducer)
406    }
407
408    /// Count by validating each visited row, including its selected index entry.
409    /// # Errors
410    /// Returns the same admission and resource errors as [`Self::count`], and
411    /// reports corruption encountered during traversal.
412    pub fn count_validated(self) -> Result<Aggregate<u64>, Error> {
413        self.scalar_count(None)
414    }
415
416    /// Count physical entries in the range selected by `using_index` and its
417    /// leading operands. This reads neither index values nor owner rows. A
418    /// dangling index entry counts as one. Use [`Self::count_validated`] to validate rows.
419    /// Like other scalar counts, the row limit does not truncate the input.
420    /// # Errors
421    /// Requires an explicit index. Row predicates, ordering, cursors, grouping
422    /// and DISTINCT refuse. Visiting the inclusive terminal key proves completion,
423    /// even on the last allowed visit. Exhausting the work budget earlier returns
424    /// `Limit`; fewer than eight output bytes returns `RowTooLarge`.
425    /// # Panics
426    /// Malformed supplier storage can trap during traversal.
427    pub fn count_index_entries(self) -> Result<Aggregate<u64>, Error> {
428        if self.request.index.is_none() || !self.request.predicates.is_empty() {
429            return Err(Error::InvalidQuery);
430        }
431        if self.request.bytes < 8 {
432            return Err(Error::RowTooLarge);
433        }
434        let plan = self.aggregate_plan(None, false)?;
435        self.count_keys(&plan)
436    }
437
438    fn count_keys(&self, plan: &Plan<'_>) -> Result<Aggregate<u64>, Error> {
439        let bounds = super::range::bounds(E::SCHEMA, &self.request, plan)?;
440        let mut examined = 0;
441        if let Some(range) = bounds {
442            let terminal = super::range::terminal_key(&range, false);
443            let entries = self.db.map.range((range.0.as_ref(), range.1.as_ref()));
444            for entry in entries.take(self.request.work) {
445                examined += 1;
446                if examined == self.request.work && terminal.is_none_or(|end| end != entry.key()) {
447                    return Err(Error::Limit);
448                }
449            }
450        }
451        Ok(Aggregate {
452            value: examined as u64,
453            examined,
454            fetched: 0,
455            output_bytes: 8,
456        })
457    }
458
459    /// Count present values of a top-level field, including non-scalar fields.
460    /// Empty input and all-absent input yield zero; present empty values count.
461    /// Limits, refusals and corruption traps are the same as [`Self::count`].
462    /// # Errors
463    /// Unknown fields refuse even for empty input. Ordinary count errors apply.
464    /// # Panics
465    /// Supplier corruption traps without repair.
466    pub fn count_field(self, field: &str) -> Result<Aggregate<u64>, Error> {
467        self.scalar_count(Some(field))
468    }
469
470    /// Count distinct projected tuples, including absent values. An indexed filter can narrow the input.
471    /// Retain at most 32 distinct comparison keys, each at most 4096 bytes, within
472    /// the shared 128-KiB state bound. No projected records are retained. `limit`
473    /// never truncates this count. The scalar result occupies eight bytes.
474    /// # Errors
475    /// Ordinary count/projection errors and distinct-state exhaustion refuse.
476    /// # Panics
477    /// Supplier corruption traps without repair.
478    pub fn count_distinct<K: Value>(self) -> Result<Aggregate<u64>, Error> {
479        if self.request.bytes < 8 {
480            return Err(Error::RowTooLarge);
481        }
482        let (mut plan, reducer) = self.distinct_plan::<K>()?;
483        self.select_aggregate_row_path(&mut plan)?;
484        self.scalar_count_value(&plan, reducer)
485    }
486
487    fn distinct_plan<K: Value>(&self) -> Result<(Plan<'static>, Reducer), Error> {
488        let plan = self.aggregate_plan(Some(K::definition()), false)?;
489        let reducer = Reducer::distinct(plan.projection.ok_or(Error::InvalidQuery)?);
490        Ok((plan, reducer))
491    }
492
493    fn select_aggregate_row_path(&self, plan: &mut Plan<'static>) -> Result<(), Error> {
494        if plan.index.is_none() && !super::range::covers_primary(plan) {
495            super::order::select_aggregate_prefix(E::SCHEMA, &self.request, plan)?;
496        }
497        Ok(())
498    }
499
500    fn scalar_count(self, field: Option<&str>) -> Result<Aggregate<u64>, Error> {
501        if self.request.bytes < 8 {
502            return Err(Error::RowTooLarge);
503        }
504        let mut plan = self.aggregate_plan(None, false)?;
505        self.select_aggregate_row_path(&mut plan)?;
506        let reducer = self.count_reducer(field)?;
507        self.scalar_count_value(&plan, reducer)
508    }
509
510    fn scalar_count_value(
511        self,
512        plan: &Plan<'_>,
513        mut reducer: Reducer,
514    ) -> Result<Aggregate<u64>, Error> {
515        let page = execute(
516            &self.db.map,
517            E::SCHEMA,
518            &self.request,
519            plan,
520            Some(Accumulator::Scalar(&mut reducer)),
521        )?;
522        Ok(Aggregate {
523            value: reducer.finish()?.into_count()?,
524            examined: page.examined,
525            fetched: page.fetched,
526            output_bytes: 8,
527        })
528    }
529
530    /// Count matching rows grouped by an exact named-record projection.
531    /// Keys support ordered scalars/unit enums, including optional absence; no index
532    /// is required. Results default to ascending tuple order in canonical field-name order.
533    /// `filter_groups` and `order_groups` select completed results before output limits.
534    /// `limit` selects the first completed groups (1..=32), never input rows.
535    /// The engine output allowance bounds key-record bytes plus eight count bytes per group.
536    /// Empty input yields no groups. At most 32 groups and 128 KiB of key/output
537    /// state may be retained, independently of the requested output limit;
538    /// each comparison key is capped at 4096 bytes, independently of index-key limits.
539    /// # Errors
540    /// Refuses unsupported key types, cursors, DISTINCT and explicit ordering.
541    /// Work, group or retained-state exhaustion returns `Limit`, never partial groups.
542    /// Completion must be established within `work_limit`; a boundary scan can refuse.
543    /// Ordinary projection/query integrity and byte-limit errors also apply.
544    /// # Panics
545    /// Supplier corruption traps without repair, as for ordinary query execution.
546    pub fn group_count<K: Value>(self) -> Result<Aggregate<Vec<GroupCount<K>>>, Error> {
547        self.group_counts(None)
548    }
549
550    /// Count present operand values within each projected group. Groups whose
551    /// operands are all absent remain in the output with count zero. The operand
552    /// may be any top-level stored field and need not belong to the group key.
553    /// Bounds and ordering are the same as [`Self::group_count`].
554    /// # Errors
555    /// Unknown operands and ordinary grouped-count errors refuse.
556    /// # Panics
557    /// Supplier corruption traps without repair.
558    pub fn group_count_field<K: Value>(
559        self,
560        field: &str,
561    ) -> Result<Aggregate<Vec<GroupCount<K>>>, Error> {
562        self.group_counts(Some(field))
563    }
564
565    /// Count distinct projected `R` tuples within each projected `K` group.
566    /// Absence is a tuple component, as for [`Self::count_distinct`]; an all-absent
567    /// group has one distinct tuple, not zero. Projections may overlap. No index
568    /// is required. Complete input before group filtering/order and output limits.
569    /// Retain at most 32 groups and 32 distinct tuples per group, with one shared
570    /// 128-KiB state allowance and a 4096-byte cap per comparison key.
571    /// # Errors
572    /// Refuses unsupported projections, cursors, DISTINCT and explicit input order.
573    /// Work, cardinality or byte exhaustion returns `Limit`, never partial counts.
574    /// Ordinary grouped-count output and integrity errors apply.
575    /// # Panics
576    /// Supplier corruption traps without repair.
577    pub fn group_count_distinct<K: Value, R: Value>(
578        self,
579    ) -> Result<Aggregate<Vec<GroupCount<K>>>, Error> {
580        let (plan, reducer) = self.group_distinct_plan::<K, R>()?;
581        self.group_count_values(plan, reducer)
582    }
583
584    fn group_distinct_plan<K: Value, R: Value>(&self) -> Result<(Plan<'static>, Reducer), Error> {
585        let mut plan = self.aggregate_plan(Some(K::definition()), true)?;
586        let operand = admit_projection(E::SCHEMA, R::definition())?;
587        admit_ordered_fields(E::SCHEMA, operand)?;
588        plan.selected |= operand;
589        Ok((plan, Reducer::distinct(operand)))
590    }
591
592    fn group_counts<K: Value>(
593        self,
594        field: Option<&str>,
595    ) -> Result<Aggregate<Vec<GroupCount<K>>>, Error> {
596        let plan = self.aggregate_plan(Some(K::definition()), true)?;
597        let reducer = self.count_reducer(field)?;
598        self.group_count_values(plan, reducer)
599    }
600
601    fn group_count_values<K: Value>(
602        self,
603        mut plan: Plan<'static>,
604        reducer: Reducer,
605    ) -> Result<Aggregate<Vec<GroupCount<K>>>, Error> {
606        self.select_aggregate_row_path(&mut plan)?;
607        let (groups, page) = self.grouped(&plan, reducer)?;
608        let (value, output_bytes) = group_output(
609            groups,
610            self.request.rows,
611            self.request.bytes,
612            decode_group_count,
613        )?;
614        Ok(Aggregate {
615            value,
616            examined: page.examined,
617            fetched: page.fetched,
618            output_bytes,
619        })
620    }
621
622    fn count_reducer(&self, field: Option<&str>) -> Result<Reducer, Error> {
623        // Presence needs only the validated tag, not an ordered scalar slot.
624        let presence = field
625            .map(|name| E::SCHEMA.field_position(name).ok_or(Error::InvalidQuery))
626            .transpose()?
627            .filter(|&i| E::SCHEMA.fields[i].optional);
628        Ok(Reducer::Count { presence, total: 0 })
629    }
630
631    fn grouped(
632        &self,
633        plan: &Plan<'_>,
634        initial: Reducer,
635    ) -> Result<(Groups, QueryPage<Vec<u8>>), Error> {
636        if self.request.groups.page.is_some() {
637            return Err(Error::InvalidQuery);
638        }
639        let mut groups = self.admit_groups(plan, initial)?;
640        let page = execute(
641            &self.db.map,
642            E::SCHEMA,
643            &self.request,
644            plan,
645            Some(Accumulator::Groups(&mut groups)),
646        )?;
647        Ok((groups, page))
648    }
649
650    fn admit_groups(&self, plan: &Plan<'_>, initial: Reducer) -> Result<Groups, Error> {
651        Ok(Groups {
652            selection: self.request.groups.admit(E::SCHEMA, &initial)?,
653            mask: plan.projection.ok_or(Error::InvalidQuery)?,
654            initial,
655            rows: BTreeMap::new(),
656            state_bytes: 0,
657        })
658    }
659
660    fn aggregate_plan(
661        &self,
662        output: Option<&TypeSchema>,
663        group_output: bool,
664    ) -> Result<Plan<'static>, Error> {
665        if self.request.cursor.is_some()
666            || self.request.distinct
667            || self.request.order.is_some()
668            || self.request.descending
669        {
670            return Err(Error::InvalidQuery);
671        }
672        self.db.schema.admits(E::SCHEMA)?;
673        let mut plan = Plan::admit(
674            E::SCHEMA,
675            &self.request,
676            self.db.schema.fingerprint(),
677            self.db.incarnation,
678            output,
679            super::PlanOutput::Aggregate {
680                grouped: group_output,
681            },
682        )?;
683        if let Some(mask) = plan.projection {
684            admit_ordered_fields(E::SCHEMA, mask)?;
685            plan.selected |= mask;
686        }
687        Ok(plan)
688    }
689}
690
691fn group_output<T>(
692    groups: Groups,
693    rows: usize,
694    bytes: usize,
695    decode: fn(&[u8], Completed) -> Result<T, Error>,
696) -> Result<(Vec<T>, usize), Error> {
697    let mut output_bytes = 0;
698    let value = groups
699        .into_output()?
700        .into_iter()
701        .take(rows)
702        .map(|(key, result)| {
703            admit_output(key.len() + result.encoded_len(), bytes, &mut output_bytes)?;
704            decode(&key, result)
705        })
706        .collect::<Result<_, Error>>()?;
707    Ok((value, output_bytes))
708}
709
710fn decode_group_count<K: Value>(key: &[u8], result: Completed) -> Result<GroupCount<K>, Error> {
711    Ok(GroupCount {
712        key: decode_record(key)?,
713        count: result.into_count()?,
714    })
715}
716
717fn decode_group_value<K: Value, R: Value>(
718    key: &[u8],
719    result: Completed,
720) -> Result<GroupValue<K, R>, Error> {
721    let value = result.into_value()?;
722    Ok(GroupValue {
723        key: decode_record(key)?,
724        value: value.as_deref().map(decode_record).transpose()?,
725    })
726}
727
728fn decode_record<R: Value>(bytes: &[u8]) -> Result<R, Error> {
729    let mut reader = Decoder::new(bytes)?;
730    let value = R::decode_from(&mut reader)?;
731    reader.finish()?;
732    Ok(value)
733}
734
735// Closed choices over delivered reducers and one grouping table, not user extensions.
736pub(super) enum Accumulator<'a> {
737    // Synchronous existence can complete on the first matching row; no job state.
738    Exists(&'a mut bool),
739    Scalar(&'a mut Reducer),
740    Groups(&'a mut Groups),
741}
742impl Accumulator<'_> {
743    pub(super) fn needs_spans(&self) -> bool {
744        !matches!(
745            self,
746            Self::Exists(_)
747                | Self::Scalar(
748                    Reducer::Numeric(_)
749                        | Reducer::Count { presence: None, .. }
750                        | Reducer::Distinct { .. }
751                )
752        )
753    }
754
755    pub(super) fn push(
756        &mut self,
757        entity: &EntitySchema,
758        values: &[IndexValue<'_>],
759        bytes: &[u8],
760        spans: &[(u16, u16)],
761    ) -> Result<(), Error> {
762        match self {
763            Self::Exists(found) => {
764                **found = true;
765                Ok(())
766            }
767            Self::Scalar(state) => state.push(entity, values, bytes, spans, MAX_STATE_BYTES),
768            Self::Groups(state) => state.push(entity, values, bytes, spans),
769        }
770    }
771}
772
773#[derive(Clone)]
774pub(super) enum Reducer {
775    Numeric(numeric::NumericAggregate),
776    Distinct {
777        mask: u16,
778        seen: BTreeSet<Vec<u8>>,
779    },
780    Count {
781        presence: Option<usize>,
782        total: u64,
783    },
784    Extremum {
785        position: usize,
786        maximum: bool,
787        selected: Option<(Vec<u8>, Vec<u8>)>,
788    },
789}
790impl Reducer {
791    fn distinct(mask: u16) -> Self {
792        Self::Distinct {
793            mask,
794            seen: BTreeSet::new(),
795        }
796    }
797
798    fn extremum(entity: &EntitySchema, mask: u16, maximum: bool) -> Result<Self, Error> {
799        if mask.count_ones() != 1 {
800            return Err(Error::InvalidQuery);
801        }
802        let position = mask.trailing_zeros() as usize;
803        if !key::supports(entity.fields[position].value) {
804            return Err(Error::InvalidQuery);
805        }
806        Ok(Self::Extremum {
807            position,
808            maximum,
809            selected: None,
810        })
811    }
812
813    fn retained_bytes(&self) -> usize {
814        match self {
815            Self::Numeric(state) => state.retained_bytes(),
816            Self::Count { .. } => 8,
817            Self::Distinct { seen, .. } => 8 + seen.iter().map(Vec::capacity).sum::<usize>(),
818            Self::Extremum { selected, .. } => selected
819                .as_ref()
820                .map_or(0, |(key, bytes)| key.capacity() + bytes.capacity()),
821        }
822    }
823
824    fn finish(self) -> Result<Completed, Error> {
825        match self {
826            Self::Count { total, .. } => Ok(Completed::Count(total)),
827            Self::Distinct { seen, .. } => Ok(Completed::Count(seen.len() as u64)),
828            Self::Extremum { selected, .. } => Ok(Completed::Value(selected)),
829            Self::Numeric(state) => state.finish(),
830        }
831    }
832
833    fn push(
834        &mut self,
835        entity: &EntitySchema,
836        values: &[IndexValue<'_>],
837        bytes: &[u8],
838        spans: &[(u16, u16)],
839        allowance: usize,
840    ) -> Result<(), Error> {
841        match self {
842            Self::Numeric(state) => state.push(values)?,
843            Self::Distinct { mask, seen } => {
844                let key = comparison_key(entity, *mask, values)?;
845                if !seen.contains(&key) {
846                    if seen.len() == MAX_DISTINCT_VALUES
847                        || 8 + seen.iter().map(Vec::capacity).sum::<usize>() + key.capacity()
848                            > allowance
849                    {
850                        return Err(Error::Limit);
851                    }
852                    seen.insert(key);
853                }
854            }
855            Self::Count { presence, total } => {
856                *total += u64::from(presence.is_none_or(|i| bytes[usize::from(spans[i].0)] != 0));
857            }
858            Self::Extremum {
859                position,
860                maximum,
861                selected,
862            } => {
863                let value = values[*position];
864                if value == IndexValue::Absent {
865                    return Ok(());
866                }
867                let mut key = Vec::new();
868                value.append_component(entity.fields[*position].optional, &mut key);
869                if key.len() > MAX_COMPARISON_KEY {
870                    return Err(Error::Limit);
871                }
872                if selected
873                    .as_ref()
874                    .is_none_or(|(old, _)| if *maximum { key > *old } else { key < *old })
875                {
876                    let (start, end) = spans[*position];
877                    if key.capacity() + usize::from(end - start) > allowance {
878                        return Err(Error::Limit);
879                    }
880                    // Copy only a winner, after admitting its replacement charge.
881                    let output = bytes[usize::from(start)..usize::from(end)].to_vec();
882                    if key.capacity() + output.capacity() > allowance {
883                        return Err(Error::Limit);
884                    }
885                    *selected = Some((key, output));
886                }
887            }
888        }
889        Ok(())
890    }
891}
892
893// Completed results cannot be advanced. All numeric range checks precede selection.
894enum Completed {
895    Count(u64),
896    Value(Option<(Vec<u8>, Vec<u8>)>),
897}
898impl Completed {
899    fn encoded_len(&self) -> usize {
900        match self {
901            Self::Count(_) => 8,
902            Self::Value(value) => value.as_ref().map_or(0, |(_, bytes)| bytes.len()),
903        }
904    }
905    // Counts need only eight stack bytes; extrema borrow their retained ordered key.
906    fn result_key<'a>(&'a self, scratch: &'a mut [u8; 8]) -> Option<&'a [u8]> {
907        match self {
908            Self::Count(total) => {
909                *scratch = total.to_be_bytes();
910                Some(scratch)
911            }
912            Self::Value(selected) => selected.as_ref().map(|(key, _)| key.as_slice()),
913        }
914    }
915
916    fn retained_bytes(&self) -> usize {
917        match self {
918            Self::Count(_) => 8,
919            Self::Value(value) => value
920                .as_ref()
921                .map_or(0, |(key, bytes)| key.capacity() + bytes.capacity()),
922        }
923    }
924
925    fn into_count(self) -> Result<u64, Error> {
926        match self {
927            Self::Count(count) => Ok(count),
928            _ => Err(Error::Corrupt),
929        }
930    }
931
932    fn into_value(self) -> Result<Option<Vec<u8>>, Error> {
933        match self {
934            Self::Value(value) => Ok(value.map(|(_, bytes)| bytes)),
935            _ => Err(Error::Corrupt),
936        }
937    }
938}
939
940pub(super) struct Groups {
941    selection: Selection,
942    mask: u16,
943    // Never advanced: each new group receives its own empty reducer state.
944    initial: Reducer,
945    rows: BTreeMap<Vec<u8>, (Vec<u8>, Reducer)>,
946    state_bytes: usize,
947}
948impl Groups {
949    fn into_output(self) -> Result<Vec<(Vec<u8>, Completed)>, Error> {
950        // Accumulation has completed. Output filtering cannot relax input/state bounds.
951        let mut state_bytes = self.state_bytes;
952        let mut rows = self
953            .rows
954            .into_iter()
955            .map(|(key, (record, reducer))| {
956                let previous = reducer.retained_bytes();
957                let result = reducer.finish()?;
958                state_bytes = state_bytes - previous + result.retained_bytes();
959                if state_bytes > MAX_STATE_BYTES {
960                    return Err(Error::Limit);
961                }
962                Ok((key, (record, result)))
963            })
964            .collect::<Result<Vec<_>, Error>>()?;
965        rows.retain(|(_, (_, result))| {
966            let mut scratch = [0; 8];
967            let actual = result.result_key(&mut scratch);
968            self.selection
969                .predicates
970                .split_inclusive(|(_, _, or_next)| !or_next)
971                .all(|clause| {
972                    clause.iter().any(|(comparison, operand, _)| {
973                        comparison.accepts(actual, operand.as_deref(), self.selection.optional)
974                    })
975                })
976        });
977        if self.selection.order == GroupOrder::KeyDescending {
978            // BTreeMap supplied ascending canonical keys; reverse the completed,
979            // filtered rows before any output limit or page is selected.
980            rows.reverse();
981        } else if self.selection.order != GroupOrder::Key {
982            rows.sort_unstable_by(|(ka, (_, a)), (kb, (_, b))| {
983                let (mut sa, mut sb) = ([0; 8], [0; 8]);
984                let order = a.result_key(&mut sa).cmp(&b.result_key(&mut sb));
985                let order = if self.selection.order == GroupOrder::AggregateDescending {
986                    order.reverse()
987                } else {
988                    order
989                };
990                order.then_with(|| ka.cmp(kb))
991            });
992        }
993        Ok(rows.into_iter().map(|(_, row)| row).collect())
994    }
995
996    fn push(
997        &mut self,
998        entity: &EntitySchema,
999        values: &[IndexValue<'_>],
1000        bytes: &[u8],
1001        spans: &[(u16, u16)],
1002    ) -> Result<(), Error> {
1003        let key = comparison_key(entity, self.mask, values)?;
1004        let size = projection_size(spans, self.mask);
1005        if let Some((_, reducer)) = self.rows.get_mut(&key) {
1006            let previous = reducer.retained_bytes();
1007            reducer.push(
1008                entity,
1009                values,
1010                bytes,
1011                spans,
1012                MAX_STATE_BYTES - self.state_bytes + previous,
1013            )?;
1014            self.state_bytes = self.state_bytes - previous + reducer.retained_bytes();
1015            return Ok(());
1016        }
1017        let base = key.capacity() + size;
1018        let available = MAX_STATE_BYTES - self.state_bytes;
1019        if self.rows.len() == MAX_GROUPS || base + self.initial.retained_bytes() > available {
1020            return Err(Error::Limit);
1021        }
1022        let mut reducer = self.initial.clone();
1023        reducer.push(entity, values, bytes, spans, available - base)?;
1024        let output = copy_projection(bytes, spans, self.mask, size);
1025        // Map/set nodes and fixed bookkeeping are independently bounded by the
1026        // group and per-group distinct cardinality caps.
1027        let retained = key.capacity() + output.capacity() + reducer.retained_bytes();
1028        if retained > available {
1029            return Err(Error::Limit);
1030        }
1031        self.state_bytes += retained;
1032        self.rows.insert(key, (output, reducer));
1033        Ok(())
1034    }
1035}