The Linux install now mirrors the Windows virtual-account hardening: the
unit runs with DynamicUser=yes (transient per-service UID, no login),
ProtectSystem=strict with only StateDirectory writable (the install dir,
so self-update can rewrite the binary), NoNewPrivileges, empty
capability sets, restricted address families and a @system-service
syscall filter. Install copies the binary to /var/lib/gpu-turnstile and
the config to /etc/gpu-turnstile.env; Remove cleans up the unit and
binary but keeps the config.
--install-service now registers the service under
NT SERVICE\gpu-turnstile (low-privilege, per-service, no password) and
grants it modify access to the install dir (for self-updates) and the
LOG_FILE dir, plus read access to an external config file. Grants run
after CreateService because the virtual account's SID does not exist
before registration; a failed grant rolls back the registration.
The service package now has a Linux implementation alongside the Windows
one: systemd unit install/remove (/etc/systemd/system), readiness
notification (READY=1 via go-systemd) sent only after the listeners are
bound, a 30s watchdog, and STOPPING on shutdown. Listeners are pre-bound
so port conflicts fail fast and the readiness signal is truthful. The
notify calls are no-ops without NOTIFY_SOCKET (containers, shells) and
on non-Linux builds. --install-service/--remove-service work on both
platforms; the 'service install|remove' subcommand remains as an alias.
OLLAMA_URL and COMFY_URL no longer have defaults; each consumer
(listener, client, startup probe, lock participation) is enabled by
setting its URL and disabled by leaving it empty. At least one must be
set. Ollama-only mode is a pure pass-through; ComfyUI-only mode skips
the unload and warm-reload steps. /metrics is now served on both
listeners. This is the extension pattern for future consumers such as
local game detection.
LLM_BUSY_MODE=reject answers blocked LLM requests immediately with
LLM_BUSY_STATUS (default 503, 429 works) and Retry-After, so routers
like LiteLLM can cool down and retry instead of holding a hung
connection. The default wait mode now also sends Retry-After when
LLM_WAIT_TIMEOUT expires. Document the service account (LocalSystem
default, NT SERVICE virtual-account hardening) and the Program
Files / ProgramData install layout.
- internal/config: .env-style config file (gpu-turnstile.env next to the
exe, -config flag or GPU_TURNSTILE_CONFIG); process env overrides file.
- internal/service: Windows service via golang.org/x/sys/windows/svc —
graceful SCM stop, 'service install/remove' commands, restart-on-failure
recovery (also applies staged updates). First external dependency,
Windows-only; Linux/Docker build unaffected (go.mod stays at 1.23).
- internal/update: polls the Gitea releases API, verifies the Ed25519
signature of the downloaded binary against an embedded public key
(openssl-signed by CI), swaps it in next to the running exe, and once
the GPU lock is idle exits with code 3 so service recovery restarts
onto the new version. Dev builds and empty pubkey never update.
- CI: tag builds additionally produce gpu-turnstile.exe + .sig + .sha256
attached to a Gitea release.
- LOG_FILE env var so the service has somewhere to log.
Backoff now covers all dial-phase errors (refused, timeout, DNS), TLS
handshake failures, and 5xx responses with replayable bodies; streamed
POSTs are never replayed to avoid duplicate work. First retry of an
episode logs at WARN, subsequent attempts at INFO.
LOGLEVEL (LOG_LEVEL kept as alias) now defaults to warn: startup logs
version plus every setting; INFO adds one line per incoming request and
per response with status/duration, ANSI-colored in text mode (bypasses
slog's escaping so colors render in docker compose logs); NO_COLOR or
LOG_FORMAT=json disables colors.
A refused dial (service down/restarting) is retried with a wait that
doubles from BACKOFF_INITIAL (1s) up to BACKOFF_MAX (60s) until the
upstream answers or the client disconnects. Handles the Windows WSA
errno (10061) as well as POSIX ECONNREFUSED. compose.yaml.example now
uses host.docker.internal like the working local deployment.
Tests still run on every push. The docker job is gated on ref_type=tag,
validates the tag is a strict vX.Y.Z semver, and publishes :<tag> plus
:latest. No branch images (sha-* / main latest) anymore.
The proxy now listens on the standard service ports (Ollama :11434,
ComfyUI :8188) and the actual services move one port up (:11435, :8189).
Metric prefix is now gpu_turnstile_.
mram