SNOW-3192362: Fix auth callback recv blind sleep with MSG_DONTWAIT

[sfc-gh-fpawlowski]

SNOW-3192362: Fix auth callback recv blind sleep with MSG_DONTWAIT

Summary

• Replace time.sleep(cooldown) with select.select([client_socket], [], [], cooldown) in the MSG_DONTWAIT retry path of _try_receive_block, making the cooldown data-aware
• Fixes flaky test_auth_callback_success failures (particularly with dontwait=true, timeout=0.05) on loaded CI runners

Root Cause

In _try_receive_block (src/snowflake/connector/auth/_http_server.py:194-201), the MSG_DONTWAIT code path uses a blind time.sleep() between non-blocking recv attempts:

# recv with MSG_DONTWAIT raises BlockingIOError instantly if no data
except BlockingIOError:
    if attempt < attempts - 1:
        cooldown = min(attempt_timeout, 0.25) if attempt_timeout else 0.25
        time.sleep(cooldown)  # <-- BLIND: data arriving during sleep is invisible

With timeout=0.05 and max_attempts=15:

• attempt_timeout = 0.05 / 15 = ~3.3ms
• Each iteration: instant recv check (0ms useful work) + 3.3ms blind sleep
• 12 recv attempts = ~0ms of data-checking time, ~36ms of blind sleeping
• On a loaded CI runner, the data arrival window is easily missed entirely

Compare with the dontwait=false path: recv blocks for up to attempt_timeout, returning as soon as data arrives. Every millisecond is spent actively waiting for data.

The core issue: time.sleep() is not socket-aware. Data arriving on the socket during the sleep cannot wake it up. The process sleeps for the full cooldown duration regardless of whether data is available.

Fix

One-line change in src/snowflake/connector/auth/_http_server.py, line 201:

# Before:
    time.sleep(cooldown)

# After:
    select.select([client_socket], [], [], cooldown)

select.select() monitors the socket's read-readiness with a timeout of cooldown. If data arrives during the wait, select returns early and the next recv succeeds immediately. If no data arrives, select times out after cooldown — same wallclock behavior as the original time.sleep.

Why this is correct and safe

1. Only affects MSG_DONTWAIT path — BlockingIOError is never raised without MSG_DONTWAIT, so the dontwait=false path is completely untouched
2. Never runs on Windows — _use_msg_dont_wait() returns False on Windows (lines 26-30), so this code path is Unix-only where select works on sockets
3. select already imported and used — module-level import at line 9, used in _try_poll at line 180
4. Spurious wakeups handled — if select returns readable but recv still raises BlockingIOError, the existing retry loop handles it gracefully
5. No signature/API changes — only the internal sleep mechanism changes
6. time import stays — still used in the bind retry loop (line 116)

Test plan

• python -m pytest test/unit/test_auth_callback_server.py -v — all 318 tests pass
• python -m pytest test/unit/test_auth_callback_server.py -v -k "test_auth_callback_success" — the specific failing tests
• Full unit suite: python -m pytest test/unit/ -x --timeout=120