code/scripts/watchdog.sh

#!/usr/bin/env bash
# Detects the two failure modes the retry loop inside run_all.sh cannot see.
#
#   (setsid nohup bash scripts/watchdog.sh > work/logs/watchdog.log 2>&1 < /dev/null &)
#
# WHAT THE RETRY LOOP ALREADY HANDLES: a step that exits non-zero. It retries three times.
#
# WHAT IT CANNOT HANDLE, AND WHY THIS EXISTS:
#   1. A DEAD TRACK. When retr_F4 failed three times, its track function returned and the
#      track simply stopped - silently, with the other track still running and the status
#      file showing nothing new. Nothing noticed for seven minutes. This restarts a track
#      whose steps are unfinished but whose process is gone. Restarting is safe because
#      every step skips work whose output already exists.
#   2. A HUNG STEP. A process that is alive but has written nothing for a long time is
#      worse than a crash: the retry loop waits on it forever. CUDA deadlocks and NFS
#      stalls both look like this. After STALL_S with no log growth the step is killed, and
#      run_all's own retry then restarts it from its checkpoint.
#
# It never touches anyone else's processes: every kill is matched on our own script paths
# under this directory, checked against the command line, never on a bare name.
set -uo pipefail
cd "$(dirname "$0")/.."
ROOT=$(pwd)
W="${BERX_WORK:-$ROOT/work}"
L="$W/logs"
STALL_S="${STALL_S:-1500}"      # 25 min with no log growth = hung
POLL_S="${POLL_S:-60}"
mkdir -p "$L"

log() { echo "[$(date '+%H:%M:%S')] $*"; }

track_alive() {   # $1 = 1 or 2
  # Per-track pidfile written by run_all.sh. Counting run_all.sh processes cannot distinguish
  # the tracks - with either one running the count is identical - so a dead track hid behind
  # a live one. That is the bug this watchdog exists to catch, so it must not repeat it.
  local pf="$L/track$1.pid"
  [ -f "$pf" ] || return 1
  local pid; pid=$(cat "$pf" 2>/dev/null)
  [ -n "$pid" ] && [ -d "/proc/$pid" ]
}

track_done() {    # track 1 is done once the submission exists; track 2 once shards exist
  case "$1" in
    1) [ -f "$W/output_v2/matching_results.tsv" ] || grep -q '"pass": false' \
         "$W/rescore/gate.json" 2>/dev/null ;;
    2) [ -f "$W/shards/test_manifest.json" ] ;;
  esac
}

restart_track() {
  local t="$1"
  log "TRACK $t appears dead with work outstanding - restarting (steps with outputs skip)"
  ( cd "$ROOT" && TRACKS="$t" setsid nohup bash scripts/run_all.sh \
      > "$L/run_track${t}_restart_$(date +%H%M%S).log" 2>&1 < /dev/null & )
}

log "watchdog up. stall=${STALL_S}s poll=${POLL_S}s work=$W"
while true; do
  sleep "$POLL_S"

  # ---- 1. hung stage, judged by CPU TIME, not by log mtime.
  #
  # A log that has not grown does NOT mean the job is stuck: Python block-buffers stdout to
  # a file, so a stage started without -u writes nothing for its entire run. t1_fit.log sat
  # at 0 bytes while the fit burned 198% CPU. Killing on log mtime alone would have
  # destroyed a healthy 40-minute job.
  #
  # A process that is genuinely wedged - CUDA deadlock, NFS stall - stops accumulating CPU
  # time. That is the signal. The log is only used as a secondary confirmation.
  for pid in $(pgrep -u "$USER" -f "scripts/[0-9][0-9]_.*\.py" 2>/dev/null); do
    cmd=$(tr '\0' ' ' < "/proc/$pid/cmdline" 2>/dev/null)
    # Only ever our own scripts, under our own root. Other users run their own pipelines on
    # this box and one of them is a scripts/run_p4_verifier.py.
    case "$cmd" in
      *"$ROOT/scripts/"*|*" scripts/0"*|*" scripts/1"*) ;;
      *) continue ;;
    esac
    # utime+stime in clock ticks, fields 14 and 15 of /proc/<pid>/stat
    cpu=$(awk '{print $14+$15}' "/proc/$pid/stat" 2>/dev/null) || continue
    seen="$L/.cpu_$pid"
    if [ -f "$seen" ]; then
      read -r oldcpu oldts < "$seen"
      if [ "$cpu" -gt "$oldcpu" ]; then
        echo "$cpu $(date +%s)" > "$seen"                 # progressing
      else
        idle=$(( $(date +%s) - oldts ))
        if [ "$idle" -ge "$STALL_S" ]; then
          log "STALL pid $pid used no CPU for ${idle}s: $cmd"
          log "      killing it so run_all retries the step from its checkpoint"
          kill -TERM "$pid" 2>/dev/null; sleep 10; kill -KILL "$pid" 2>/dev/null
          rm -f "$seen"
        fi
      fi
    else
      echo "$cpu $(date +%s)" > "$seen"
    fi
  done
  # forget bookkeeping for processes that have exited
  for f in "$L"/.cpu_*; do
    [ -f "$f" ] || continue
    p=${f##*/.cpu_}
    [ -d "/proc/$p" ] || rm -f "$f"
  done

  # ---- 2. dead track with work outstanding
  for t in 1 2; do
    track_done "$t" && continue
    if ! track_alive "$t"; then
      # confirm over two polls so a momentary gap between steps is not mistaken for death
      sleep 15
      track_alive "$t" && continue
      restart_track "$t"
      sleep 30
    fi
  done

  # ---- 3. announce the milestones, so the log alone tells the story
  if [ -f "$W/rescore/gate.json" ] && [ ! -f "$L/.announced_gate" ]; then
    log "GATE A: $(tr -d '\n ' < "$W/rescore/gate.json")"; touch "$L/.announced_gate"
  fi
  if [ -f "$W/shards/test_manifest.json" ] && [ ! -f "$L/.announced_shards" ]; then
    log "CE SHARDS READY - Colab lanes can start"; touch "$L/.announced_shards"
  fi
  if [ -f "$W/output_v2/matching_results.tsv" ] && [ ! -f "$L/.announced_sub" ]; then
    log "SUBMISSION WRITTEN: $W/output_v2/matching_results.tsv"; touch "$L/.announced_sub"
  fi
done