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/calc/src/eval.rs

@@ -39,15 +39,6 @@ const MAX_DEPTH: usize = 400;
 /// every legitimate call; hostile work that grows past one dimension (a LAMBDA that recurses
 /// or nest `MAP` into `MAP`) then stops at the same bound instead of squaring.
 const LAMBDA_FUEL: u64 = MAX_CELLS;
-/// Work one cell's formula may do: one unit per expression evaluated, per cell read from a
-/// reference, per array value handed to a function, per pair of areas intersected and per
-/// character of a defined name's formula parsed where the name is used. Each limit
-/// above bounds one kind of step, but a compact formula could still repeat bounded steps
-/// exponentially often (defined names that each use the one before twice, a computed criterion
-/// whose database function reruns it, an aggregate over a large array inside every LAMBDA call).
-/// Past this the formula is `#CALC!`. It is far above what a formula that reads every cell of a
-/// full-size array a few times needs.
-const EVAL_WORK: u64 = 16 * MAX_CELLS;
 
 #[derive(Clone, Copy, Debug, PartialEq, Eq, Hash)]
 pub struct Area {
@@ -148,16 +139,6 @@ pub struct Evaluator<'h> {
     /// runs out of work after the same number of calls that a legitimately wide one could
     /// make, instead of an amount of work that squares with array nesting (#2382).
     fuel: u64,
-    /// Work this cell still has left (see [`EVAL_WORK`]), shared by every branch of the formula
-    /// and by the formulas it computes on the way (names, computed criteria).
-    work: u64,
-    /// A spend found too little work left. A spend that leaves exactly nothing is fine, so this
-    /// can't be told from `work` alone; a database function checks it after its computed
-    /// criteria ran, whose failures it can't otherwise see.
-    exhausted: bool,
-    /// Computed criteria being evaluated, outermost first: one that is reached again from inside
-    /// itself counts as not met instead of repeating its work at every level.
-    active_criteria: Vec<(usize, CellRef)>,
 }
 
 struct FnCtx<'a, 'h> {
@@ -187,9 +168,6 @@ impl Ctx for FnCtx<'_, '_> {
     fn computed_criterion(&mut self, row: usize, col: usize, db: &Array, record: usize) -> Option<bool> {
         let area = self.criteria?;
         let at = area.range.start.offset(row as i64, col as i64).filter(|c| area.range.contains(*c))?;
-        if self.ev.active_criteria.contains(&(area.sheet, at)) {
-            return Some(false);
-        }
         let expr = self.ev.host.workbook().sheet(area.sheet)?.cell(at)?.formula.as_ref()?.expr_arc()?;
         let shifted = gridcraft_formula::adjust::shift_relative((*expr).clone(), record as i64 - 1, 0);
         let env = (0..db.cols)
@@ -209,19 +187,10 @@ impl Ctx for FnCtx<'_, '_> {
             calls: self.ev.calls,
             stack_top: self.ev.stack_top,
             fuel: self.ev.fuel,
-            work: self.ev.work,
-            exhausted: false,
-            active_criteria: std::mem::take(&mut self.ev.active_criteria),
         };
-        ev.active_criteria.push((area.sheet, at));
         let v = matches!(ev.value(&shifted).scalar().to_bool(), Ok(true));
         // The child's spend counts against this cell's budget.
         self.ev.fuel = ev.fuel;
-        self.ev.work = ev.work;
-        self.ev.exhausted |= ev.exhausted;
-        let mut active = std::mem::take(&mut ev.active_criteria);
-        active.pop();
-        self.ev.active_criteria = active;
         Some(v)
     }
     fn has_lambda(&self) -> bool {
@@ -241,42 +210,7 @@ impl Ctx for FnCtx<'_, '_> {
 
 impl<'h> Evaluator<'h> {
     pub fn new(host: &'h mut dyn Host, sheet: usize, at: CellRef) -> Self {
-        Evaluator {
-            host,
-            sheet,
-            at,
-            env: Vec::new(),
-            depth: 0,
-            calls: 0,
-            stack_top: 0,
-            fuel: LAMBDA_FUEL,
-            work: EVAL_WORK,
-            exhausted: false,
-            active_criteria: Vec::new(),
-        }
-    }
-
-    /// Spends `units` of this cell's work; `false` (and nothing left) when there isn't enough.
-    pub(crate) fn spend(&mut self, units: u64) -> bool {
-        match self.work.checked_sub(units) {
-            Some(w) => {
-                self.work = w;
-                true
-            }
-            None => {
-                self.work = 0;
-                self.exhausted = true;
-                false
-            }
-        }
-    }
-
-    /// Charges the values of an array (one unit per value; one for anything else).
-    pub(crate) fn spend_on(&mut self, v: &Value) -> bool {
-        match v {
-            Value::Array(a) => self.spend(a.data.len() as u64),
-            _ => self.spend(1),
-        }
+        Evaluator { host, sheet, at, env: Vec::new(), depth: 0, calls: 0, stack_top: 0, fuel: LAMBDA_FUEL }
     }
 
     /// Evaluates to a value (references are dereferenced; multi-cell → array).
@@ -326,27 +260,18 @@ impl<'h> Evaluator<'h> {
         }
         // A reference keeps its size, so the array has its shape and its blanks count
         // (`=A1:A10`, `SUMPRODUCT(--(A1:A10=""))`).
-        let range = self.read_range(a);
-        let (h, w) = (range.height() as usize, range.width() as usize);
-        if (h as u64) * (w as u64) > MAX_CELLS {
-            return Value::Error(CellError::Num);
-        }
-        if !self.spend((h as u64) * (w as u64)) {
-            return Value::Error(CellError::Calc);
-        }
-        let data = self.host.range_values(a.sheet, range);
-        Array::new(h, w, data).map(Value::from).unwrap_or(Value::Error(CellError::Value))
-    }
-
-    /// The block that reading `a` gives: `a` itself, or when it is large (or runs to the sheet's
-    /// edge) the part inside the sheet's used range.
-    pub(crate) fn read_range(&mut self, a: &Area) -> RangeRef {
         let cells = u64::from(a.range.height()) * u64::from(a.range.width());
         // `A:A`, `1:1` and `A2:A1048576` run to the edge of the sheet: they stand for "the rest of
         // the column", not for a block of a chosen size.
         let to_edge = a.range.end.row == MAX_ROWS - 1 || a.range.end.col == MAX_COLS - 1;
         let full_size = cells <= FULL_SIZE_CELLS || (!to_edge && cells <= FULL_SIZE_BOUNDED_CELLS);
-        if full_size { a.range } else { self.trim(a) }
+        let range = if full_size { a.range } else { self.trim(a) };
+        let (h, w) = (range.height() as usize, range.width() as usize);
+        if (h as u64) * (w as u64) > MAX_CELLS {
+            return Value::Error(CellError::Num);
+        }
+        let data = self.host.range_values(a.sheet, range);
+        Array::new(h, w, data).map(Value::from).unwrap_or(Value::Error(CellError::Value))
     }
 
     /// Trims the bottom/right of an area to the sheet's used range (keeps the top-left so
@@ -379,9 +304,6 @@ impl<'h> Evaluator<'h> {
             self.host.stack_exhausted();
             return err(CellError::Num);
         }
-        if !self.spend(1) {
-            return err(CellError::Calc);
-        }
         self.depth += 1;
         if self.depth > MAX_DEPTH {
             self.depth -= 1;
@@ -448,11 +370,6 @@ impl<'h> Evaluator<'h> {
         if let Some(def) = wb.name(bare, scope_sheet) {
             let text = def.formula.clone();
             let scope = def.scope.unwrap_or(self.sheet);
-            // Each use parses and evaluates the definition again: names that each use the one
-            // before twice double the work per name, so it is charged.
-            if !self.spend(text.len() as u64) {
-                return err(CellError::Calc);
-            }
             return match gridcraft_formula::parse_name(&text) {
                 Ok(expr) => {
                     // Relative references in a name are relative to A1: used here, they move by
@@ -645,10 +562,6 @@ impl<'h> Evaluator<'h> {
                 let (x, y) = (self.eval(a), self.eval(b));
                 match (x, y) {
                     (Ev::R(mut xa), Ev::R(ya)) => {
-                        // A union reused through LET doubles at each step: its areas count.
-                        if !self.spend((xa.len() as u64).saturating_add(ya.len() as u64)) {
-                            return err(CellError::Calc);
-                        }
                         xa.extend(ya);
                         Ev::R(xa)
                     }
@@ -660,11 +573,6 @@ impl<'h> Evaluator<'h> {
                 let (x, y) = (self.eval(a), self.eval(b));
                 match (x, y) {
                     (Ev::R(xa), Ev::R(ya)) => {
-                        // Every pair is looked at, and each may be kept: charged before the loop,
-                        // so unions squared through LET can't build an enormous list.
-                        if !self.spend((xa.len() as u64).saturating_mul(ya.len() as u64)) {
-                            return err(CellError::Calc);
-                        }
                         let mut out = Vec::new();
                         for p in &xa {
                             for q in &ya {
@@ -758,113 +666,6 @@ mod fuel_tests {
         ev.value(&gridcraft_formula::parser::parse(src).unwrap_or_else(|e| panic!("{e}")))
     }
 
-    /// Evaluates `src` in `wb` with `work` units of work, and the work left.
-    fn eval_work(wb: Workbook, work: u64, src: &str) -> (Value, u64) {
-        let mut host = Empty { wb };
-        let mut ev = Evaluator::new(&mut host, 0, CellRef::new(0, 0));
-        ev.work = work;
-        let v = ev.value(&gridcraft_formula::parser::parse(src).unwrap_or_else(|e| panic!("{e}")));
-        (v, ev.work)
-    }
-
-    fn is_calc(v: &Value) -> bool {
-        matches!(v, Value::Error(CellError::Calc))
-    }
-
-    /// Serves the workbook's stored values, for database functions to read.
-    struct Stored {
-        wb: Workbook,
-    }
-    impl Host for Stored {
-        fn workbook(&self) -> &Workbook {
-            &self.wb
-        }
-        fn cell_value(&mut self, sheet: usize, c: CellRef) -> Value {
-            self.wb.sheet(sheet).and_then(|s| s.cell(c)).map_or(Value::Empty, |cell| cell.value.clone())
-        }
-        fn now_serial(&self) -> f64 {
-            0.0
-        }
-        fn random(&mut self) -> f64 {
-            0.0
-        }
-    }
-
-    /// A computed criterion that runs out of work counts as not met for its record, which the
-    /// database function can't see: it must answer #CALC!, not a count that leaves records out.
-    #[test]
-    fn a_computed_criterion_out_of_work_makes_the_database_function_calc() {
-        use gridcraft_model::{Cell, Formula};
-        let mut wb = Workbook::new();
-        let name = |name: String, formula: String| gridcraft_model::DefinedName { name, scope: None, formula, comment: String::new(), hidden: false };
-        wb.names.push(name("Step0".into(), "1".into()));
-        for i in 1..=60 {
-            wb.names.push(name(format!("Step{i}"), format!("Step{}+Step{}", i - 1, i - 1)));
-        }
-        let at = |s: &str| CellRef::parse(s).unwrap();
-        let sheet = wb.sheet_mut(0).unwrap();
-        for (r, (n, age)) in [("name", Value::text("age")), ("a", Value::number(1.0)), ("b", Value::number(2.0))].into_iter().enumerate() {
-            sheet.cells.set(at(&format!("A{}", r + 1)), Cell::value(Value::text(n)));
-            sheet.cells.set(at(&format!("B{}", r + 1)), Cell::value(age));
-        }
-        sheet.cells.set(at("D1"), Cell::value(Value::text("Formula")));
-        let host_wb = |criterion: &str| {
-            let mut wb = wb.clone();
-            // A criteria cell is read by its value first (a blank one matches everything): give it
-            // the value a recalculation would have stored.
-            let mut cell = Cell::formula(Formula::new(criterion));
-            cell.value = Value::Bool(true);
-            wb.sheet_mut(0).unwrap().cells.set(at("D2"), cell);
-            wb
-        };
-        let run = |criterion: &str| {
-            let mut host = Stored { wb: host_wb(criterion) };
-            let mut ev = Evaluator::new(&mut host, 0, CellRef::new(9, 9));
-            ev.work = 100_000;
-            ev.value(&gridcraft_formula::parser::parse("=DCOUNT(A1:B3,\"age\",D1:D2)").unwrap())
-        };
-        assert_eq!(run("=B2>0"), Value::number(2.0), "a cheap criterion counts every record");
-        assert!(is_calc(&run("=Step60>0")), "{:?}", run("=Step60>0"));
-    }
-
-    #[test]
-    fn names_that_double_at_each_step_run_out_of_work() {
-        // Step1 = Step0+Step0, Step2 = Step1+Step1, ...: Step60 would evaluate Step0 2^60 times.
-        let mut wb = Workbook::default();
-        let name = |name: String, formula: String| gridcraft_model::DefinedName { name, scope: None, formula, comment: String::new(), hidden: false };
-        wb.names.push(name("Step0".into(), "1".into()));
-        for i in 1..=60 {
-            wb.names.push(name(format!("Step{i}"), format!("Step{}+Step{}", i - 1, i - 1)));
-        }
-        let (v, _) = eval_work(wb.clone(), EVAL_WORK, "=Step5");
-        assert_eq!(v, Value::number(32.0));
-        let (v, left) = eval_work(wb, 100_000, "=Step60");
-        assert!(is_calc(&v), "{v:?}");
-        assert_eq!(left, 0);
-    }
-
-    #[test]
-    fn intersections_of_unions_are_charged_per_pair() {
-        let wb = Workbook::default();
-        // 4 x 4 pairs, 4 of which meet.
-        let src = "=AREAS((A1,B1,C1,D1) (A1,B1,C1,D1))";
-        let (v, _) = eval_work(wb.clone(), EVAL_WORK, src);
-        assert_eq!(v, Value::number(4.0));
-        let (v, _) = eval_work(wb, 20, src);
-        assert!(is_calc(&v), "{v:?}");
-    }
-
-    #[test]
-    fn an_aggregate_inside_every_lambda_call_spends_work() {
-        let wb = Workbook::default();
-        // 100 calls, each summing the 10,000 values of a captured array: about a million units.
-        let src = "=LET(big,SEQUENCE(10000),SUM(MAP(SEQUENCE(100),LAMBDA(x,SUM(big)))))";
-        let (v, _) = eval_work(wb.clone(), 2_000_000, src);
-        assert_eq!(v, Value::number(100.0 * 50_005_000.0));
-        let (v, _) = eval_work(wb, 200_000, src);
-        assert!(is_calc(&v), "{v:?}");
-    }
-
     #[test]
     fn lambda_calls_spend_fuel_and_run_out_quietly() {
         // Three applications, one charge each: 3 inside the budget. (A nested MAP wouldn't do

crates/calc/src/eval/call.rs

@@ -353,21 +353,8 @@ impl<'h> Evaluator<'h> {
             };
             evaluated.push(Arg { value, from_ref });
         }
-        // The function may go through every value it is given (a SUM inside each call of a
-        // LAMBDA over a large array): the values count against the cell's work.
-        let values = evaluated.iter().fold(0u64, |n, a| n.saturating_add(if let Value::Array(x) = &a.value { x.data.len() as u64 } else { 1 }));
-        if !self.spend(values) {
-            return err(CellError::Calc);
-        }
-        let database = criteria.is_some();
         let mut ctx = FnCtx { ev: self, criteria, lambda };
-        let v = gridcraft_functions::call(spec, &evaluated, &mut ctx);
-        // A computed criterion that ran out of work counted as not met: the result leaves
-        // records out.
-        if database && self.exhausted {
-            return err(CellError::Calc);
-        }
-        Ev::V(v)
+        gridcraft_functions::call(spec, &evaluated, &mut ctx).pipe(Ev::V)
     }
 
     /// Performs aggregation operations without copying its ranges: blocks are folded over the stored
@@ -396,12 +383,6 @@ impl<'h> Evaluator<'h> {
     /// Adds one area to `acc`; `Err` is the result the function ends with.
     fn aggregate_area(&mut self, kind: Agg, a: &Area, acc: &mut Acc) -> Result<(), Ev> {
         if !a.range.is_single() {
-            // Folded without building an array, but still walked: the part inside the used
-            // range counts against the cell's work.
-            let r = self.read_range(a);
-            if !self.spend(u64::from(r.height()) * u64::from(r.width())) {
-                return Err(err(CellError::Calc));
-            }
             let mut error = None;
             let walked = self.host.visit_range(a.sheet, a.range, &mut |v| match acc.add_cell(kind, v) {
                 Ok(()) => true,
@@ -421,9 +402,6 @@ impl<'h> Evaluator<'h> {
     /// Adds one argument's value by calling `SUM` and `COUNT` on it, so coercion and error rules
     /// stay in one place (`AVERAGE` of an argument is its `SUM` over its `COUNT`).
     fn aggregate_value(&mut self, kind: Agg, value: Value, from_ref: bool, acc: &mut Acc) -> Result<(), Ev> {
-        if !from_ref && !self.spend_on(&value) {
-            return Err(err(CellError::Calc));
-        }
         let (Some(sum), Some(count)) = (gridcraft_functions::lookup("SUM"), gridcraft_functions::lookup("COUNT")) else {
             return Err(err(CellError::Name));
         };

crates/calc/src/eval/reference_fns.rs

@@ -118,29 +118,11 @@ impl<'h> Evaluator<'h> {
                 for c in 0..w {
                     let pos = (num(&Some(ra.get_broadcast(r, c))), num(&Some(ca.get_broadcast(r, c))));
                     let v = match pos {
-                        // A position that picks several cells is #CALC!. That is decided from the
-                        // shape: reading the block first, for every position, let a short formula
-                        // read one large range over and over only to discard it.
-                        (Ok(row), Ok(col)) if index_picks_several_values(&src, row, col, two) => Value::Error(CellError::Calc),
                         (Ok(row), Ok(col)) => {
                             let one = self.index_one(src.clone(), row, col, area_n, two);
-                            let cells = match &one {
-                                Ev::R(areas) => match areas.as_slice() {
-                                    [a] => self.area_cells(a),
-                                    _ => 0,
-                                },
-                                _ => 0,
-                            };
-                            if cells > MAX_CELLS {
-                                // As reading it would answer.
-                                Value::Error(CellError::Num)
-                            } else if cells > 1 {
-                                Value::Error(CellError::Calc)
-                            } else {
-                                match self.deref(one) {
-                                    Value::Array(a) if a.rows * a.cols > 1 => Value::Error(CellError::Calc),
-                                    v => v.scalar(),
-                                }
+                            match self.deref(one) {
+                                Value::Array(a) if a.rows * a.cols > 1 => Value::Error(CellError::Calc),
+                                v => v.scalar(),
                             }
                         }
                         (Err(e), _) | (_, Err(e)) => Value::Error(e),
@@ -157,16 +139,6 @@ impl<'h> Evaluator<'h> {
         self.index_one(src, row, col, area_n, two)
     }
 
-    /// Cells that reading `a` gives: its own size, or for a large block the part inside the
-    /// sheet's used range (what [`Evaluator::area_value`] reads).
-    fn area_cells(&mut self, a: &Area) -> u64 {
-        if a.range.is_single() {
-            return 1;
-        }
-        let r = self.read_range(a);
-        u64::from(r.height()) * u64::from(r.width())
-    }
-
     /// INDEX for one row and column position (0 = the whole column or row).
     fn index_one(&mut self, src: Ev, row: i64, col: i64, area_n: i64, two_args: bool) -> Ev {
         if row < 0 || col < 0 {
@@ -435,71 +407,3 @@ impl<'h> Evaluator<'h> {
         Ev::V(gridcraft_functions::aggregate_values(func, &values))
     }
 }
-
-/// Whether INDEX on an array picks a whole row or column of several values at position
-/// (`row`, `col`), which the array-of-positions form answers with #CALC!.
-fn index_picks_several_values(src: &Ev, row: i64, col: i64, two_args: bool) -> bool {
-    let Ev::V(Value::Array(arr)) = src else { return false };
-    let (row, col) = if two_args && arr.rows == 1 { (1, row) } else { (row, col) };
-    if row < 0 || col < 0 || row as u64 > arr.rows as u64 || col as u64 > arr.cols as u64 {
-        return false;
-    }
-    (row == 0 && arr.rows > 1) || (col == 0 && arr.cols > 1)
-}
-
-#[cfg(test)]
-mod tests {
-    use super::super::Host;
-    use super::*;
-    use gridcraft_model::Workbook;
-
-    /// A sheet of blanks that counts the blocks it is asked for.
-    struct Counting {
-        wb: Workbook,
-        cells_read: usize,
-    }
-    impl Host for Counting {
-        fn workbook(&self) -> &Workbook {
-            &self.wb
-        }
-        fn cell_value(&mut self, _sheet: usize, _c: CellRef) -> Value {
-            Value::Empty
-        }
-        fn now_serial(&self) -> f64 {
-            0.0
-        }
-        fn random(&mut self) -> f64 {
-            0.0
-        }
-        fn range_values(&mut self, _sheet: usize, range: RangeRef) -> Vec<Value> {
-            self.cells_read += range.count() as usize;
-            vec![Value::Empty; range.count() as usize]
-        }
-    }
-
-    fn eval(src: &str) -> (Value, usize) {
-        let mut host = Counting { wb: Workbook::default(), cells_read: 0 };
-        let mut ev = Evaluator::new(&mut host, 0, CellRef::new(0, 0));
-        let v = ev.value(&gridcraft_formula::parser::parse(src).unwrap_or_else(|e| panic!("{e}")));
-        (v, host.cells_read)
-    }
-
-    fn all_calc(v: &Value, n: usize) -> bool {
-        matches!(v, Value::Array(a) if a.data.len() == n && a.data.iter().all(|x| *x == Value::Error(CellError::Calc)))
-    }
-
-    #[test]
-    fn index_positions_that_pick_several_cells_are_calc_without_reading_them() {
-        // Each zero row position picks the whole of A1:A50000: #CALC!, decided from the shape. It
-        // used to read the 50,000 cells once per position before discarding them.
-        let (v, read) = eval("=INDEX(A1:A50000,{0,0,0,0},1)");
-        assert!(all_calc(&v, 4), "{v:?}");
-        assert_eq!(read, 0);
-        // The same for an array source: a whole column of a 3x3 array.
-        let (v, _) = eval("=INDEX({1,2,3;4,5,6;7,8,9},{0,0},2)");
-        assert!(all_calc(&v, 2), "{v:?}");
-        // Single cells are still read.
-        let (v, _) = eval("=INDEX({1,2,3;4,5,6;7,8,9},{1,3},2)");
-        assert!(matches!(&v, Value::Array(a) if a.data == vec![Value::number(2.0), Value::number(8.0)]), "{v:?}");
-    }
-}

crates/calc/src/tests.rs

@@ -2372,19 +2372,3 @@ fn formulas_read_a_dirty_python_cell_after_it_is_laid_out() {
     assert_eq!((t.num("B1"), t.num("B2")), (5.0, 6.0));
     assert_eq!((t.num("A1"), t.num("A2"), t.num("A3")), (5.0, 100.0, 111.0));
 }
-
-/// A computed criterion whose formula runs a database function over the same criteria counts as
-/// not met inside itself: before, each record's check ran the function again for every record, at
-/// every level, so the work grew exponentially with the nesting allowed.
-#[test]
-fn computed_criterion_that_reenters_itself_is_not_met() {
-    let mut t = T::new();
-    for (r, (name, age)) in [("name", "age"), ("a", "1"), ("b", "2"), ("c", "3"), ("d", "4"), ("e", "5")].iter().enumerate() {
-        t.set(&format!("F{}", r + 2), name);
-        t.set(&format!("G{}", r + 2), age);
-    }
-    t.set("N46", "Formula");
-    t.set("N47", "=DCOUNT($F$2:$G$7,\"age\",$N$46:$N$47)>0");
-    t.set("B60", "=DCOUNT($F$2:$G$7,\"age\",N46:N47)");
-    assert_eq!(t.num("B60"), 0.0);
-}

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}");
+    }
+}