Summary

Fixes #25.

FlushWorker.step decides whether to log an upload failure with:

if response.status_code == 500 and getattr(response, "exception") != None:

getattr(response, "exception") is called without a default. The two things step can receive are asymmetric:

• On a network error, Uploader.__call__ catches requests.RequestException and returns Fake500(e) — which has an .exception attribute.
• On a genuine server HTTP 500, session.post(...) returns a real requests.Response — which has no .exception attribute.

So on a real 500, status_code == 500 is True, then getattr(response, "exception") raises AttributeError. That propagates out of step() → run(), so the background flush thread dies. Every subsequent log is enqueued but never sent for the rest of the process — silent log loss. This matches the reporter's "happens occasionally" (real 500s are intermittent).

Fix

-            if response.status_code == 500 and getattr(response, "exception") != None:
+            if response.status_code == 500 and getattr(response, "exception", None) is not None:

A real Response now yields None and skips the branch; the network-error Fake500 path still logs its exception as before. (Also switched != None → is not None to match idiom.)

Tests

Two regression tests in tests/test_flusher.py:

• test_step_does_not_crash_on_real_server_500 — returns a real requests.Response with status_code = 500; step() must not raise. This fails on the current code with the exact AttributeError: 'Response' object has no attribute 'exception' from Flusher fails to log: AttributeError: 'Response' object has no attribute 'exception' #25.
• test_step_still_logs_network_error_fake500 — the Fake500 network-error path still logs.

⚠️ Worth noting: the existing tests use mock.MagicMock(status_code=500), and MagicMock auto-vivifies .exception, which is why they never caught this — the new test deliberately uses a real Response.

Full tests/test_flusher.py suite passes (11 tests).