[kesmit13]

When a connection is opened with pure_python=False and the C extension failed to import, Connection.__init__ retries the import only so it can print why, then raises NotSupportedError. The retry misspelled the module (_singlestortedb_accel), so stderr always showed:

ModuleNotFoundError: No module named '_singlestortedb_accel'

That hid the real loader error, such as a missing symbol, the wrong architecture or a too-old glibc. It only matters with an explicit pure_python=False; the default None falls back to pure Python without printing anything. But that is exactly what publish.yml's wheel tests use (?pure_python=0 in SINGLESTOREDB_URL), so a broken wheel in CI would have reported the wrong cause.

Why this isn't a one-letter fix

Correcting the spelling of the import statement breaks every connection. An import statement anywhere in __init__ makes _singlestoredb_accel local to the whole function, so the _singlestoredb_accel is None check before it and the elif _singlestoredb_accel is not None after it raise UnboundLocalError. The typo was avoiding that, most likely by accident. This retries with importlib.import_module('_singlestoredb_accel') instead, which doesn't bind any name.

Tests

singlestoredb/tests/test_accel_import.py; neither test needs a server:

• test_reports_the_real_import_error makes the extension import fail with a specific ImportError by adding a finder to sys.meta_path, which an import statement and importlib both go through. It then checks that the error appears on stderr. It fails with the old typo.
• test_loads_when_available builds a pure_python=False connection (defer_connect=True) while the extension is available and checks that the C result class is used. It fails with the one-letter fix, raising UnboundLocalError. It's skipped when the extension isn't built.

I ran it locally against a singlestoredb-dev container. test_basics.py, test_accel_leaks.py and the new tests pass, and test_basics.py plus the new tests also pass with SINGLESTOREDB_PURE_PYTHON=1. Pre-commit is clean.

🤖 Generated with Claude Code

---

Note

Low Risk
Narrow change to error reporting and import retry when pure_python=False; behavior for default connections is unchanged.

Overview
When pure_python=False and the C extension did not load at import time, Connection.__init__ retries the load only to print the failure to stderr before raising NotSupportedError. That retry used a misspelled module name (_singlestortedb_accel), so CI and users saw ModuleNotFoundError instead of the real loader error (e.g. missing GLIBC).

The fix loads _singlestoredb_accel via importlib.import_module, which retries the correct module without introducing a local binding that would break the existing _singlestoredb_accel is None checks in __init__ (a plain import there would cause UnboundLocalError on every connection).

Adds test_accel_import.py: one test asserts stderr contains a simulated import failure; another (when the extension is built) asserts pure_python=False still selects the C result class.

^{Reviewed by Cursor Bugbot for commit ce6c69c. Bugbot is set up for automated code reviews on this repo. Configure here.}

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Copilot]

🟢 Approval recommended

No active review issues are reported for the change improving C-extension import diagnostics.

No open issues 🎉

_{Review effort: Balanced}