You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Add related source code locations to errors #13662
Is your feature request related to a problem or challenge?
DataFusionError does not provide information about the location inside the SQL query that caused it. It's difficult for the end user to understand where, in a large query, to go look for the cause of the error.
Example:
WITH products_with_price AS (
SELECT
id,
name,
price_usd
FROM
products
JOIN
prices
ONproducts.id=prices.product_id
)
SELECT
id,
name,
price_usd +'$'AS display_price_usd
FROM
products_with_price
When creating a logical plan for this query, we'd return:
Cannot coerce arithmetic expression Int64 + Utf8 to valid types
It's up to the user to figure out that the error lies on the third-to-last line.
Describe the solution you'd like
With the AST nodes from the parser now containing the source code location (thanks to apache/datafusion-sqlparser-rs#1435), it now becomes possible to add the same information to errors. This could seriously improve the experience for end users, since they would be able to clearly see where in the queries the errors lie.
In the example above, the desired feature of this ticket would be to add a reference to the third-to-last line, which is the cause of the error: price_usd + '$' AS display_price_usd.
Furthermore, it would be helpful to link to additonal segments of code note that, despite not representing the root cause of the error, help the user understand the context around the error and the elements involved. In this example, we'd highlight the projected price_usd column in the CTE to help the user understand where it comes from.
Describe alternatives you've considered
DataFusionError::Context only adds additional debug strings, but doesn't attach source code location info.
That's true, and I think it's an invaluable tool for debugging and developing DataFusion or software that relies on it. But in my opinion it's not really the best end user facing error reporting solution.
It's not really readable on its own (for end users, not developers)
Can't extract a nicer message from it
References DataFusion's source code locations, not the SQL queries' locations
Take this trace:
3: datafusion_common::error::DataFusionError::get_back_trace
at /Users/eper/code/datafusion/datafusion/common/src/error.rs:410:30
4: datafusion_sql::utils::check_column_satisfies_expr
at /Users/eper/code/datafusion/datafusion/sql/src/utils.rs:153:16
5: datafusion_sql::utils::check_columns_satisfy_exprs
at /Users/eper/code/datafusion/datafusion/sql/src/utils.rs:141:18
6: datafusion_sql::select::<impl datafusion_sql::planner::SqlToRel<S>>::aggregate
at /Users/eper/code/datafusion/datafusion/sql/src/select.rs:807:9
7: datafusion_sql::select::<impl datafusion_sql::planner::SqlToRel<S>>::select_to_plan
at /Users/eper/code/datafusion/datafusion/sql/src/select.rs:200:13
8: datafusion_sql::query::<impl datafusion_sql::planner::SqlToRel<S>>::query_to_plan
at /Users/eper/code/datafusion/datafusion/sql/src/query.rs:55:28
9: datafusion_sql::statement::<impl datafusion_sql::planner::SqlToRel<S>>::sql_statement_to_plan_with_context_impl
at /Users/eper/code/datafusion/datafusion/sql/src/statement.rs:220:40
10: datafusion_sql::statement::<impl datafusion_sql::planner::SqlToRel<S>>::sql_statement_to_plan
at /Users/eper/code/datafusion/datafusion/sql/src/statement.rs:183:9
11: datafusion_sql::statement::<impl datafusion_sql::planner::SqlToRel<S>>::statement_to_plan
at /Users/eper/code/datafusion/datafusion/sql/src/statement.rs:171:42
12: datafusion::execution::session_state::SessionState::statement_to_plan::{{closure}}
at /Users/eper/code/datafusion/datafusion/core/src/execution/session_state.rs:560:9
13: datafusion::execution::session_state::SessionState::create_logical_plan::{{closure}}
at /Users/eper/code/datafusion/datafusion/core/src/execution/session_state.rs:593:54
The query was SELECT name FROM users GROUP BY id, so the error is that name must be in the GROUP BY. The closest hint we have from the backtrace is the call to SqlToRel::aggregate but that's it.
With the changes that I'm contributing in #13664, the same query would produce diagnostics data that look like this:
(line 1, column 8) error: 'users.name' in projection does not appear in GROUP BY clause
(line 1, column 33) note: GROUP BY clause is here
(line 1, column 33) Help: Add 'users.name' to GROUP BY clause
(line 1, column 8) error: 'users.name' in projection does not appear in GROUP BY clause
(line 1, column 33) note: GROUP BY clause is here
(line 1, column 33) Help: Add 'users.name' to GROUP BY clause
(line 1, column 8) error: 'users.name' in projection does not appear in GROUP BY clause
(line 1, column 33) note: GROUP BY clause is here
(line 1, column 33) Help: Add 'users.name' to GROUP BY clause
I think the parser and the logical planner capture different categories of errors. They do it already now, just not in a user-friendly way.
The parser would catch errors such as a syntactically invalid query (e.g. SELECT id name FROM users, missing a comma), while the logical planner would catch more sophisticated errors that appear in a grammatically correct query (e.g. SELECT id + name FROM users, trying to do an arithmetic sum between an integer and a string).
Though, you are right that the source location information must come from the parser, and we have contributed that here datafusion-sqlparser-rs#1435. It's just that the parser alone doesn't understand how the AST tree is being interpreted, the logical planner does that.
One key consideration is that in #13664 I'm enhancing errors from the logical planner only: syntactically invalid queries would still result in a user-unfriendly error from datafusion-sqlparser-rs. Perhaps we can, at some point, do the same to datafusion-sqlparser-rs too?
Activity
comphead commentedon Dec 5, 2024
Datafusion supports backtraces to find where the problem happens in the first place https://datafusion.apache.org/user-guide/crate-configuration.html#enable-backtraces
eliaperantoni commentedon Dec 6, 2024
That's true, and I think it's an invaluable tool for debugging and developing DataFusion or software that relies on it. But in my opinion it's not really the best end user facing error reporting solution.
Take this trace:
The query was
SELECT name FROM users GROUP BY id, so the error is thatnamemust be in theGROUP BY. The closest hint we have from the backtrace is the call toSqlToRel::aggregatebut that's it.With the changes that I'm contributing in #13664, the same query would produce diagnostics data that look like this:
comphead commentedon Dec 6, 2024
Thanks @eliaperantoni
But this more looks for as a parse challenge 🤔 shouldn't be it in the parser phase?
https://github.com/apache/datafusion-sqlparser-rs
eliaperantoni commentedon Dec 9, 2024
I think the parser and the logical planner capture different categories of errors. They do it already now, just not in a user-friendly way.
The parser would catch errors such as a syntactically invalid query (e.g.
SELECT id name FROM users, missing a comma), while the logical planner would catch more sophisticated errors that appear in a grammatically correct query (e.g.SELECT id + name FROM users, trying to do an arithmetic sum between an integer and a string).Though, you are right that the source location information must come from the parser, and we have contributed that here datafusion-sqlparser-rs#1435. It's just that the parser alone doesn't understand how the AST tree is being interpreted, the logical planner does that.
One key consideration is that in #13664 I'm enhancing errors from the logical planner only: syntactically invalid queries would still result in a user-unfriendly error from datafusion-sqlparser-rs. Perhaps we can, at some point, do the same to datafusion-sqlparser-rs too?