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