CI: build and run sqllogictests binary directly in extended workflow

[kosiew]

Which issue does this PR close?

• Closes Extended tests now take over 2 hours to run #20052.

Rationale for this change

Extended CI runs have been taking 2+ hours, with sqllogictests identified as a major contributor to the regression (thanks to @nuno-faria for identifying the contributor). This PR adjusts the extended workflow to separate building the sqllogictests test binary from executing it, allowing the run step to execute the already-built test executable directly.

This helps reduce workflow overhead from re-invoking cargo test for execution (and makes it easier to distinguish compilation time from test runtime when investigating performance regressions).

What changes are included in this PR?

• In the sqllogictests job:

  • Build the sqllogictests test binary once using cargo test build and capture the produced executable path.
  • Run the captured test binary directly from datafusion/sqllogictest with --include-sqlite.
  • Fail fast with a clear error if the test executable cannot be located.

Are these changes tested?

Yes.

Before fix

source: without current PR

After fix

source: current PR

(These changes are CI-only; functional behavior of DataFusion is unchanged.)

Are there any user-facing changes?

No. This PR only updates CI workflow behavior.

LLM-generated code disclosure

This PR includes LLM-generated code and comments. All LLM-generated content has been manually reviewed and tested.

[alamb]

Rather than cargo test another idea would be to use cargo build 🤔

Also, what is the reason to use --message-format=json?

Finally, what is the head command for? I didn't see any obvious reason in

https://github.com/apache/datafusion/actions/runs/21890300850/job/63194519053

Perhaps you could add a comment explaining what those commands are for

[kosiew]

Thanks, good suggestions. I agree cargo build is clearer here since this step only compiles the sqllogictests target and does not run it.

I used --message-format=json to capture the exact emitted test binary path because Cargo places test executables under target/.../deps with a hash suffix. The head -n 1 was a simple way to select the first extracted executable path from the stream.

I’ll simplify and document this more clearly in the workflow so the intent is explicit.

[alamb]

Why the ( and )?
Also, why does it need to do cd when the current version doesn't?

[kosiew]

The cd is needed because sqllogictests resolves test data via relative paths (for example test_files/ and ../../datafusion-testing/data/ in datafusion/sqllogictest/bin/sqllogictests.rs), so running from repo root can point it at the wrong locations. I added it after an earlier version without cd errored with test_files/ not found.

The subshell (...) was only used to scope cd so it would not affect the subsequent cargo clean. It is not strictly necessary.

I’ll remove the subshell and switch to a cleaner step-level working-directory: datafusion/sqllogictest for the test execution step.

[alamb]

thanks for looking into this @kosiew