mbutrovich opened a new issue, #26948: URL: https://github.com/apache/datafusion/issues/26948
# `repeat` on `LargeUtf8` aborts the process on a result size no allocation can satisfy ### Describe the bug `repeat` caps the result at `i32::MAX` bytes for `Utf8` and `Utf8View`, and returns an error above that. For `LargeUtf8` the cap is `i64::MAX`, in both the [scalar path](https://github.com/apache/datafusion/blob/926901f9001977e85df617b96328801b42f6af4a/datafusion/functions/src/string/repeat.rs#L127-L128) and the [array path](https://github.com/apache/datafusion/blob/926901f9001977e85df617b96328801b42f6af4a/datafusion/functions/src/string/repeat.rs#L227-L228). That cap admits sizes no allocation can satisfy. A result of 2^60 bytes (1 EiB) is larger than the virtual address space of x86_64 (48 or 57 bits) and aarch64 (at most 52 bits), so the allocator refuses it at once, whatever memory the machine has: - With both arguments scalar, [`s.repeat(count)`](https://github.com/apache/datafusion/blob/926901f9001977e85df617b96328801b42f6af4a/datafusion/functions/src/string/repeat.rs#L167) cannot fail gracefully, so the refused allocation aborts the process: `memory allocation of 1152921504606846976 bytes failed`. - With an array of strings, the builder panics: `failed to allocate memory for layout Layout { size: 1152921504606846976, align: 64 (1 << 6) }`. The existing `test_repeat_string_array_overflow` uses a count of `i64::MAX`, which the overflow check catches before any allocation. ### To Reproduce On `main` at 926901f90: ```rust use std::sync::Arc; use arrow::datatypes::Field; use datafusion_common::ScalarValue; use datafusion_common::config::ConfigOptions; use datafusion_expr::{ColumnarValue, ReturnFieldArgs, ScalarFunctionArgs}; use datafusion_functions::string::repeat; let udf = repeat(); let s = ScalarValue::LargeUtf8(Some("a".to_string())); // With ScalarValue::Utf8: Err(Execution("string size overflow on repeat, max size is 2147483647, but got 1152921504606846976")) let n = ScalarValue::Int64(Some(1 << 60)); let arg_fields = vec![ Arc::new(Field::new("s", s.data_type(), true)), Arc::new(Field::new("n", n.data_type(), true)), ]; let return_field = udf .return_field_from_args(ReturnFieldArgs { arg_fields: &arg_fields, scalar_arguments: &[Some(&s), Some(&n)], }) .unwrap(); let result = udf.invoke_with_args(ScalarFunctionArgs { args: vec![ColumnarValue::Scalar(s), ColumnarValue::Scalar(n)], arg_fields, number_rows: 1, return_field, config_options: Arc::new(ConfigOptions::default()), }); // aborts: memory allocation of 1152921504606846976 bytes failed ``` ### Expected behavior An error, as for `Utf8`, instead of an abort. Two fixes seem possible: - A fixed cap for `LargeUtf8`. PostgreSQL's `repeat` returns "requested length too large" above [`MaxAllocSize`](https://github.com/postgres/postgres/blob/884aca9b96989b9bc1b65d42e731e90f098ee2ee/src/include/utils/memutils.h#L40), 1 GB minus 1 ([`oracle_compat.c`](https://github.com/postgres/postgres/blob/884aca9b96989b9bc1b65d42e731e90f098ee2ee/src/backend/utils/adt/oracle_compat.c#L1177-L1182)). DuckDB's strings are limited to `u32::MAX` bytes ([`MAX_STRING_SIZE`](https://github.com/duckdb/duckdb/blob/bca3953c739d90390fdbf9f1c295909d89d87fa8/src/include/duckdb/common/types/string_type.hpp#L31)). - Fallible allocation, such as `try_reserve`, so that a refused allocation becomes an error. Which one should `repeat` use? Neither guards against an allocation that succeeds under overcommit and fails later when the pages are touched. Both turn this case, where the allocator refuses up front, into an error. #22688 tracks checked size accounting for string builders in general. ### Additional context Found by property-based tests that call each function with generated arguments, including large counts. ### Willingness to contribute I cannot contribute a fix for this bug at this time -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
