[wellparth]

• This fix issue: archinstall freezes when selecting France for mirror region #4858

PR Description:

During mirror ranking, selecting regions with large numbers of mirrors (such as France or Germany) previously caused archinstall to hang or take several minutes when encountering unresponsive, slow, or flaky mirrors.

Root Cause:

1. Unguarded connection attempts: DownloadTimer was entered after urllib.request.urlopen(). Consequently, SIGALRM was not active during DNS resolution, TCP connection establishment, or TLS handshakes, relying solely on socket timeouts which could block for 5–10+ seconds per address.
2. Compound retries on failing mirrors: Mirrors encountering ConnectionResetError or IncompleteRead retried up to 3 times (_speedtest_retries = 3). For dead or flaky mirrors, this multiplied the delay significantly.
3. Overly generous timeout: Probing a ~131 KB core.db file had a 5-second socket timeout and 5-second timer. Healthy mirrors download this in <1.5s, while failing mirrors blocked for 5+ seconds each.

Changes:

• Wrap urlopen inside DownloadTimer: DownloadTimer is entered before urllib.request.urlopen, ensuring that DNS resolution, TCP connection, TLS handshake, headers, and body read are all strictly bounded by SIGALRM.
• Fast-fail without retries: Eliminated the retry loop on mirror speed probing. Unreachable, timing out, or flaky mirrors fail fast and are immediately marked with speed = 0 so the installer moves on to the next mirror without delay.
• Lower timeout to 3s: Reduced the socket and download timer timeout to 3 seconds, which is ample headroom for viable mirrors while minimizing wait time on unresponsive endpoints.
• Sequential sorting maintained: Maintained sequential speed checks as recommended by maintainers to avoid bandwidth contention skewing speed measurements.
• Guarded calculations: Guarded against ZeroDivisionError when timer.time == 0.
• Robust DownloadTimer: Handled AttributeError and ValueError safely when signal handling is unsupported.
• Unit Tests: Added test coverage in tests/test_mirrorlist.py for mirror speed ranking order, error/timeout handling, and timer behavior.

Tests and Checks

• I have tested the code!
  • Unit tests in tests/test_mirrorlist.py passing (6/6 passed).
  • Code formatting and linting verified with ruff check and ruff format (0 errors).

[Torxed]

Would this not cause an unfair speed measurement considering max_workers compete for the available speed?

[h8d13]

I don't think we want concurrency at all for this. Sequentially even with 100s of mirrors in a region, it should breeze through them (10-20 seconds), if it's getting stuck somewhere it's likely due to how we currently handle flaky/dead mirrors, not a need for workers.

If anything its additional packages that should be threaded (#4083)

The title of this PR also doesn't make sense, sorting happens after the TUI. I sense this might be automated with the escapes leaking in the PR xd

[wellparth]

Thanks for the feedback @Torxed and @h8d13!

I've updated the PR to pivot away from concurrency and address the root cause of the hang:

1. Sequential probing maintained: Removed ThreadPoolExecutor to keep bandwidth measurements fair and sequential.
2. Proper timeout bounding: DownloadTimer was previously entered after urlopen(), leaving DNS resolution and connection handshakes completely unguarded by SIGALRM. It now wraps urlopen directly with a 3s timeout so stalled endpoints fail fast.
3. No retries on failed probes: Dropped the retry loop on speed checks—unresponsive, dropping, or resetting mirrors are immediately assigned speed 0 and skipped rather than retrying up to 3 times.
4. Cleaned up PR: Updated the title and description to accurately reflect the fix.

[h8d13]

Scope looks much better IMO