System Resource & I/O Optimization Questions

#!/bin/bash
svc=$1
threshold_pct=75
sustain_checks=3
check_interval=10
drain_timeout=60
# get cgroup path (systemd-run style)
while true; do
  pid=$(systemctl show -p MainPID --value $svc)
  [ "$pid" -gt 0 ] || { sleep $check_interval; continue; }
  mem_bytes=$(cat /sys/fs/cgroup/system.slice/${svc}.service/memory.current 2>/dev/null || \
              awk '/VmRSS/ {print $2*1024}' /proc/$pid/status)
  mem_max=$(cat /sys/fs/cgroup/system.slice/${svc}.service/memory.max 2>/dev/null || echo 0)
  if [ "$mem_max" -gt 0 ]; then
    pct=$((100 * mem_bytes / mem_max))
  else
    # derive sensible limit or use physical memory
    pct=$((100 * mem_bytes / (1024*1024*1024) )) # assume 1GB baseline or fetch from config
  fi
  if [ "$pct" -ge "$threshold_pct" ]; then
    hits=$((hits+1)); else hits=0; fi
  if [ "$hits" -ge "$sustain_checks" ]; then
    # Stage 1: mark draining (LB / Consul / pacemaker API)
    curl -s -X POST http://lb-api/drain/$(hostname) || true
    # wait for inflight to drop or timeout
    start=$(date +%s)
    while in_flight_requests_above_zero && [ $(( $(date +%s)-start )) -lt $drain_timeout ]; do sleep 2; done
    # Stage 2: graceful restart
    systemctl kill --signal=SIGTERM $svc
    # Wait for stop; systemd will restart due to Restart=on-failure
    # If service doesn't stop, escalate after TimeoutStopSec
    # Stage 3: clear drain after service active and healthy
    sleep 5
    curl -s -X POST http://lb-api/ready/$(hostname) || true
    hits=0
  fi
  sleep $check_interval
done

#!/usr/bin/env python3
import sys

def parse_vmstat(stream):
    headers = None
    si_idx = so_idx = time_idx = None
    max_si = {'val': None, 'time': None}
    max_so = {'val': None, 'time': None}

    for line in stream:
        line = line.strip()
        if not line:
            continue
        parts = line.split()
        # detect header line (contains 'si' and 'so')
        if 'si' in parts and 'so' in parts:
            headers = parts
            si_idx = parts.index('si')
            so_idx = parts.index('so')
            # optional timestamp header name (common: 'time' or 'timestamp' or first column not numeric)
            if parts[0].lower() in ('time', 'timestamp'):
                time_idx = 0
            continue

        # skip kernel repeated header lines or summary lines
        if headers is None:
            continue

        # if there's an extra timestamp column before standard columns
        # detect if first token looks like hh:mm:ss
        if time_idx is None:
            if ':' in parts[0]:
                time_idx = 0
                # shift indices right by 1 if header didn't include time
                # assume header had no time column -> si_idx/so_idx correspond to parts without time
                # so adjust using length difference
                # if parts have same length as headers+1, shift
                if len(parts) == len(headers) + 1:
                    si_idx += 1
                    so_idx += 1

        try:
            si = int(parts[si_idx])
            so = int(parts[so_idx])
        except (IndexError, ValueError):
            continue

        ts = parts[time_idx] if time_idx is not None else None

        if max_si['val'] is None or si > max_si['val']:
            max_si.update(val=si, time=ts)
        if max_so['val'] is None or so > max_so['val']:
            max_so.update(val=so, time=ts)

    return max_si, max_so

if __name__ == '__main__':
    max_si, max_so = parse_vmstat(sys.stdin)
    print(f"max si: {max_si['val']} at {max_si['time']}")
    print(f"max so: {max_so['val']} at {max_so['time']}")