# Canonical example unit for BaseRunnerMulti.py - see docs/BaseRunnerMulti.md's
# "Deployment" section for the copy/daemon-reload/enable/start sequence.
# Adjust paths if this host's checkout doesn't live at /usr/local/basestation3.
[Unit]
Description=Consolidated multi-site glider account runner
After=network.target baserunnerprivexec.service
# Wants=, not Requires=: still pulls baserunnerprivexec.service in and
# orders this unit's start after it (see After= above), but - unlike
# Requires= - does NOT stop this unit whenever baserunnerprivexec.service
# is stopped or restarted (e.g. to pick up a sites.yaml/cgroup edit, or a
# unit-file change - see "Installing the units" below). This unit's own
# dispatch-failure requeue (Dispatcher._dispatch_one_queued) already
# absorbs the helper being briefly unreachable during that restart, so
# there's nothing to gain from also taking this unit down - Requires=
# previously did exactly that, silently, since Requires= propagates a
# stop to units that declare it, not only a start.
Wants=baserunnerprivexec.service

[Service]
User=baserunner
Group=baserunner
# baserunner has no home directory (--no-create-home), so matplotlib's
# default $HOME/.config/matplotlib cache dir isn't writable; it falls
# back to a throwaway /tmp dir with a startup warning if left unset.
CacheDirectory=baserunner
Environment=MPLCONFIGDIR=/var/cache/baserunner
ExecStart=/opt/basestation/bin/python /usr/local/basestation3/BaseRunnerMulti.py \
    --sites_config /usr/local/basestation3/etc/sites.yaml \
    --priv_exec_socket /run/baserunner/priv_exec.sock \
    --cgroup_root /sys/fs/cgroup/baserunner.slice \
    --base_log /var/log/baserunner/baserunnermulti.log
#    --job_completions_log /var/log/baserunner/job_completions.jsonl
# Read-only: BaseRunnerMulti.py only reads memory.current/memory.peak
# stat files under this path for per-job memory tracking, it never
# writes to cgroupfs. Must match baserunnerprivexec.service's own
# --cgroup_root exactly, or memory reporting silently reads nothing
# (fail-open, not fatal) - see "Per-site memory limits" in
# docs/BaseRunnerMulti.md.
# --job_completions_log (commented out above - opt in by uncommenting
# and pointing local/collect_cgroup_stats.py's own --job_completions_log
# at the same path) is a per-job record that never misses a short-lived
# job, unlike the periodic cgroup sampler - see "Per-site memory limits"
# in docs/BaseRunnerMulti.md for why that gap exists.
# Creates /var/log/baserunner/ owned baserunner:baserunner (mode 0750) on
# every start, recreating it if it's ever missing - no manual mkdir/chown
# of the log directory needed. Requires systemd >= 235.
LogsDirectory=baserunner
WatchdogSec=30
Restart=always
Type=notify

[Install]
WantedBy=multi-user.target