linux-mirror-bench 0.2.0__tar.gz → 0.3.0__tar.gz

{linux_mirror_bench-0.2.0 → linux_mirror_bench-0.3.0}/.github/workflows/integration.yml

@@ -28,8 +28,8 @@ jobs:
     runs-on: ubuntu-latest
     timeout-minutes: 15
     steps:
-      - uses: actions/checkout@v6
-      - uses: actions/setup-python@v6
+      - uses: actions/checkout@v7
+      - uses: actions/setup-python@v7
         with:
           python-version: ${{ env.PYTHON_VERSION }}
       - uses: astral-sh/setup-uv@v7

linux_mirror_bench-0.3.0/.release-please-manifest.json

@@ -0,0 +1,3 @@
+{
+  ".": "0.3.0"
+}

{linux_mirror_bench-0.2.0 → linux_mirror_bench-0.3.0}/AGENTS.md

@@ -92,6 +92,9 @@ Run a single test: `uv run pytest tests/test_scorer.py::test_weights_shift_ranki
 ## Adding a new discoverer
 
 1. `src/mirror_bench/discovery/<name>.py` implementing `MirrorDiscoverer`.
+   `probe_path` must name a small file that upstream rewrites as packages
+   change, and `sync_time` must read its generation time, or stale mirrors
+   cannot be detected.
 2. Register in `discovery/__init__.py::DISCOVERERS`.
 3. Add ID aliases in `distro.py::_ALIASES`.
 4. Respx-mocked test in `tests/test_discovery.py` + fixture file under

{linux_mirror_bench-0.2.0 → linux_mirror_bench-0.3.0}/CHANGELOG.md

@@ -5,6 +5,24 @@ All notable changes to this project are documented here.
 The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and
 this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [0.3.0](https://github.com/MysticRyuujin/mirror-bench/compare/v0.2.0...v0.3.0) (2026-11-13)
+
+
+### Features
+
+* measure mirror freshness and drop stale mirrors ([#18](https://github.com/MysticRyuujin/mirror-bench/issues/18)) ([462ac42](https://github.com/MysticRyuujin/mirror-bench/commit/462ac42017f1026c6892701aaa0452b791a0c498))
+
+
+### Bug Fixes
+
+* make completion work without click, and reject error pages as probes ([#19](https://github.com/MysticRyuujin/mirror-bench/issues/19)) ([0c2427d](https://github.com/MysticRyuujin/mirror-bench/commit/0c2427dcb949f394e0118a386407d215d732465b))
+* try https for mirrors listed as http, and write plain output with no TTY ([#17](https://github.com/MysticRyuujin/mirror-bench/issues/17)) ([07d38ee](https://github.com/MysticRyuujin/mirror-bench/commit/07d38eec00a0c5669a728723a13a9d805f6113c8))
+
+
+### Dependencies
+
+* **actions:** match action versions in integration.yml ([#15](https://github.com/MysticRyuujin/mirror-bench/issues/15)) ([35da73d](https://github.com/MysticRyuujin/mirror-bench/commit/35da73d052acf4f1ea67c2f14f4df0b09a8b8c98))
+
 ## [0.2.0](https://github.com/MysticRyuujin/mirror-bench/compare/v0.1.2...v0.2.0) (2026-11-13)
 
 

{linux_mirror_bench-0.2.0 → linux_mirror_bench-0.3.0}/Dockerfile

@@ -20,16 +20,22 @@ ENV PYTHONDONTWRITEBYTECODE=1 \
 COPY --from=uv /uv /usr/local/bin/uv
 
 WORKDIR /app
-COPY pyproject.toml README.md LICENSE ./
+COPY pyproject.toml uv.lock README.md LICENSE ./
 COPY src ./src
 
-# Build the wheel, then install it (+ runtime deps) into a self-contained venv
-# under /opt/venv. No dev deps, no sources, no uv binary in the final image.
+# Build the wheel, then install it into a self-contained venv under /opt/venv.
+# Runtime deps come from uv.lock, so the image runs the same versions that CI
+# tests; resolving at build time would pull whatever is newest that day.
+# No dev deps, no sources, no uv binary in the final image.
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv venv /opt/venv \
+ && uv export --frozen --no-dev --no-emit-project --format requirements.txt \
+      --output-file /tmp/requirements.txt \
+ && VIRTUAL_ENV=/opt/venv uv pip install --no-cache --require-hashes \
+      -r /tmp/requirements.txt \
  && uv build --wheel \
  && wheel=$(find dist -maxdepth 1 -name '*.whl' -print -quit) \
- && VIRTUAL_ENV=/opt/venv uv pip install --no-cache "${wheel}"
+ && VIRTUAL_ENV=/opt/venv uv pip install --no-cache --no-deps "${wheel}"
 
 # ---- Runtime stage -------------------------------------------------------
 FROM python:${PYTHON_VERSION}-slim-bookworm AS runtime

{linux_mirror_bench-0.2.0 → linux_mirror_bench-0.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: linux-mirror-bench
-Version: 0.2.0
+Version: 0.3.0
 Summary: Cross-distribution Linux package mirror benchmarking tool
 Project-URL: Homepage, https://github.com/MysticRyuujin/mirror-bench
 Project-URL: Issues, https://github.com/MysticRyuujin/mirror-bench/issues
@@ -144,6 +144,7 @@ Running `mirror-bench` with no subcommand is equivalent to `mirror-bench bench`.
 | `--tls13-only`                | Hard-exclude mirrors that don't negotiate TLS 1.3.                                       |
 | `--weights lat=…,thr=…,sec=…` | Override scoring weights (auto-normalized).                                              |
 | `--no-throughput`             | Skip phase 2. Phase-1 latency screen only.                                               |
+| `--max-age HOURS`             | Drop mirrors more than this far behind the newest mirror (default 36). `0` keeps all.    |
 | `--http3`                     | Probe over QUIC the mirrors that advertise HTTP/3. Needs the `http3` extra.              |
 | `--json`                      | Emit JSON to stdout. Suppresses the table + progress.                                    |
 | `--csv`                       | Emit CSV to stdout. Suppresses the table + progress.                                     |
@@ -164,8 +165,9 @@ directly.
 > The image renders a colored, properly-widened table even without `-it`.
 > If you want **rich to use your host terminal's exact width** (and react to
 > resizes), add `-it`; otherwise the image falls back to a sensible 120-col
-> table. Set `NO_COLOR=1` if you'd rather have plain text — e.g. when
-> redirecting to a file.
+> table. Set `NO_COLOR=1` if you'd rather have plain text with no escape
+> codes — e.g. when redirecting to a file. Progress bars show only when
+> stderr is a terminal.
 
 ### Basic invocation
 
@@ -402,6 +404,9 @@ The mirror lists for those distributions cover x86_64 mirrors only.
      multi-record format). The list does not say which sites serve HTTPS, so
      each site is tried as an `https://` and an `http://` candidate.
    - Fedora: `mirrors.fedoraproject.org/mirrorlist?repo=&arch=&country=`.
+   - The Ubuntu and Fedora lists name one scheme for each mirror, often
+     `http://` for a host that also serves HTTPS. Each `http://` entry is
+     also tried as `https://`.
    - Linux Mint: scraped from `linuxmint.com/mirrors.php`.
    - Arch: `archlinux.org/mirrors/status/json/` (JSON, filtered to active
      HTTP and HTTPS entries).
@@ -409,16 +414,22 @@ The mirror lists for those distributions cover x86_64 mirrors only.
    If the list cannot be fetched, the command prints the reason and exits
    with code 1.
 
-2. **Phase 1 — latency screen.** Concurrent ranged GET (`bytes=0-1023`) on the
-   distro's canonical small metadata file (`InRelease` for Ubuntu and Debian,
-   `Release` for Mint, `repomd.xml` for Fedora, `lastsync` for Arch). Three
-   samples per mirror, **median TTFB** used for ranking.
+2. **Phase 1 — latency screen.** Concurrent ranged GET (`bytes=0-1023`) on a
+   small metadata file that the distribution rewrites as packages change:
+   `InRelease` of the `-updates` suite for Ubuntu and Debian, `Release` for
+   Mint, `repomd.xml` of the updates repository for Fedora, `lastupdate` for
+   Arch. Three samples per mirror, **median TTFB** used for ranking.
 3. **Phase 2 — throughput test.** Streaming GET of a large file
    (`Contents-<arch>.gz` for apt, `images/install.img` for Fedora, `extra.db`
    for Arch) on the top N from phase 1, capped at 5 MiB and 30 seconds.
    `bytes_per_sec` is measured from first-byte to last-byte so steady-state
    bandwidth isn't polluted by connect/TLS overhead.
 
+   A mirror that answers a missing file with status 200 and an HTML error
+   page fails phase 1 with `soft_404`. A complete phase 2 body under 256 KiB
+   fails with `short_body`. If no mirror passes phase 1, a line on stderr
+   gives the most common failure; a wrong `--release` is the usual cause.
+
    Both phases follow redirects, so the timing includes the extra hop that a
    package manager also pays. Only a 2xx final response counts. An `https://`
    mirror that redirects to `http://` fails with `https_downgrade`. An
@@ -436,6 +447,23 @@ The mirror lists for those distributions cover x86_64 mirrors only.
      is kept.
 5. **Render.** A `rich.Table`, or JSON / CSV for scripting.
 
+### Freshness
+
+A fast mirror that stopped syncing serves old packages. The phase 1 file
+carries the time its content was generated, so each run also reads that
+time, with no extra request.
+
+- The `Age` column shows how far a mirror is behind the newest mirror in the
+  run: `0`, `23h`, `12d`, or `-` when the file gave no time. JSON and CSV
+  carry it as `sync_age_s`.
+- Age is relative to the other mirrors, not to your clock. If the
+  distribution published nothing for a week, every current mirror still
+  shows `0`.
+- Mirrors more than 36 hours behind are dropped before ranking, and a line on
+  stderr says how many. Change the limit with `--max-age HOURS`; `0` keeps
+  all mirrors.
+- Age does not change the score. A stale mirror is wrong, not slow.
+
 ### HTTP/3
 
 Phase 1 reads the `Alt-Svc` response header. The `HTTP/3` column shows `adv`

{linux_mirror_bench-0.2.0 → linux_mirror_bench-0.3.0}/README.md

@@ -117,6 +117,7 @@ Running `mirror-bench` with no subcommand is equivalent to `mirror-bench bench`.
 | `--tls13-only`                | Hard-exclude mirrors that don't negotiate TLS 1.3.                                       |
 | `--weights lat=…,thr=…,sec=…` | Override scoring weights (auto-normalized).                                              |
 | `--no-throughput`             | Skip phase 2. Phase-1 latency screen only.                                               |
+| `--max-age HOURS`             | Drop mirrors more than this far behind the newest mirror (default 36). `0` keeps all.    |
 | `--http3`                     | Probe over QUIC the mirrors that advertise HTTP/3. Needs the `http3` extra.              |
 | `--json`                      | Emit JSON to stdout. Suppresses the table + progress.                                    |
 | `--csv`                       | Emit CSV to stdout. Suppresses the table + progress.                                     |
@@ -137,8 +138,9 @@ directly.
 > The image renders a colored, properly-widened table even without `-it`.
 > If you want **rich to use your host terminal's exact width** (and react to
 > resizes), add `-it`; otherwise the image falls back to a sensible 120-col
-> table. Set `NO_COLOR=1` if you'd rather have plain text — e.g. when
-> redirecting to a file.
+> table. Set `NO_COLOR=1` if you'd rather have plain text with no escape
+> codes — e.g. when redirecting to a file. Progress bars show only when
+> stderr is a terminal.
 
 ### Basic invocation
 
@@ -375,6 +377,9 @@ The mirror lists for those distributions cover x86_64 mirrors only.
      multi-record format). The list does not say which sites serve HTTPS, so
      each site is tried as an `https://` and an `http://` candidate.
    - Fedora: `mirrors.fedoraproject.org/mirrorlist?repo=&arch=&country=`.
+   - The Ubuntu and Fedora lists name one scheme for each mirror, often
+     `http://` for a host that also serves HTTPS. Each `http://` entry is
+     also tried as `https://`.
    - Linux Mint: scraped from `linuxmint.com/mirrors.php`.
    - Arch: `archlinux.org/mirrors/status/json/` (JSON, filtered to active
      HTTP and HTTPS entries).
@@ -382,16 +387,22 @@ The mirror lists for those distributions cover x86_64 mirrors only.
    If the list cannot be fetched, the command prints the reason and exits
    with code 1.
 
-2. **Phase 1 — latency screen.** Concurrent ranged GET (`bytes=0-1023`) on the
-   distro's canonical small metadata file (`InRelease` for Ubuntu and Debian,
-   `Release` for Mint, `repomd.xml` for Fedora, `lastsync` for Arch). Three
-   samples per mirror, **median TTFB** used for ranking.
+2. **Phase 1 — latency screen.** Concurrent ranged GET (`bytes=0-1023`) on a
+   small metadata file that the distribution rewrites as packages change:
+   `InRelease` of the `-updates` suite for Ubuntu and Debian, `Release` for
+   Mint, `repomd.xml` of the updates repository for Fedora, `lastupdate` for
+   Arch. Three samples per mirror, **median TTFB** used for ranking.
 3. **Phase 2 — throughput test.** Streaming GET of a large file
    (`Contents-<arch>.gz` for apt, `images/install.img` for Fedora, `extra.db`
    for Arch) on the top N from phase 1, capped at 5 MiB and 30 seconds.
    `bytes_per_sec` is measured from first-byte to last-byte so steady-state
    bandwidth isn't polluted by connect/TLS overhead.
 
+   A mirror that answers a missing file with status 200 and an HTML error
+   page fails phase 1 with `soft_404`. A complete phase 2 body under 256 KiB
+   fails with `short_body`. If no mirror passes phase 1, a line on stderr
+   gives the most common failure; a wrong `--release` is the usual cause.
+
    Both phases follow redirects, so the timing includes the extra hop that a
    package manager also pays. Only a 2xx final response counts. An `https://`
    mirror that redirects to `http://` fails with `https_downgrade`. An
@@ -409,6 +420,23 @@ The mirror lists for those distributions cover x86_64 mirrors only.
      is kept.
 5. **Render.** A `rich.Table`, or JSON / CSV for scripting.
 
+### Freshness
+
+A fast mirror that stopped syncing serves old packages. The phase 1 file
+carries the time its content was generated, so each run also reads that
+time, with no extra request.
+
+- The `Age` column shows how far a mirror is behind the newest mirror in the
+  run: `0`, `23h`, `12d`, or `-` when the file gave no time. JSON and CSV
+  carry it as `sync_age_s`.
+- Age is relative to the other mirrors, not to your clock. If the
+  distribution published nothing for a week, every current mirror still
+  shows `0`.
+- Mirrors more than 36 hours behind are dropped before ranking, and a line on
+  stderr says how many. Change the limit with `--max-age HOURS`; `0` keeps
+  all mirrors.
+- Age does not change the score. A stale mirror is wrong, not slow.
+
 ### HTTP/3
 
 Phase 1 reads the `Alt-Svc` response header. The `HTTP/3` column shows `adv`

{linux_mirror_bench-0.2.0 → linux_mirror_bench-0.3.0}/pyproject.toml

@@ -7,7 +7,7 @@ build-backend = "hatchling.build"
 # neighbors hit PyPI's similarity check. The installed command + import module
 # are still `mirror-bench` / `mirror_bench` via [project.scripts] below.
 name = "linux-mirror-bench"
-version = "0.2.0"
+version = "0.3.0"
 description = "Cross-distribution Linux package mirror benchmarking tool"
 readme = "README.md"
 license = { text = "AGPL-3.0-or-later" }

{linux_mirror_bench-0.2.0 → linux_mirror_bench-0.3.0}/src/mirror_bench/benchmark/probe.py

@@ -18,6 +18,8 @@ if TYPE_CHECKING:
 # Phase 1 samples a small ranged window (1 KiB) to keep the latency probe cheap.
 _LATENCY_RANGE = "bytes=0-1023"
 _LATENCY_MAX_BYTES = 1024
+# identity: the body is read raw and parsed for the mirror's sync time.
+_LATENCY_HEADERS = {"Range": _LATENCY_RANGE, "Accept-Encoding": "identity"}
 _LATENCY_DEADLINE_S = 10.0
 
 # Phase 2 downloads this many bytes max — caps the cost for slow/misbehaving mirrors.
@@ -97,6 +99,11 @@ async def throughput_test(
     return out
 
 
+def _is_html(body: bytes | bytearray) -> bool:
+    # None of the probed files is HTML; Fedora's repomd.xml starts `<?xml`.
+    return bytes(body[:64]).lstrip().lower().startswith((b"<!doctype html", b"<html"))
+
+
 def _downgraded(mirror: Mirror, resp: httpx.Response) -> bool:
     """Did an HTTPS mirror send any hop of this exchange over plain HTTP?"""
     return mirror.is_https and any(
@@ -152,26 +159,29 @@ async def _latency_probe(
     host: HostInfo,
     samples: int,
 ) -> ProbeResult:
-    url = mirror.url + discoverer.probe_path(host)
+    url = discoverer.probe_url(mirror, host)
     ttfbs: list[float] = []
     totals: list[float] = []
     last_resp: httpx.Response | None = None
     last_bytes = 0
+    last_body = b""
     last_error = "no_samples"
 
     for _ in range(samples):
         started = time.perf_counter()
         nbytes = 0
+        body = bytearray()
         try:
             async with (
                 asyncio.timeout(_LATENCY_DEADLINE_S),
-                client.stream("GET", url, headers={"Range": _LATENCY_RANGE}) as resp,
+                client.stream("GET", url, headers=_LATENCY_HEADERS) as resp,
             ):
                 ttfb_s = time.perf_counter() - started
                 if _downgraded(mirror, resp):
                     return _failed(mirror, "https_downgrade", resp)
                 async for chunk in resp.aiter_raw():
                     nbytes += len(chunk)
+                    body += chunk
                     # Only a server that ignored Range sends more than the
                     # window. A ranged reply is read to its end, which keeps
                     # the connection reusable for the next sample.
@@ -184,8 +194,12 @@ async def _latency_probe(
         else:
             ttfbs.append(ttfb_s * 1000.0)
             totals.append((time.perf_counter() - started) * 1000.0)
+            if resp.is_success and _is_html(body):
+                # Some mirrors answer a missing file with 200 and an error page.
+                return _failed(mirror, "soft_404", resp)
             last_resp = resp
             last_bytes = nbytes
+            last_body = bytes(body[:_LATENCY_MAX_BYTES])
             continue
         # A mirror that fails before any sample succeeds is not worth the
         # remaining connect timeouts.
@@ -212,6 +226,7 @@ async def _latency_probe(
         error=None,
         final_url=str(last_resp.url),
         h3_port=_h3_port(last_resp),
+        sync_time=discoverer.sync_time(mirror, last_body),
     )
 
 
@@ -255,13 +270,17 @@ async def _throughput_probe(
 
     if resp is None or first_byte_at is None:
         return _failed(mirror, "empty_body", resp)
+    if bytes_read < _THROUGHPUT_MIN_BYTES:
+        # Each throughput file is megabytes long. A complete body this short
+        # is an error page or a stub, and its rate would mean nothing.
+        return _failed(mirror, "short_body", resp, bytes_read=bytes_read)
 
     total_s = last_byte_at - started
     # Measure throughput from TTFB to the last byte, skipping connect/TLS overhead,
     # so the number reflects steady-state bandwidth rather than connection setup cost.
     transfer_s = last_byte_at - first_byte_at
-    if bytes_read >= _THROUGHPUT_MIN_BYTES and transfer_s > 0:
-        bps = bytes_read / transfer_s
+    if transfer_s > 0:
+        bps: float | None = bytes_read / transfer_s
     elif total_s > 0:
         bps = bytes_read / total_s
     else:

{linux_mirror_bench-0.2.0 → linux_mirror_bench-0.3.0}/src/mirror_bench/benchmark/scorer.py

@@ -1,5 +1,6 @@
 """Weighted composite mirror ranking."""
 
+import time
 from dataclasses import replace
 from typing import TYPE_CHECKING
 
@@ -11,11 +12,16 @@ if TYPE_CHECKING:
 # Maximum security points we can award. See _security_score for the breakdown.
 _MAX_SECURITY = 2.5
 
+# A sync time this far past the local clock is a broken mirror, not the
+# newest one.
+_FUTURE_SLACK_S = 24 * 3600.0
+
 
 def score_results(
     latency_results: Iterable[ProbeResult],
     throughput_results: dict[str, ProbeResult] | None,
     weights: Weights,
+    ages: dict[str, float] | None = None,
 ) -> list[Score]:
     """Rank mirrors. `throughput_results=None` means phase 2 did not run."""
     throughput_map = throughput_results or {}
@@ -73,6 +79,7 @@ def score_results(
                 composite=composite,
                 probe=r,
                 throughput_probe=tput_probe,
+                sync_age_s=(ages or {}).get(r.mirror.url),
             )
         )
 
@@ -123,6 +130,38 @@ def _security_score(probe: ProbeResult, tput_probe: ProbeResult | None) -> float
     return score
 
 
+def sync_ages(
+    results: Iterable[ProbeResult], *, now: float | None = None
+) -> dict[str, float]:
+    """Seconds each mirror is behind the newest one, keyed by mirror URL.
+
+    Mirrors are compared with each other, not with the local clock: the
+    distribution may not have published anything for days, and that is not
+    the mirrors' fault. Mirrors with no known sync time are left out.
+    """
+    limit = (time.time() if now is None else now) + _FUTURE_SLACK_S
+    times = {
+        r.mirror.url: r.sync_time for r in results if r.ok and r.sync_time is not None
+    }
+    plausible = [t for t in times.values() if t <= limit]
+    if not plausible:
+        return {}
+    newest = max(plausible)
+    return {url: max(0.0, newest - t) for url, t in times.items()}
+
+
+def drop_stale(
+    results: Iterable[ProbeResult], ages: dict[str, float], max_age_s: float
+) -> list[ProbeResult]:
+    """Drop mirrors more than `max_age_s` behind; 0 keeps all.
+
+    A mirror with no known age is kept: nothing shows it is stale.
+    """
+    if max_age_s <= 0:
+        return list(results)
+    return [r for r in results if ages.get(r.mirror.url, 0.0) <= max_age_s]
+
+
 def prefer_https(results: Iterable[ProbeResult]) -> list[ProbeResult]:
     """Drop an http:// result when the same site answered over https://.
 

{linux_mirror_bench-0.2.0 → linux_mirror_bench-0.3.0}/src/mirror_bench/cli.py

@@ -2,6 +2,7 @@
 
 import asyncio
 import math
+from collections import Counter
 from dataclasses import replace
 
 import typer
@@ -13,6 +14,7 @@ from rich.progress import (
     TextColumn,
     TimeElapsedColumn,
 )
+from typer._completion_shared import get_completion_script
 from typer.core import TyperGroup
 
 from mirror_bench import display
@@ -94,6 +96,14 @@ Http3Opt = typer.Option(
         'Needs the optional extra: "linux-mirror-bench[http3]".'
     ),
 )
+MaxAgeOpt = typer.Option(
+    36.0,
+    "--max-age",
+    min=0.0,
+    help=(
+        "Drop mirrors more than this many hours behind the newest mirror. 0 keeps all."
+    ),
+)
 SkipThroughputOpt = typer.Option(
     False,
     "--no-throughput",
@@ -104,7 +114,13 @@ SkipThroughputOpt = typer.Option(
 def _parse_countries(raw: str | None) -> tuple[str, ...]:
     if not raw:
         return ()
-    return tuple(c.strip().upper() for c in raw.split(",") if c.strip())
+    codes = tuple(c.strip().upper() for c in raw.split(",") if c.strip())
+    for code in codes:
+        # Codes go into mirror-list URLs, so only real alpha-2 shapes pass.
+        if len(code) != 2 or not code.isascii() or not code.isalpha():
+            msg = f"bad --country code {code!r}; expected two letters, e.g. US"
+            raise typer.BadParameter(msg)
+    return codes
 
 
 def _parse_weights(raw: str | None) -> Weights:
@@ -179,9 +195,13 @@ def bench_cmd(
     csv_out: bool = CsvOpt,
     no_throughput: bool = SkipThroughputOpt,
     http3: bool = Http3Opt,
+    max_age: float = MaxAgeOpt,
 ) -> None:
     """Discover mirrors, run a two-phase benchmark, print the ranked table."""
     _check_output_flags(json_out, csv_out)
+    if not math.isfinite(max_age):
+        msg = "--max-age must be a finite number of hours"
+        raise typer.BadParameter(msg)
     if http3:
         # Before discovery, so the flag fails the same way whether or not
         # any mirror turns out to advertise HTTP/3.
@@ -206,6 +226,7 @@ def bench_cmd(
         weights=_parse_weights(weights),
         skip_throughput=no_throughput,
         http3=http3,
+        max_age_hours=max_age,
     )
     if no_throughput and cfg.weights.latency + cfg.weights.security <= 0:
         msg = "--no-throughput needs a latency or security weight above zero"
@@ -238,8 +259,6 @@ def completion_cmd(
     depend on shell auto-detection and works reliably in sandboxes, CI, and
     containers. Pipe the output to your shell's completion directory.
     """
-    from click.shell_completion import shell_complete
-
     shell = shell.lower()
     if shell not in {"bash", "zsh", "fish"}:
         typer.secho(
@@ -249,12 +268,16 @@ def completion_cmd(
         )
         raise typer.Exit(code=2)
 
-    prog = "mirror-bench"
-    env_var = "_MIRROR_BENCH_COMPLETE"
-    cli = typer.main.get_command(app)
-    rc = shell_complete(cli, {}, prog, env_var, f"{shell}_source")
-    if rc != 0:
-        raise typer.Exit(code=rc)
+    # typer's own generator: recent typer releases bundle their CLI layer and
+    # no longer install click, so click cannot be imported here.
+    typer.echo(
+        get_completion_script(
+            prog_name="mirror-bench",
+            complete_var="_MIRROR_BENCH_COMPLETE",
+            # The keyword names the target shell; no command is executed.
+            shell=shell,  # nosec B604
+        )
+    )
 
 
 @app.command("list")
@@ -310,6 +333,7 @@ async def _run_bench(host: HostInfo, cfg: BenchConfig, *, quiet: bool) -> list[S
             TimeElapsedColumn(),
         )
         console = display.stderr_console()
+        quiet = quiet or not console.is_terminal
 
         if quiet:
             latency_results = await probe.latency_screen(
@@ -337,7 +361,33 @@ async def _run_bench(host: HostInfo, cfg: BenchConfig, *, quiet: bool) -> list[S
                     on_progress=on_prog,
                 )
 
-        latency_results = scorer.prefer_https(latency_results)
+        if not any(r.ok for r in latency_results):
+            # An empty table with no reason is the usual result of a wrong
+            # --release, so say what the mirrors answered.
+            reasons = Counter(r.error or f"http_{r.status}" for r in latency_results)
+            typer.secho(
+                f"no mirror passed the latency probe ({len(latency_results)} tried; "
+                f"most common failure: {reasons.most_common(1)[0][0]}). "
+                "Check the release name.",
+                fg=typer.colors.YELLOW,
+                err=True,
+            )
+
+        # Stale endpoints go first: prefer_https would otherwise keep a stale
+        # https:// endpoint and discard its current http:// twin.
+        ages = scorer.sync_ages(latency_results)
+        fresh = scorer.drop_stale(latency_results, ages, cfg.max_age_hours * 3600)
+        fresh_hosts = {r.mirror.host for r in fresh}
+        stale_hosts = {r.mirror.host for r in latency_results} - fresh_hosts
+        if stale_hosts:
+            typer.secho(
+                f"dropped {len(stale_hosts)} "
+                f"{'mirror' if len(stale_hosts) == 1 else 'mirrors'} more than "
+                f"{cfg.max_age_hours:g} h behind the newest",
+                fg=typer.colors.YELLOW,
+                err=True,
+            )
+        latency_results = scorer.prefer_https(fresh)
         if cfg.tls13_only:
             latency_results = scorer.filter_for_policy(
                 latency_results, https_only=False, tls13_only=True
@@ -377,7 +427,7 @@ async def _run_bench(host: HostInfo, cfg: BenchConfig, *, quiet: bool) -> list[S
                         on_progress=on_prog2,
                     )
 
-        scores = scorer.score_results(top_for_phase2, throughput_map, cfg.weights)
+        scores = scorer.score_results(top_for_phase2, throughput_map, cfg.weights, ages)
         if cfg.http3:
             scores = await _with_h3_probes(scores, min(8, cfg.concurrency))
         return scores

{linux_mirror_bench-0.2.0 → linux_mirror_bench-0.3.0}/src/mirror_bench/discovery/arch.py

@@ -1,5 +1,6 @@
 """Arch Linux mirror discovery via archlinux.org/mirrors/status/json/."""
 
+import re
 from typing import TYPE_CHECKING
 
 from mirror_bench.discovery.base import (
@@ -7,6 +8,7 @@ from mirror_bench.discovery.base import (
     MirrorDiscoverer,
     ensure_trailing_slash,
     fetch,
+    first_int,
     host_of,
 )
 from mirror_bench.models import HostInfo, Mirror
@@ -15,6 +17,7 @@ if TYPE_CHECKING:
     import httpx
 
 _STATUS_JSON = "https://archlinux.org/mirrors/status/json/"
+_UNIX_TIME_RE = re.compile(rb"\A\s*(\d+)\s*\Z")
 
 
 class ArchDiscoverer(MirrorDiscoverer):
@@ -59,7 +62,12 @@ class ArchDiscoverer(MirrorDiscoverer):
         return mirrors
 
     def probe_path(self, host: HostInfo) -> str:
-        return "lastsync"
+        # `lastupdate` changes when the content does; `lastsync` changes on
+        # each sync run, so two mirrors with equal content would differ.
+        return "lastupdate"
+
+    def sync_time(self, mirror: Mirror, body: bytes) -> float | None:
+        return first_int(body, _UNIX_TIME_RE)
 
     def throughput_path(self, host: HostInfo) -> str:
         # Official Arch mirrors carry x86_64 only, whatever the probing host is.