# /etc/systemd/system/passkeeper.service
#
# Install / update:
#   sudo cp scripts/passkeeper.service /etc/systemd/system/passkeeper.service
#   sudo systemctl daemon-reload
#   sudo systemctl enable passkeeper
#   sudo systemctl restart passkeeper
#   journalctl -xeu passkeeper.service
#
# Phase 5 additions vs original:
#   - Restart=on-failure: systemd restarts Gunicorn if the master exits non-zero
#   - Gunicorn --timeout: workers that don't respond within 25 s are replaced
#   - Gunicorn --graceful-timeout: allows in-flight requests to finish on reload
#   - PrivateTmp, NoNewPrivileges, ProtectSystem: basic systemd sandboxing
#   - StartLimitIntervalSec / StartLimitBurst: caps restart storm

[Unit]
Description=PassKeeper Gunicorn daemon
After=network.target mysql.service
Wants=mysql.service

# Restart policy: cap to 5 restarts in 60 s to prevent restart storms
StartLimitIntervalSec=60
StartLimitBurst=5

[Service]
User=www-data
Group=www-data
WorkingDirectory=/home/spuser/PassKeeper
EnvironmentFile=/home/spuser/PassKeeper/.env

ExecStart=/home/spuser/.venv/bin/gunicorn \
          --workers 4 \
          --bind 127.0.0.1:5000 \
          --timeout 25 \
          --graceful-timeout 20 \
          --keep-alive 5 \
          --access-logfile /home/spuser/logs/access.log \
          --error-logfile  /home/spuser/logs/error.log \
          --log-level warning \
          wsgi:app

# Reload (zero-downtime): send USR2 to Gunicorn master
ExecReload=/bin/kill -s USR2 $MAINPID

# NO WatchdogSec here — deliberately.
#
# WatchdogSec requires the service to send WATCHDOG=1 keepalives over the sd_notify
# socket. Gunicorn only does that when systemd exports NOTIFY_SOCKET, which happens
# only under Type=notify (+ NotifyAccess=main). This unit is Type=simple (the
# default), so no keepalive was ever sent, systemd treated the service as hung, and
# SIGKILLed it every ~30 s. Restart=on-failure then brought it back after RestartSec,
# producing a repeating window of 502s from Nginx.
#
# Hung *workers* are already handled by Gunicorn's own --timeout above; a crashed
# *master* is already handled by Restart=on-failure below. The watchdog added no
# coverage, only outages.
#
# To re-enable it properly (optional), all three lines are required:
#   Type=notify
#   NotifyAccess=main
#   WatchdogSec=30s

Restart=on-failure
RestartSec=5s

# Systemd sandboxing
PrivateTmp=true
NoNewPrivileges=true
ProtectSystem=strict
ReadWritePaths=/home/spuser/PassKeeper /home/spuser/logs /home/spuser/backups

[Install]
WantedBy=multi-user.target