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 `byte_limit`; 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) = value_output(groups, self.request.rows, self.request.bytes)?;
328        Ok(Aggregate {
329            value,
330            examined: page.examined,
331            fetched: page.fetched,
332            output_bytes,
333        })
334    }
335
336    /// Whether at least one row matches, without materializing an entity.
337    /// Stops after validating the first matching row; later rows are not inspected.
338    /// Returns false only after observing the end of the selected input.
339    /// `limit` does not truncate input. The boolean result occupies one byte.
340    /// # Errors
341    /// Refuses cursors, DISTINCT, explicit ordering and group modifiers, as for
342    /// `count`. Exhausting `work_limit` without a match returns `Limit`, never
343    /// false; an exact-boundary no-match scan can conservatively refuse.
344    /// Ordinary admission and examined-row integrity errors still apply.
345    /// # Panics
346    /// Supplier corruption traps without repair, as for ordinary query execution.
347    pub fn exists(self) -> Result<Aggregate<bool>, Error> {
348        let plan = self.aggregate_plan(None, false)?;
349        let mut found = false;
350        let page = execute(
351            &self.db.map,
352            E::SCHEMA,
353            &self.request,
354            &plan,
355            Some(Accumulator::Exists(&mut found)),
356        )?;
357        Ok(Aggregate {
358            value: found,
359            examined: page.examined,
360            fetched: page.fetched,
361            output_bytes: 1,
362        })
363    }
364
365    /// Count all matching rows in one synchronous execution. Empty input yields zero.
366    /// Uses maintained cardinality for unfiltered input and physical index keys for
367    /// covered equality prefixes or primary-key comparison ranges. Other predicates use row execution. Use
368    /// [`Self::count_validated`] to check each visited row and selected index owner.
369    /// `limit` does not limit input; the scalar output occupies eight bytes.
370    /// Completion must be observed before `work_limit` is reached; an exact-boundary
371    /// scan can conservatively refuse. Use `start_count`/`resume_count` for retained work.
372    /// # Errors
373    /// Refuses cursors, DISTINCT, explicit ordering and the ordinary query failures.
374    /// Work exhaustion returns `Limit`, never a partial count. Too few output bytes
375    /// returns `RowTooLarge`.
376    /// # Panics
377    /// Supplier corruption traps without repair, as for ordinary query execution.
378    pub fn count(self) -> Result<Aggregate<u64>, Error> {
379        if self.request.bytes < 8 {
380            return Err(Error::RowTooLarge);
381        }
382        let mut plan = self.aggregate_plan(None, false)?;
383        if self.request.predicates.is_empty() && self.request.index.is_none() {
384            return Ok(Aggregate {
385                value: self.db.cardinality::<E>()?,
386                examined: 0,
387                fetched: 0,
388                output_bytes: 8,
389            });
390        }
391        if super::range::covers_primary(&plan)
392            || super::order::count_prefix(E::SCHEMA, &self.request, &mut plan)?
393        {
394            return self.count_keys(&plan);
395        }
396        let reducer = self.count_reducer(None)?;
397        self.scalar_count_value(&plan, reducer)
398    }
399
400    /// Count by validating each visited row, including its selected index entry.
401    /// # Errors
402    /// Returns the same admission and resource errors as [`Self::count`], and
403    /// reports corruption encountered during traversal.
404    pub fn count_validated(self) -> Result<Aggregate<u64>, Error> {
405        self.scalar_count(None)
406    }
407
408    /// Count physical entries in the range selected by `using_index` and its
409    /// leading operands. This reads neither index values nor owner rows. A
410    /// dangling index entry counts as one. Use [`Self::count_validated`] to validate rows.
411    /// Like other scalar counts, the row limit does not truncate the input.
412    /// # Errors
413    /// Requires an explicit index. Row predicates, ordering, cursors, grouping
414    /// and DISTINCT refuse. Reaching the work limit before observing completion
415    /// returns `Limit`; fewer than eight output bytes returns `RowTooLarge`.
416    /// # Panics
417    /// Malformed supplier storage can trap during traversal.
418    pub fn count_index_entries(self) -> Result<Aggregate<u64>, Error> {
419        if self.request.index.is_none() || !self.request.predicates.is_empty() {
420            return Err(Error::InvalidQuery);
421        }
422        if self.request.bytes < 8 {
423            return Err(Error::RowTooLarge);
424        }
425        let plan = self.aggregate_plan(None, false)?;
426        self.count_keys(&plan)
427    }
428
429    fn count_keys(&self, plan: &Plan<'_>) -> Result<Aggregate<u64>, Error> {
430        let bounds = super::range::bounds(E::SCHEMA, &self.request, plan)?;
431        let examined = bounds.map_or(0, |range| {
432            self.db.map.range(range).take(self.request.work).count()
433        });
434        if examined == self.request.work {
435            return Err(Error::Limit);
436        }
437        Ok(Aggregate {
438            value: examined as u64,
439            examined,
440            fetched: 0,
441            output_bytes: 8,
442        })
443    }
444
445    /// Count present values of a top-level field, including non-scalar fields.
446    /// Empty input and all-absent input yield zero; present empty values count.
447    /// Limits, refusals and corruption traps are the same as [`Self::count`].
448    /// # Errors
449    /// Unknown fields refuse even for empty input. Ordinary count errors apply.
450    /// # Panics
451    /// Supplier corruption traps without repair.
452    pub fn count_field(self, field: &str) -> Result<Aggregate<u64>, Error> {
453        self.scalar_count(Some(field))
454    }
455
456    /// Count distinct projected tuples, including absent values, without an index.
457    /// Retain at most 32 distinct comparison keys, each at most 4096 bytes, within
458    /// the shared 128-KiB state bound. No projected records are retained. `limit`
459    /// never truncates this count. The scalar result occupies eight bytes.
460    /// # Errors
461    /// Ordinary count/projection errors and distinct-state exhaustion refuse.
462    /// # Panics
463    /// Supplier corruption traps without repair.
464    pub fn count_distinct<K: Value>(self) -> Result<Aggregate<u64>, Error> {
465        if self.request.bytes < 8 {
466            return Err(Error::RowTooLarge);
467        }
468        let (plan, reducer) = self.distinct_plan::<K>()?;
469        self.scalar_count_value(&plan, reducer)
470    }
471
472    fn distinct_plan<K: Value>(&self) -> Result<(Plan<'static>, Reducer), Error> {
473        let plan = self.aggregate_plan(Some(K::definition()), false)?;
474        let reducer = Reducer::distinct(plan.projection.ok_or(Error::InvalidQuery)?);
475        Ok((plan, reducer))
476    }
477
478    fn scalar_count(self, field: Option<&str>) -> Result<Aggregate<u64>, Error> {
479        if self.request.bytes < 8 {
480            return Err(Error::RowTooLarge);
481        }
482        let plan = self.aggregate_plan(None, false)?;
483        let reducer = self.count_reducer(field)?;
484        self.scalar_count_value(&plan, reducer)
485    }
486
487    fn scalar_count_value(
488        self,
489        plan: &Plan<'_>,
490        mut reducer: Reducer,
491    ) -> Result<Aggregate<u64>, Error> {
492        let page = execute(
493            &self.db.map,
494            E::SCHEMA,
495            &self.request,
496            plan,
497            Some(Accumulator::Scalar(&mut reducer)),
498        )?;
499        Ok(Aggregate {
500            value: reducer.finish()?.into_count()?,
501            examined: page.examined,
502            fetched: page.fetched,
503            output_bytes: 8,
504        })
505    }
506
507    /// Count matching rows grouped by an exact named-record projection.
508    /// Keys support ordered scalars/unit enums, including optional absence; no index
509    /// is required. Results default to ascending tuple order in canonical field-name order.
510    /// `filter_groups` and `order_groups` select completed results before output limits.
511    /// `limit` selects the first completed groups (1..=32), never input rows.
512    /// `byte_limit` bounds key-record bytes plus eight count bytes per group.
513    /// Empty input yields no groups. At most 32 groups and 128 KiB of key/output
514    /// state may be retained, independently of the requested output limit;
515    /// each comparison key is capped at 4096 bytes, independently of index-key limits.
516    /// # Errors
517    /// Refuses unsupported key types, cursors, DISTINCT and explicit ordering.
518    /// Work, group or retained-state exhaustion returns `Limit`, never partial groups.
519    /// Completion must be observed before `work_limit`; a boundary scan can refuse.
520    /// Ordinary projection/query integrity and byte-limit errors also apply.
521    /// # Panics
522    /// Supplier corruption traps without repair, as for ordinary query execution.
523    pub fn group_count<K: Value>(self) -> Result<Aggregate<Vec<GroupCount<K>>>, Error> {
524        self.group_counts(None)
525    }
526
527    /// Count present operand values within each projected group. Groups whose
528    /// operands are all absent remain in the output with count zero. The operand
529    /// may be any top-level stored field and need not belong to the group key.
530    /// Bounds and ordering are the same as [`Self::group_count`].
531    /// # Errors
532    /// Unknown operands and ordinary grouped-count errors refuse.
533    /// # Panics
534    /// Supplier corruption traps without repair.
535    pub fn group_count_field<K: Value>(
536        self,
537        field: &str,
538    ) -> Result<Aggregate<Vec<GroupCount<K>>>, Error> {
539        self.group_counts(Some(field))
540    }
541
542    /// Count distinct projected `R` tuples within each projected `K` group.
543    /// Absence is a tuple component, as for [`Self::count_distinct`]; an all-absent
544    /// group has one distinct tuple, not zero. Projections may overlap. No index
545    /// is required. Complete input before group filtering/order and output limits.
546    /// Retain at most 32 groups and 32 distinct tuples per group, with one shared
547    /// 128-KiB state allowance and a 4096-byte cap per comparison key.
548    /// # Errors
549    /// Refuses unsupported projections, cursors, DISTINCT and explicit input order.
550    /// Work, cardinality or byte exhaustion returns `Limit`, never partial counts.
551    /// Ordinary grouped-count output and integrity errors apply.
552    /// # Panics
553    /// Supplier corruption traps without repair.
554    pub fn group_count_distinct<K: Value, R: Value>(
555        self,
556    ) -> Result<Aggregate<Vec<GroupCount<K>>>, Error> {
557        let (plan, reducer) = self.group_distinct_plan::<K, R>()?;
558        self.group_count_values(&plan, reducer)
559    }
560
561    fn group_distinct_plan<K: Value, R: Value>(&self) -> Result<(Plan<'static>, Reducer), Error> {
562        let mut plan = self.aggregate_plan(Some(K::definition()), true)?;
563        let operand = admit_projection(E::SCHEMA, R::definition())?;
564        admit_ordered_fields(E::SCHEMA, operand)?;
565        plan.selected |= operand;
566        Ok((plan, Reducer::distinct(operand)))
567    }
568
569    fn group_counts<K: Value>(
570        self,
571        field: Option<&str>,
572    ) -> Result<Aggregate<Vec<GroupCount<K>>>, Error> {
573        let plan = self.aggregate_plan(Some(K::definition()), true)?;
574        let reducer = self.count_reducer(field)?;
575        self.group_count_values(&plan, reducer)
576    }
577
578    fn group_count_values<K: Value>(
579        self,
580        plan: &Plan<'_>,
581        reducer: Reducer,
582    ) -> Result<Aggregate<Vec<GroupCount<K>>>, Error> {
583        let (groups, page) = self.grouped(plan, reducer)?;
584        let (value, output_bytes) = count_output(groups, self.request.rows, self.request.bytes)?;
585        Ok(Aggregate {
586            value,
587            examined: page.examined,
588            fetched: page.fetched,
589            output_bytes,
590        })
591    }
592
593    fn count_reducer(&self, field: Option<&str>) -> Result<Reducer, Error> {
594        // Presence needs only the validated tag, not an ordered scalar slot.
595        let presence = field
596            .map(|name| {
597                E::SCHEMA
598                    .fields
599                    .iter()
600                    .enumerate()
601                    .find(|(_, f)| f.name == name)
602                    .map(|(i, f)| f.optional.then_some(i))
603                    .ok_or(Error::InvalidQuery)
604            })
605            .transpose()?
606            .flatten();
607        Ok(Reducer::Count { presence, total: 0 })
608    }
609
610    fn grouped(
611        &self,
612        plan: &Plan<'_>,
613        initial: Reducer,
614    ) -> Result<(Groups, QueryPage<Vec<u8>>), Error> {
615        if self.request.groups.page.is_some() {
616            return Err(Error::InvalidQuery);
617        }
618        let mut groups = self.admit_groups(plan, initial)?;
619        let page = execute(
620            &self.db.map,
621            E::SCHEMA,
622            &self.request,
623            plan,
624            Some(Accumulator::Groups(&mut groups)),
625        )?;
626        Ok((groups, page))
627    }
628
629    fn admit_groups(&self, plan: &Plan<'_>, initial: Reducer) -> Result<Groups, Error> {
630        Ok(Groups {
631            selection: self.request.groups.admit(E::SCHEMA, &initial)?,
632            mask: plan.projection.ok_or(Error::InvalidQuery)?,
633            initial,
634            rows: BTreeMap::new(),
635            state_bytes: 0,
636        })
637    }
638
639    fn aggregate_plan(
640        &self,
641        output: Option<&TypeSchema>,
642        group_output: bool,
643    ) -> Result<Plan<'static>, Error> {
644        if self.request.cursor.is_some()
645            || self.request.distinct
646            || self.request.order.is_some()
647            || self.request.descending
648        {
649            return Err(Error::InvalidQuery);
650        }
651        self.db.schema.admits(E::SCHEMA)?;
652        let mut plan = Plan::admit(
653            E::SCHEMA,
654            &self.request,
655            self.db.schema.fingerprint(),
656            self.db.incarnation,
657            output,
658            group_output,
659        )?;
660        if let Some(mask) = plan.projection {
661            admit_ordered_fields(E::SCHEMA, mask)?;
662            plan.selected |= mask;
663        }
664        Ok(plan)
665    }
666}
667
668fn value_output<K: Value, R: Value>(
669    groups: Groups,
670    rows: usize,
671    bytes: usize,
672) -> Result<(Vec<GroupValue<K, R>>, usize), Error> {
673    let mut output_bytes = 0;
674    let value = groups
675        .into_output()?
676        .into_iter()
677        .take(rows)
678        .map(|(key, reducer)| {
679            admit_output(key.len() + reducer.encoded_len(), bytes, &mut output_bytes)?;
680            decode_group_value(&key, reducer)
681        })
682        .collect::<Result<_, Error>>()?;
683    Ok((value, output_bytes))
684}
685
686fn count_output<K: Value>(
687    groups: Groups,
688    rows: usize,
689    bytes: usize,
690) -> Result<(Vec<GroupCount<K>>, usize), Error> {
691    let mut output_bytes = 0;
692    let value = groups
693        .into_output()?
694        .into_iter()
695        .take(rows)
696        .map(|(key, reducer)| {
697            admit_output(key.len() + 8, bytes, &mut output_bytes)?;
698            decode_group_count(&key, reducer)
699        })
700        .collect::<Result<_, Error>>()?;
701    Ok((value, output_bytes))
702}
703
704fn decode_group_count<K: Value>(key: &[u8], result: Completed) -> Result<GroupCount<K>, Error> {
705    Ok(GroupCount {
706        key: decode_record(key)?,
707        count: result.into_count()?,
708    })
709}
710
711fn decode_group_value<K: Value, R: Value>(
712    key: &[u8],
713    result: Completed,
714) -> Result<GroupValue<K, R>, Error> {
715    let value = result.into_value()?;
716    Ok(GroupValue {
717        key: decode_record(key)?,
718        value: value.as_deref().map(decode_record).transpose()?,
719    })
720}
721
722fn decode_record<R: Value>(bytes: &[u8]) -> Result<R, Error> {
723    let mut reader = Decoder::new(bytes)?;
724    let value = R::decode_from(&mut reader)?;
725    reader.finish()?;
726    Ok(value)
727}
728
729// Closed choices over delivered reducers and one grouping table, not user extensions.
730pub(super) enum Accumulator<'a> {
731    // Synchronous existence can complete on the first matching row; no job state.
732    Exists(&'a mut bool),
733    Scalar(&'a mut Reducer),
734    Groups(&'a mut Groups),
735}
736impl Accumulator<'_> {
737    pub(super) fn needs_spans(&self) -> bool {
738        !matches!(
739            self,
740            Self::Exists(_)
741                | Self::Scalar(
742                    Reducer::Numeric(_)
743                        | Reducer::Count { presence: None, .. }
744                        | Reducer::Distinct { .. }
745                )
746        )
747    }
748
749    pub(super) fn push(
750        &mut self,
751        entity: &EntitySchema,
752        values: &[IndexValue<'_>],
753        bytes: &[u8],
754        spans: &[(u16, u16)],
755    ) -> Result<(), Error> {
756        match self {
757            Self::Exists(found) => {
758                **found = true;
759                Ok(())
760            }
761            Self::Scalar(state) => state.push(entity, values, bytes, spans, MAX_STATE_BYTES),
762            Self::Groups(state) => state.push(entity, values, bytes, spans),
763        }
764    }
765}
766
767#[derive(Clone)]
768pub(super) enum Reducer {
769    Numeric(numeric::NumericAggregate),
770    Distinct {
771        mask: u16,
772        seen: BTreeSet<Vec<u8>>,
773    },
774    Count {
775        presence: Option<usize>,
776        total: u64,
777    },
778    Extremum {
779        position: usize,
780        maximum: bool,
781        selected: Option<(Vec<u8>, Vec<u8>)>,
782    },
783}
784impl Reducer {
785    fn distinct(mask: u16) -> Self {
786        Self::Distinct {
787            mask,
788            seen: BTreeSet::new(),
789        }
790    }
791
792    fn extremum(entity: &EntitySchema, mask: u16, maximum: bool) -> Result<Self, Error> {
793        if mask.count_ones() != 1 {
794            return Err(Error::InvalidQuery);
795        }
796        let position = mask.trailing_zeros() as usize;
797        if !key::supports(entity.fields[position].value) {
798            return Err(Error::InvalidQuery);
799        }
800        Ok(Self::Extremum {
801            position,
802            maximum,
803            selected: None,
804        })
805    }
806
807    fn retained_bytes(&self) -> usize {
808        match self {
809            Self::Numeric(state) => state.retained_bytes(),
810            Self::Count { .. } => 8,
811            Self::Distinct { seen, .. } => 8 + seen.iter().map(Vec::capacity).sum::<usize>(),
812            Self::Extremum { selected, .. } => selected
813                .as_ref()
814                .map_or(0, |(key, bytes)| key.capacity() + bytes.capacity()),
815        }
816    }
817
818    fn finish(self) -> Result<Completed, Error> {
819        match self {
820            Self::Count { total, .. } => Ok(Completed::Count(total)),
821            Self::Distinct { seen, .. } => Ok(Completed::Count(seen.len() as u64)),
822            Self::Extremum { selected, .. } => Ok(Completed::Value(selected)),
823            Self::Numeric(state) => state.finish(),
824        }
825    }
826
827    fn push(
828        &mut self,
829        entity: &EntitySchema,
830        values: &[IndexValue<'_>],
831        bytes: &[u8],
832        spans: &[(u16, u16)],
833        allowance: usize,
834    ) -> Result<(), Error> {
835        match self {
836            Self::Numeric(state) => state.push(values)?,
837            Self::Distinct { mask, seen } => {
838                let key = comparison_key(entity, *mask, values)?;
839                if !seen.contains(&key) {
840                    if seen.len() == MAX_DISTINCT_VALUES
841                        || 8 + seen.iter().map(Vec::capacity).sum::<usize>() + key.capacity()
842                            > allowance
843                    {
844                        return Err(Error::Limit);
845                    }
846                    seen.insert(key);
847                }
848            }
849            Self::Count { presence, total } => {
850                *total += u64::from(presence.is_none_or(|i| bytes[usize::from(spans[i].0)] != 0));
851            }
852            Self::Extremum {
853                position,
854                maximum,
855                selected,
856            } => {
857                let value = values[*position];
858                if value == IndexValue::Absent {
859                    return Ok(());
860                }
861                let mut key = Vec::new();
862                value.append_component(entity.fields[*position].optional, &mut key);
863                if key.len() > MAX_COMPARISON_KEY {
864                    return Err(Error::Limit);
865                }
866                if selected
867                    .as_ref()
868                    .is_none_or(|(old, _)| if *maximum { key > *old } else { key < *old })
869                {
870                    let (start, end) = spans[*position];
871                    if key.capacity() + usize::from(end - start) > allowance {
872                        return Err(Error::Limit);
873                    }
874                    // Copy only a winner, after admitting its replacement charge.
875                    let output = bytes[usize::from(start)..usize::from(end)].to_vec();
876                    if key.capacity() + output.capacity() > allowance {
877                        return Err(Error::Limit);
878                    }
879                    *selected = Some((key, output));
880                }
881            }
882        }
883        Ok(())
884    }
885}
886
887// Completed results cannot be advanced. All numeric range checks precede selection.
888enum Completed {
889    Count(u64),
890    Value(Option<(Vec<u8>, Vec<u8>)>),
891}
892impl Completed {
893    fn encoded_len(&self) -> usize {
894        match self {
895            Self::Count(_) => 8,
896            Self::Value(value) => value.as_ref().map_or(0, |(_, bytes)| bytes.len()),
897        }
898    }
899    // Counts need only eight stack bytes; extrema borrow their retained ordered key.
900    fn result_key<'a>(&'a self, scratch: &'a mut [u8; 8]) -> Option<&'a [u8]> {
901        match self {
902            Self::Count(total) => {
903                *scratch = total.to_be_bytes();
904                Some(scratch)
905            }
906            Self::Value(selected) => selected.as_ref().map(|(key, _)| key.as_slice()),
907        }
908    }
909
910    fn retained_bytes(&self) -> usize {
911        match self {
912            Self::Count(_) => 8,
913            Self::Value(value) => value
914                .as_ref()
915                .map_or(0, |(key, bytes)| key.capacity() + bytes.capacity()),
916        }
917    }
918    fn into_count(self) -> Result<u64, Error> {
919        match self {
920            Self::Count(count) => Ok(count),
921            _ => Err(Error::Corrupt),
922        }
923    }
924    fn into_value(self) -> Result<Option<Vec<u8>>, Error> {
925        match self {
926            Self::Value(value) => Ok(value.map(|(_, bytes)| bytes)),
927            _ => Err(Error::Corrupt),
928        }
929    }
930}
931
932pub(super) struct Groups {
933    selection: Selection,
934    mask: u16,
935    // Never advanced: each new group receives its own empty reducer state.
936    initial: Reducer,
937    rows: BTreeMap<Vec<u8>, (Vec<u8>, Reducer)>,
938    state_bytes: usize,
939}
940impl Groups {
941    fn into_output(self) -> Result<Vec<(Vec<u8>, Completed)>, Error> {
942        // Accumulation has completed. Output filtering cannot relax input/state bounds.
943        let mut state_bytes = self.state_bytes;
944        let mut rows = self
945            .rows
946            .into_iter()
947            .map(|(key, (record, reducer))| {
948                let previous = reducer.retained_bytes();
949                let result = reducer.finish()?;
950                state_bytes = state_bytes - previous + result.retained_bytes();
951                if state_bytes > MAX_STATE_BYTES {
952                    return Err(Error::Limit);
953                }
954                Ok((key, (record, result)))
955            })
956            .collect::<Result<Vec<_>, Error>>()?;
957        rows.retain(|(_, (_, result))| {
958            let mut scratch = [0; 8];
959            let actual = result.result_key(&mut scratch);
960            self.selection
961                .predicates
962                .split_inclusive(|(_, _, or_next)| !or_next)
963                .all(|clause| {
964                    clause.iter().any(|(comparison, operand, _)| {
965                        comparison.accepts(actual, operand.as_deref(), self.selection.optional)
966                    })
967                })
968        });
969        if self.selection.order == GroupOrder::KeyDescending {
970            // BTreeMap supplied ascending canonical keys; reverse the completed,
971            // filtered rows before any output limit or page is selected.
972            rows.reverse();
973        } else if self.selection.order != GroupOrder::Key {
974            rows.sort_unstable_by(|(ka, (_, a)), (kb, (_, b))| {
975                let (mut sa, mut sb) = ([0; 8], [0; 8]);
976                let order = a.result_key(&mut sa).cmp(&b.result_key(&mut sb));
977                let order = if self.selection.order == GroupOrder::AggregateDescending {
978                    order.reverse()
979                } else {
980                    order
981                };
982                order.then_with(|| ka.cmp(kb))
983            });
984        }
985        Ok(rows.into_iter().map(|(_, row)| row).collect())
986    }
987
988    fn push(
989        &mut self,
990        entity: &EntitySchema,
991        values: &[IndexValue<'_>],
992        bytes: &[u8],
993        spans: &[(u16, u16)],
994    ) -> Result<(), Error> {
995        let key = comparison_key(entity, self.mask, values)?;
996        let size = projection_size(spans, self.mask);
997        if let Some((_, reducer)) = self.rows.get_mut(&key) {
998            let previous = reducer.retained_bytes();
999            reducer.push(
1000                entity,
1001                values,
1002                bytes,
1003                spans,
1004                MAX_STATE_BYTES - self.state_bytes + previous,
1005            )?;
1006            self.state_bytes = self.state_bytes - previous + reducer.retained_bytes();
1007            return Ok(());
1008        }
1009        let base = key.capacity() + size;
1010        let available = MAX_STATE_BYTES - self.state_bytes;
1011        if self.rows.len() == MAX_GROUPS || base + self.initial.retained_bytes() > available {
1012            return Err(Error::Limit);
1013        }
1014        let mut reducer = self.initial.clone();
1015        reducer.push(entity, values, bytes, spans, available - base)?;
1016        let output = copy_projection(bytes, spans, self.mask, size);
1017        // Map/set nodes and fixed bookkeeping are independently bounded by the
1018        // group and per-group distinct cardinality caps.
1019        let retained = key.capacity() + output.capacity() + reducer.retained_bytes();
1020        if retained > available {
1021            return Err(Error::Limit);
1022        }
1023        self.state_bytes += retained;
1024        self.rows.insert(key, (output, reducer));
1025        Ok(())
1026    }
1027}