DuckDB 0.9.0 query_timeout pragma – impact on cancellation and interrupt_on_timeout behavior

DuckDB 0.9.0 query_timeout pragma: interrupt_on_timeout and safe-point cancellation

When the PRAGMA query_timeout expires, DuckDB dispatches an interrupt that travels the query's cancellation pathway. The interrupt is only inspected at safe points embedded in the execution plan. Below are the three behavioral questions submitted, with the likely explanation separated from confirmed facts.

1. Setting interrupt_on_timeout = 0: the interrupt signal is still dispatched when the timeout elapses, but the query continues running until the current execution step completes. The timeout error is postponed, not suppressed permanently; once the step finishes, the error is raised. In practice this can increase resource usage because the query keeps processing.

2. Default interrupt_on_timeout = 1: the query is aborted immediately when the timeout is reached. DuckDB raises a runtime error indicating that the query exceeded the time limit. The exact message text depends on the client driver, but it consistently signals that no partial results are returned.

3. Safe-point check and observed duration: because the interrupt is only checked at safe points, a query that spends most of its time in a tight loop or an external callback may continue past the configured timeout until the next safe point is reached. The delay equals the time until the next safe point, which can be unpredictable without inspecting the plan.

To verify the behavior in your environment, run the following scoped commands (replace 5 with your desired seconds):

PRAGMA query_timeout = 5;SELECT * FROM large_table;

Repeat with PRAGMA interrupt_on_timeout = false; and compare the point at which the error surfaces. If the timeout appears significantly longer than the configured value, the query likely contains steps without safe points.

If you observe a timeout duration that does not match the configured seconds with a specific query pattern, please share the query text and its execution plan. That single diagnostic detail would help determine whether the delay is due to safe-point placement or another factor.