Stream aggregates with bounded memory instead of buffering all rows

.claude/CLAUDE.md

@@ -56,6 +56,7 @@ The main flow inside `executeSql({ tables, query, functions, signal })`:
 **Executor** (`src/execute/`):
 - `execute.js` - `executeSql()` and `executePlan()`, dispatches to node-specific executors
 - `join.js`, `aggregates.js`, `sort.js` - executors for join, aggregate, and sort nodes
+- `streamingAggregate.js` - bounded-memory aggregation for streamable aggregates (COUNT, COUNTIF, SUM, AVG, MIN, MAX); other aggregates buffer rows in `aggregates.js`
 
 **Expression Evaluator** (`src/expression/`):
 - `evaluate.js` - evaluates expression AST nodes against rows

src/execute/accumulator.js

@@ -0,0 +1,89 @@
+import { keyify } from './utils.js'
+
+/**
+ * @import { SqlPrimitive } from '../types.js'
+ */
+
+/**
+ * Incremental state for one streamable aggregate. Shared by the streaming
+ * aggregate executor and the scanColumn fast path so COUNT/COUNTIF/SUM/AVG/
+ * MIN/MAX fold semantics live in one place.
+ *
+ * @typedef {{
+ *   count: number,
+ *   sum: number,
+ *   min: SqlPrimitive,
+ *   max: SqlPrimitive,
+ *   seen: Set<unknown> | null,
+ * }} Accumulator
+ */
+
+/**
+ * @param {string} funcName
+ * @param {boolean} [distinct]
+ * @returns {Accumulator}
+ */
+export function newAccumulator(funcName, distinct) {
+  return {
+    count: 0,
+    sum: 0,
+    min: null,
+    max: null,
+    seen: funcName === 'COUNT' && distinct ? new Set() : null,
+  }
+}
+
+/**
+ * Folds one value into an accumulator, matching the buffered semantics in
+ * evaluate.js: COUNT counts non-null, COUNTIF counts truthy, MIN/MAX compare
+ * raw values, SUM/AVG only accumulate finite numbers.
+ *
+ * @param {string} funcName
+ * @param {Accumulator} acc
+ * @param {SqlPrimitive} value
+ */
+export function updateAccumulator(funcName, acc, value) {
+  switch (funcName) {
+  case 'COUNT':
+    if (value == null) break
+    if (acc.seen) acc.seen.add(keyify(value))
+    else acc.count++
+    break
+  case 'COUNTIF':
+    if (value) acc.count++
+    break
+  case 'MIN':
+    if (value != null && (acc.min === null || value < acc.min)) acc.min = value
+    break
+  case 'MAX':
+    if (value != null && (acc.max === null || value > acc.max)) acc.max = value
+    break
+  default: { // SUM, AVG
+    if (value == null) break
+    const num = Number(value)
+    if (Number.isFinite(num)) {
+      acc.sum += num
+      acc.count++
+    }
+  }
+  }
+}
+
+/**
+ * Reduces an accumulator to its final aggregate value.
+ *
+ * @param {string} funcName
+ * @param {Accumulator} acc
+ * @returns {SqlPrimitive}
+ */
+export function finalizeAccumulator(funcName, acc) {
+  switch (funcName) {
+  case 'COUNT': return acc.seen ? acc.seen.size : acc.count
+  case 'COUNTIF': return acc.count
+  case 'SUM': return acc.count === 0 ? null : acc.sum
+  case 'AVG': return acc.count === 0 ? null : acc.sum / acc.count
+  case 'MIN': return acc.min
+  case 'MAX': return acc.max
+  default: return null
+  }
+}

src/execute/aggregates.js

@@ -1,7 +1,9 @@
 import { derivedAlias } from '../expression/alias.js'
 import { evaluateExpr } from '../expression/evaluate.js'
+import { finalizeAccumulator, newAccumulator, updateAccumulator } from './accumulator.js'
 import { executePlan, selectColumnNames } from './execute.js'
 import { sortEntriesByTerms } from './sort.js'
+import { planStreamingAggregates, streamingHashAggregateRows, streamingScalarAggregateRows } from './streamingAggregate.js'
 import { keyify } from './utils.js'
 import { yieldToEventLoop } from './yield.js'
 
@@ -80,6 +82,14 @@ function aggregateContextRow(group, aggregateRow) {
  */
 export function executeHashAggregate(plan, context) {
   const child = executePlan({ plan: plan.child, context })
+  const streaming = planStreamingAggregates(plan, child.columns)
+  if (streaming) {
+    return {
+      columns: selectColumnNames(plan.columns, child.columns),
+      maxRows: child.maxRows,
+      rows: streamingHashAggregateRows({ plan, streaming, child, context }),
+    }
+  }
   return {
     columns: selectColumnNames(plan.columns, child.columns),
     maxRows: child.maxRows,
@@ -196,6 +206,15 @@ export function executeScalarAggregate(plan, context) {
   }
 
   const child = executePlan({ plan: plan.child, context })
+  const streaming = planStreamingAggregates(plan, child.columns)
+  if (streaming) {
+    return {
+      columns: selectColumnNames(plan.columns, child.columns),
+      numRows: plan.having ? undefined : 1,
+      maxRows: 1,
+      rows: streamingScalarAggregateRows({ plan, streaming, child, context }),
+    }
+  }
   return {
     columns: selectColumnNames(plan.columns, child.columns),
     numRows: plan.having ? undefined : 1,
@@ -344,17 +363,6 @@ function extractColumnAggSpec({ expr, alias }) {
   }
 }
 
-/**
- * @typedef {{
- *   spec: ColumnAggSpec,
- *   count: number,
- *   sum: number,
- *   min: SqlPrimitive,
- *   max: SqlPrimitive,
- *   seen: Set<unknown> | null,
- * }} AggAccumulator
- */
-
 /**
  * Scans a column once and computes every aggregate over it in a single pass.
  * All specs share the one scanColumn walk, so MIN(x)/MAX(x)/AVG(x) decode x once.
@@ -371,67 +379,24 @@ async function scanColumnGroup({ table, specs, limit, offset, signal }) {
   const { column } = specs[0]
   const values = table.scanColumn({ column, limit, offset, signal })
 
-  /** @type {AggAccumulator[]} */
-  const accs = specs.map(spec => /** @type {AggAccumulator} */ ({
-    spec,
-    count: 0,
-    sum: 0,
-    min: null,
-    max: null,
-    seen: spec.funcName === 'COUNT' && spec.distinct ? new Set() : null,
-  }))
+  const accs = specs.map(spec => ({ spec, acc: newAccumulator(spec.funcName, spec.distinct) }))
 
   for await (const chunk of values) {
     signal?.throwIfAborted()
     for (let i = 0; i < chunk.length; i++) {
       const v = chunk[i]
       if (v == null) continue
-      for (const acc of accs) {
-        switch (acc.spec.funcName) {
-        case 'COUNT':
-          if (acc.seen) acc.seen.add(keyify(v))
-          else acc.count++
-          break
-        case 'MIN':
-          if (acc.min === null || v < acc.min) acc.min = v
-          break
-        case 'MAX':
-          if (acc.max === null || v > acc.max) acc.max = v
-          break
-        default: { // SUM, AVG
-          const num = Number(v)
-          if (Number.isFinite(num)) {
-            acc.sum += num
-            acc.count++
-          }
-        }
-        }
+      for (const { spec, acc } of accs) {
+        updateAccumulator(spec.funcName, acc, v)
       }
     }
   }
   signal?.throwIfAborted()
 
   /** @type {Map<string, SqlPrimitive>} */
   const result = new Map()
-  for (const acc of accs) {
-    result.set(acc.spec.alias, finalizeAgg(acc))
+  for (const { spec, acc } of accs) {
+    result.set(spec.alias, finalizeAccumulator(spec.funcName, acc))
   }
   return result
 }
-
-/**
- * Reduces one accumulator to its final aggregate value.
- *
- * @param {AggAccumulator} acc
- * @returns {SqlPrimitive}
- */
-function finalizeAgg(acc) {
-  switch (acc.spec.funcName) {
-  case 'COUNT': return acc.seen ? acc.seen.size : acc.count
-  case 'SUM': return acc.count === 0 ? null : acc.sum
-  case 'AVG': return acc.count === 0 ? null : acc.sum / acc.count
-  case 'MIN': return acc.min
-  case 'MAX': return acc.max
-  default: return null
-  }
-}

src/execute/sort.js

@@ -4,16 +4,21 @@ import { executePlan } from './execute.js'
 import { compareForTerm } from './utils.js'
 
 /**
- * @import { AsyncRow, ExecuteContext, OrderByItem, QueryResults, SqlPrimitive } from '../types.js'
+ * @import { AsyncRow, ExecuteContext, ExprNode, OrderByItem, QueryResults, SqlPrimitive } from '../types.js'
  * @import { SortNode } from '../plan/types.js'
  */
 
 const MAX_CHUNK = 256
 
 /**
+ * When exprs is set, exprs[i] is evaluated for ORDER BY term i instead of the
+ * term's own expression, so callers can pre-substitute per-entry values (for
+ * example, finalized aggregates) while keeping tie-aware term evaluation.
+ *
  * @typedef {{
  *   row: AsyncRow,
  *   rows?: AsyncRow[],
+ *   exprs?: ExprNode[],
  * }} SortEntry
  */
 
@@ -63,7 +68,7 @@ export async function sortEntriesByTerms({ entries, orderBy, context, cacheValue
         const chunk = missing.slice(start, start + chunkSize)
         const values = await Promise.all(chunk.map(idx =>
           evaluateExpr({
-            node: term.expr,
+            node: entries[idx].exprs?.[orderByIdx] ?? term.expr,
             row: entries[idx].row,
             rows: entries[idx].rows,
             context,