Get Data: refuse an import whose dense rows would be too large

data.getData and data.fromTextCsv built every row densely before
writing it: JSON records were padded to the union of all keys (found
with a quadratic Vec::contains scan), and CSV cells were expanded over
the whole used rectangle. A few sparse records or one far-away CSV
cell therefore allocated billions of slots.

The importer now collects JSON keys with a HashSet, refuses more keys
than a sheet has columns, and checks the projected rows x columns
against MAX_IMPORT_CELLS (10,000,000) before building any row.

crates/engine/src/cmd/extra/data.rs

@@ -42,6 +42,15 @@ pub(super) fn table_resize(s: &mut Session, p: &Json) -> Result<Json> {
     })
 }
 
+/// Most cells (rows × columns, blanks included) Get Data imports. The rows are built densely
+/// before they are written, so a few sparse records or one far-away CSV cell must not expand
+/// into a rectangle of billions of slots.
+const MAX_IMPORT_CELLS: u64 = 10_000_000;
+
+fn too_many_cells() -> EngineError {
+    EngineError::Other(format!("The data is too large to import ({MAX_IMPORT_CELLS} cells at most)."))
+}
+
 pub(super) fn from_csv(s: &mut Session, p: &Json) -> Result<Json> {
     let bytes = if let Some(t) = str_param(p, "text") {
         t.as_bytes().to_vec()
@@ -57,15 +66,22 @@ pub(super) fn from_csv(s: &mut Session, p: &Json) -> Result<Json> {
         if text.trim_start().starts_with('[') {
             let arr: Vec<Json> = serde_json::from_str(&text).map_err(|e| bad("data.getData", e.to_string()))?;
             let mut keys: Vec<String> = Vec::new();
+            let mut seen: std::collections::HashSet<&str> = std::collections::HashSet::new();
             for o in &arr {
                 if let Some(m) = o.as_object() {
                     for k in m.keys() {
-                        if !keys.contains(k) {
+                        if seen.insert(k.as_str()) {
                             keys.push(k.clone());
                         }
                     }
                 }
             }
+            if keys.len() > gridcraft_core::MAX_COLS as usize {
+                return Err(EngineError::Other("The data has more fields than a sheet has columns.".into()));
+            }
+            if (arr.len() as u64 + 1).saturating_mul(keys.len().max(1) as u64) > MAX_IMPORT_CELLS {
+                return Err(too_many_cells());
+            }
             let mut out = vec![keys.iter().map(|k| json!(k)).collect::<Vec<_>>()];
             for o in &arr {
                 out.push(keys.iter().map(|k| o.get(k).cloned().unwrap_or(Json::Null)).collect());
@@ -79,6 +95,9 @@ pub(super) fn from_csv(s: &mut Session, p: &Json) -> Result<Json> {
             let wb = gridcraft_xlsx::read_csv(&bytes, &opts).map_err(|e| EngineError::Other(e.to_string()))?;
             let sh = wb.sheet(0).ok_or(EngineError::NoDocument)?;
             let Some(u) = sh.used_range() else { return Err(EngineError::Other("The file is empty.".into())) };
+            if u.count() > MAX_IMPORT_CELLS {
+                return Err(too_many_cells());
+            }
             (u.start.row..=u.end.row)
                 .map(|r| {
                     (u.start.col..=u.end.col)
@@ -129,3 +148,35 @@ pub(super) fn queries(s: &mut Session, _: &Json) -> Result<Json> {
             .collect(),
     ))
 }
+
+#[cfg(test)]
+mod tests {
+    use serde_json::json;
+
+    use crate::Session;
+
+    fn s() -> Session {
+        let mut s = Session::new();
+        s.new_workbook();
+        s
+    }
+
+    #[test]
+    fn sparse_imports_are_refused_before_their_rows_are_built() {
+        let mut s = s();
+        let sheets = s.doc().unwrap().wb.sheets.len();
+        // 5,000 records with one field each: 5,001 × 5,000 slots once every record gets every field.
+        let records: Vec<_> = (0..5_000).map(|i| json!({ format!("k{i}"): i })).collect();
+        let text = serde_json::to_string(&records).unwrap();
+        let err = s.execute("data.getData", json!({"text": text})).expect_err("too many cells");
+        assert!(err.to_string().contains("too large"), "{err}");
+        // A CSV whose last line reaches 16,001 columns, 1,000 lines down.
+        let text = format!("1{}{}x", "\n".repeat(999), ",".repeat(16_000));
+        let err = s.execute("data.fromTextCsv", json!({"text": text})).expect_err("too many cells");
+        assert!(err.to_string().contains("too large"), "{err}");
+        assert_eq!(s.doc().unwrap().wb.sheets.len(), sheets);
+        // Small imports still work.
+        let out = s.execute("data.getData", json!({"text": r#"[{"a": 1}, {"b": 2}]"#})).unwrap();
+        assert_eq!((out["rows"].as_u64(), out["cols"].as_u64()), (Some(3), Some(2)), "{out}");
+    }
+}

crates/functions/src/database.rs

@@ -70,42 +70,27 @@ fn cell_criterion(v: &Value) -> Option<Criterion> {
 /// when the label is unknown, criterion).
 type CriteriaRow = (usize, Vec<(usize, Option<usize>, Criterion)>);
 
-/// Most condition tests one selection may make (records times conditions): every record can be
-/// tested against every criteria row when none matches, so a modest database and criteria table
-/// multiply into billions of tests. Past it the function is `#NUM!`, as lifted lookups are; 5e8
-/// is a few seconds of matching.
-const SELECTION_WORK_BUDGET: usize = 500_000_000;
-
 /// Rows (1-based within the database, excluding the header) of matching records.
-fn matching_records(db: &Array, crit: &Array, ctx: &mut dyn Ctx) -> R<Vec<usize>> {
-    // The database column under each criteria label, found once per label rather than per cell.
-    let columns: Vec<Option<usize>> = (0..crit.cols)
-        .map(|c| match crit.get(0, c) {
-            Some(Value::Text(t)) => find_column(db, t),
-            Some(Value::Number(n)) => find_column(db, &gridcraft_core::number_to_text(*n)),
-            _ => None,
-        })
-        .collect();
+fn matching_records(db: &Array, crit: &Array, ctx: &mut dyn Ctx) -> Vec<usize> {
     // Per criteria row: its index and a list of (criteria column, database column or None when
     // the label is unknown, criterion).
     let mut rows: Vec<CriteriaRow> = Vec::new();
-    let mut tests = 0usize;
     for r in 1..crit.rows {
         let mut conds = Vec::new();
         for c in 0..crit.cols {
             let Some(cell) = crit.get(r, c) else { continue };
             let Some(cr) = cell_criterion(cell) else { continue };
-            conds.push((c, columns.get(c).copied().flatten(), cr));
+            let col = match crit.get(0, c) {
+                Some(Value::Text(t)) => find_column(db, t),
+                Some(Value::Number(n)) => find_column(db, &gridcraft_core::number_to_text(*n)),
+                _ => None,
+            };
+            conds.push((c, col, cr));
         }
-        tests = tests.saturating_add(conds.len().max(1));
         rows.push((r, conds));
     }
-    // Computed criteria (a formula evaluated per record) count like any other test.
-    if db.rows.saturating_sub(1).saturating_mul(tests) > SELECTION_WORK_BUDGET {
-        return Err(CellError::Num);
-    }
     let all_match = rows.is_empty();
-    Ok((1..db.rows)
+    (1..db.rows)
         .filter(|&rec| {
             all_match
                 || rows.iter().any(|(r, conds)| {
@@ -117,7 +102,7 @@ fn matching_records(db: &Array, crit: &Array, ctx: &mut dyn Ctx) -> R<Vec<usize>
                     })
                 })
         })
-        .collect())
+        .collect()
 }
 
 struct Selection<'a> {
@@ -138,7 +123,7 @@ fn select<'a>(a: &'a [Arg], ctx: &mut dyn Ctx) -> R<Selection<'a>> {
     let db = as_array(dbv);
     let crit = as_array(critv);
     let field = field_index(&db, a)?;
-    let records = matching_records(&db, &crit, ctx)?;
+    let records = matching_records(&db, &crit, ctx);
     Ok(Selection { db, field, records })
 }
 
@@ -367,24 +352,6 @@ mod tests {
         is_err(d("DSUM", n(9.0)), CellError::Value);
     }
 
-    /// Security: records times criteria conditions past the selection budget is `#NUM!` before
-    /// any record is tested; a nonmatching criteria table used to test every record against
-    /// every criteria row.
-    #[test]
-    fn selection_work_is_bounded() {
-        use gridcraft_core::Array;
-        let column = |label: &str, len: usize, v: Value| {
-            let mut data = vec![tv(label)];
-            data.extend(std::iter::repeat_n(v, len));
-            Value::from(Array::column(data))
-        };
-        let db = column("Tree", 100_000, tv("Apple"));
-        let wide = column("Tree", 5_001, tv("=Plum"));
-        is_err(ev("DCOUNTA", vec![rf(db.clone()), t("Tree"), rf(wide)]), CellError::Num);
-        let narrow = column("Tree", 2, tv("=Apple"));
-        close(ev("DCOUNTA", vec![rf(db), t("Tree"), rf(narrow)]), 100_000.0);
-    }
-
     #[test]
     fn simple_criteria() {
         let c = |cells: Vec<Value>| arr(vec![vec![tv("Tree"), tv("Height")], cells]);

crates/functions/src/date.rs

@@ -416,17 +416,13 @@ fn workday_core(a: &[Arg], c: &mut dyn Ctx, mask: [bool; 7], hol: Option<&Arg>)
     if days == 0 {
         return num_val(start as f64);
     }
-    // The holidays on working days, still sorted (`holidays` sorts them), so each step counts
-    // the ones it passed with two binary searches rather than a scan of every holiday: a run of
-    // consecutive holidays takes one step per holiday.
-    let hs: Vec<i64> = hs.into_iter().filter(|&h| !mask[weekday0(sys, h)]).collect();
     let dir = days.signum();
     let mut cur = start;
     let mut rem = days.abs();
     while rem > 0 {
         let next = advance(sys, cur, rem, dir, &mask, per_week);
         let (lo, hi) = if dir > 0 { (cur + 1, next) } else { (next, cur - 1) };
-        rem = hs.partition_point(|&h| h <= hi).saturating_sub(hs.partition_point(|&h| h < lo)) as i64;
+        rem = hs.iter().filter(|&&h| h >= lo && h <= hi && !mask[weekday0(sys, h)]).count() as i64;
         cur = next;
         if !(0..=MAX_SERIAL as i64).contains(&cur) {
             return Err(CellError::Num);
@@ -759,23 +755,6 @@ mod tests {
         is_err(ev("TIMEVALUE", vec![t("x")]), CellError::Value);
     }
 
-    /// Security: WORKDAY counts the holidays each step passes by binary search, so a long run of
-    /// consecutive holidays is linear work, not one scan of every holiday per holiday.
-    #[test]
-    fn workday_over_many_consecutive_holidays() {
-        let start = d(2000.0, 1.0, 3.0);
-        let hol: Vec<f64> = (1..=200_000).map(|i| start + f64::from(i)).collect();
-        let out = ev("WORKDAY.INTL", vec![n(start), n(1.0), n(11.0), av(col(&hol))]);
-        // Sundays are the only weekend (11): 200,000 days of holidays are skipped, landing on the
-        // first working day after them.
-        let after = start + 200_001.0;
-        let sunday = ev("WEEKDAY", vec![n(after), n(2.0)]) == nv(7.0);
-        close(out, if sunday { after + 1.0 } else { after });
-        // Holidays on a weekend don't push the result out.
-        let fri = d(2026.0, 10.0, 9.0);
-        close(ev("WORKDAY", vec![n(fri), n(1.0), av(col(&[fri + 1.0, fri + 3.0]))]), fri + 4.0);
-    }
-
     #[test]
     fn day_counts() {
         close(ev("DAYS", vec![n(d(2021.0, 3.0, 15.0)), n(d(2021.0, 2.0, 1.0))]), 42.0);

crates/functions/src/dynamic.rs

@@ -178,10 +178,6 @@ fn sort(args: &[Arg], _c: &mut dyn Ctx) -> R<Value> {
         return Err(CellError::Value);
     }
     let orders: Vec<i64> = (0..idx.len()).map(|i| *ord.get(i).or(ord.first()).unwrap_or(&1)).collect();
-    // A column repeated in sort_index can't break a tie its first use left, so only first uses
-    // stay: the keys copied per row are then at most the row itself, however long sort_index is.
-    let mut seen = std::collections::HashSet::new();
-    let (idx, orders): (Vec<i64>, Vec<i64>) = idx.into_iter().zip(orders).filter(|(i, _)| seen.insert(*i)).unzip();
     let mut lines: Vec<(Vec<Value>, Vec<Value>)> =
         rows_of(&a).into_iter().map(|r| (idx.iter().map(|&i| r.get(i as usize - 1).cloned().unwrap_or(Value::Empty)).collect(), r)).collect();
     lines.sort_by(|x, y| cmp_keys(&x.0, &y.0, &orders));
@@ -838,16 +834,9 @@ fn pivotby(args: &[Arg], ctx: &mut dyn Ctx) -> R<Value> {
     let ckeys: Vec<Vec<Value>> = p.rows.iter().map(|&r| key_row(&cf, r)).collect();
     let rk = distinct_keys(&rkeys, if row_sort.unsigned_abs() as usize <= rf.cols { row_sort } else { 0 });
     let ck = distinct_keys(&ckeys, if col_sort.unsigned_abs() as usize <= cf.cols { col_sort } else { 0 });
-    // Source rows of every row key, column key and pair, built in one pass, so a total reads
-    // its rows instead of rescanning the whole source once per total.
     let mut cell: HashMap<(String, String), Vec<usize>> = HashMap::new();
-    let mut by_row: HashMap<String, Vec<usize>> = HashMap::new();
-    let mut by_col: HashMap<String, Vec<usize>> = HashMap::new();
     for ((r, c), &i) in rkeys.iter().zip(ckeys.iter()).zip(p.rows.iter()) {
-        let (r, c) = (row_key(r), row_key(c));
-        by_row.entry(r.clone()).or_default().push(i);
-        by_col.entry(c.clone()).or_default().push(i);
-        cell.entry((r, c)).or_default().push(i);
+        cell.entry((row_key(r), row_key(c))).or_default().push(i);
     }
     let mut agg = |rows: &[usize], c: usize| -> Value {
         if rows.is_empty() {
@@ -857,12 +846,14 @@ fn pivotby(args: &[Arg], ctx: &mut dyn Ctx) -> R<Value> {
         aggregate_with(f, &vs, ctx)
     };
     let rows_for = |r: Option<&Vec<Value>>, c: Option<&Vec<Value>>| -> Vec<usize> {
-        match (r, c) {
-            (Some(r), Some(c)) => cell.get(&(row_key(r), row_key(c))).cloned().unwrap_or_default(),
-            (Some(r), None) => by_row.get(&row_key(r)).cloned().unwrap_or_default(),
-            (None, Some(c)) => by_col.get(&row_key(c)).cloned().unwrap_or_default(),
-            (None, None) => p.rows.clone(),
-        }
+        let rkey = r.map(|k| row_key(k));
+        let ckey = c.map(|k| row_key(k));
+        p.rows
+            .iter()
+            .zip(rkeys.iter().zip(ckeys.iter()))
+            .filter(|(_, (a, b))| rkey.as_ref().is_none_or(|k| *k == row_key(a)) && ckey.as_ref().is_none_or(|k| *k == row_key(b)))
+            .map(|(&i, _)| i)
+            .collect()
     };
     let mut cols: Vec<Option<&Vec<Value>>> = ck.iter().map(Some).collect();
     if col_total > 0 {
@@ -912,7 +903,10 @@ fn pivotby(args: &[Arg], ctx: &mut dyn Ctx) -> R<Value> {
             }
         };
         for c in &cols {
-            let rows = rows_for(r, *c);
+            let rows = match (r, c) {
+                (Some(rk), Some(ck)) => cell.get(&(row_key(rk), row_key(ck))).cloned().unwrap_or_default(),
+                _ => rows_for(r, *c),
+            };
             for vc in 0..vals.cols {
                 line.push(agg(&rows, vc));
             }
@@ -1116,21 +1110,6 @@ mod tests {
         assert_eq!(rows_of(&r), vec![vec![nv(0.0), nv(5.0)], vec![nv(1.0), nv(2.0)], vec![nv(1.0), nv(1.0)]]);
     }
 
-    /// Security: a column repeated in sort_index keys the sort once. The keys copied per row stay
-    /// within the row, however long sort_index is, and a repeat (even with the other order) can't
-    /// change the result, since its first use already ordered the rows.
-    #[test]
-    fn sort_drops_repeated_sort_indices() {
-        let m = || av(arr(vec![vec![nv(1.0), nv(2.0)], vec![nv(1.0), nv(1.0)], vec![nv(0.0), nv(5.0)]]));
-        let want = vec![vec![nv(0.0), nv(5.0)], vec![nv(1.0), nv(2.0)], vec![nv(1.0), nv(1.0)]];
-        let r = ev("SORT", vec![m(), av(row(&[1.0, 1.0, 2.0, 1.0])), av(row(&[1.0, -1.0, -1.0, 1.0]))]);
-        assert_eq!(rows_of(&r), want);
-        let idx = vec![1.0; 1_000_000];
-        let tall = av(gridcraft_core::Array::filled(2_000, 1, nv(1.0)).into());
-        let r = ev("SORT", vec![tall, av(row(&idx))]);
-        assert_eq!(rows_of(&r).len(), 2_000);
-    }
-
     #[test]
     fn sortby() {
         let r = ev("SORTBY", vec![av(col(&[1.0, 2.0, 3.0])), av(col(&[30.0, 10.0, 20.0]))]);
@@ -1302,16 +1281,6 @@ mod tests {
         );
     }
 
-    /// Security: PIVOTBY totals read their source rows from indexes built in the grouping pass,
-    /// rather than rescanning the source once per total, so many distinct keys stay linear.
-    #[test]
-    fn pivotby_totals_with_many_groups() {
-        let keys: Vec<f64> = (0..20_000).map(f64::from).collect();
-        let r = rows_of(&ev("PIVOTBY", vec![av(col(&keys)), av(col(&vec![1.0; 20_000])), av(col(&vec![2.0; 20_000])), t("SUM")]));
-        assert_eq!(r.get(1), Some(&vec![nv(0.0), nv(2.0), nv(2.0)]));
-        assert_eq!(r.last(), Some(&vec![tv("Total"), nv(40_000.0), nv(40_000.0)]));
-    }
-
     #[test]
     fn pivotby_size_check_counts_column_header_rows() {
         // One row, n column fields and n value columns: n header rows + value-name row + body row + total row,