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