From d58fdb41980eba43d3b1ab14e6b1455b938aaf65 Mon Sep 17 00:00:00 2001
From: sora <2075279110@qq.com>
Date: Thu, 17 Sep 2026 05:39:58 +0000
Subject: [PATCH] Pull chain: 5-minute idle watchdog per mirror

A dead mirror emitted zero bytes and the fixed 3600s attempt timeout
let it stall the whole chain for an hour (third occurrence tonight:
15m49s frozen at 0 bytes). select() with a 300s idle timeout now kills
the attempt and moves to the next mirror; a trickling source (1 line/s
Downloading updates) resets the timer and stays alive.

Co-Authored-By: Claude <noreply@anthropic.com>
---
 evalharness/sandbox/prefetch.py | 24 ++++++++++++++++++++----
 1 file changed, 20 insertions(+), 4 deletions(-)

diff --git a/evalharness/sandbox/prefetch.py b/evalharness/sandbox/prefetch.py
index e59d24e..c3aa8e7 100644
--- a/evalharness/sandbox/prefetch.py
+++ b/evalharness/sandbox/prefetch.py
@@ -115,6 +115,15 @@ def _pull_one(image: str, verbose: bool = True) -> None:
         proc = subprocess.Popen(['docker', 'pull', ref],
                                 stdout=subprocess.PIPE, stderr=subprocess.STDOUT,
                                 bufsize=0)
+        # idle watchdog: a dead mirror emits NOTHING (0 bytes/s for hours);
+        # the old fixed 3600s timeout let one dead source stall the whole
+        # chain. 5 minutes of silence = kill and try the next mirror.
+        import os as _os
+        import selectors as _sel
+
+        _selr = _sel.DefaultSelector()
+        _selr.register(proc.stdout, _sel.EVENT_READ)
+        _fd = proc.stdout.fileno()
         last_line = ''
         last_show = 0.0
         layer_prog = {}   # layer id -> (done_bytes, total_bytes): the raw
@@ -134,20 +143,27 @@ def _pull_one(image: str, verbose: bool = True) -> None:
         # \n -- readline() buffered them until a rare \n arrived, so the
         # Downloading snapshots (and the ⏳ summary fed by them) never
         # surfaced. Split on BOTH terminators, chunked reads.
-        import io as _io
-        stream = _io.TextIOWrapper(proc.stdout, errors='replace')
         buf = ''
         _lines_iter = iter(lambda: None, 1)  # placeholder, replaced below
 
+        _IDLE_S = 300   # no output for 5 minutes -> dead mirror
+
         def _iter_lines():
             nonlocal buf
             while True:
-                chunk = stream.read(512)
+                if not _selr.select(timeout=_IDLE_S):
+                    proc.kill()
+                    if verbose:
+                        print(f'· no data for {_IDLE_S // 60}min -- dead '
+                              'mirror, trying next source ...',
+                              file=_sys.stderr, flush=True)
+                    return
+                chunk = _os.read(_fd, 4096)
                 if not chunk:
                     if buf:
                         yield buf
                     return
-                buf += chunk
+                buf += chunk.decode('utf-8', 'replace')
                 parts = buf.split('\r') if '\r' in buf else buf.split('\n')
                 if len(parts) > 1:
                     for p in parts[:-1]: