> Raw Markdown twin (generated at build time from the source Markdown). Rendered page: https://docs.gatellm.io/en/console/settings · Doc index: https://docs.gatellm.io/en/llms.txt


# Settings: system and telemetry

Console → **Settings** (visible to admin and monitor) is organized into sub-tabs. This page covers the **System** and **Telemetry** sub-tabs; for web search see [Web search](/en/console/search.md), for SSO see [Configure SSO enterprise login](/en/howto/configure-sso.md), for billing see [Configure model pricing and view monthly costs](/en/howto/setup-pricing-and-billing.md), for audit see [Audit and security configuration](/en/reference/audit-and-security-config.md).

## System (#settings/system)

An operations diagnostics panel that reads like `kubectl describe` / `htop`: shows this gateway instance's **build / runtime identity**, **upstream health**, and **real-time memory diagnostics**.

- **Build / runtime identity**: version number, build information, instance identifier, etc., used during troubleshooting to confirm "which build and which instance is running".
- **License information**: the panel for the current license (`LICENSE_KEY`) — plan (Professional / Enterprise), billing cycle (monthly / yearly), total memory, instance count limit, per-instance memory, license number, and the online validation's **last online time** and **last result** (valid / invalid). Two alert states: **grace period** (the last online validation was judged invalid, with a countdown of remaining grace days; on expiry it downgrades to the free edition, prompting you to renew or change the key as soon as possible); **cluster over limit** (reconciliation detected instance count / total memory exceeding the licensed quota — alert only, the gateway keeps serving, prompting you to reduce instances or raise the quota). For the free edition and activation, see [Licensing and authorization status](/en/reference/licensing.md).
- **Upstream health**: per-instance readings of admission resources + per-upstream bad-link status, so on failure, opening the console immediately locates which upstream is problematic:
  - The **status conclusion** is shown to the right of the panel title (green dot = all upstreams healthy; yellow/red dot = number of affected/saturated upstreams); the bad-link budget and judgment threshold are displayed side by side as badges.
  - **Upstream table**: lists only **affected** upstreams (bad-link count > 0 or saturated), showing per row the upstream name, bad-link count (`current / budget`), and status (saturated = red, counting = yellow), and appears only during failures.
  - **Per-instance admission table**: one row per instance — in-flight connections, queue depth, processing passes, fd usage, **database connection pool** (open/limit and waiting columns, hover for per-pool breakdown), bad links. **Pool waiting > 0 is a precursor signal of the login cliff** — the data plane is squeezing DB connections. In multi-instance mode, the cluster total is appended as a small footnote below the table.
  - **Multi-instance (Redis) mode**: each instance broadcasts its above summary every 3 seconds (including its own connection-pool readings); the endpoint does cluster-level aggregation — bad-link counts are **summed** per upstream, saturation is **saturated if any instance is saturated**, and additive readings such as in-flight connections / queue depth / passes are summed; fd usage and connection pool are process-level readings where summing is meaningless, so they only appear in the per-instance table. On Redis read failure, it degrades to the local view of the instance hit. Memory (single-instance) mode semantics are unchanged.
  - The data source is the read-only endpoint `/console/api/upstream-health` (polled every 3 seconds); counts are **real-time in-flight values**, cleared as soon as a request ends, with no failure memory.
- **Memory diagnostics**: RSS / current usage / limit + a usage bar; decomposes the OOM trigger value into anonymous memory (RSS) + page cache + a small residual, exposing "the gap you can't see just by looking"; plus two diagnostic bars for memory-pressure level and heap-return health, turning abstract numbers into readable judgments.
- **Diagnostic download**: one-click export of a diagnostic bundle (`protoflux-diagnostic-<timestamp>.tar.gz`, which archive tools on any OS can double-click to unpack directly), for offline analysis of memory or admission issues. Files in the bundle:

  | File | Contents |
  |------|------|
  | `meta.txt` | Build / runtime identity + allocator configuration |
  | `heap.txt` | jemalloc heap profile |
  | `jemalloc-stats.json` | Allocator statistics |
  | `upstream-health.json` | Upstream bad-link / saturation snapshot + connection budget / fd usage (multi-instance mode: cluster aggregation + per-instance detail) |

  Each section is **collected independently** — if a section's collection fails, only a note is written into that file, without affecting the whole bundle.

For environment variables related to memory admission and spill, see [Environment variable reference → Memory admission and spill](/en/reference/configuration.md#memory-admission-spill); for upstream fault isolation and fd budget configuration, see [Listening and request limits](/en/reference/configuration.md#listen-and-request-limits) and [Upstream connection retry and routing affinity](/en/reference/configuration.md#upstream-retry-affinity) on the same page; for troubleshooting abnormal memory increases, see [Troubleshoot from symptoms](/en/usecases/troubleshooting.md).

## Telemetry (#settings/telemetry)

One switch: **Issue reporting** — whether to allow this gateway to send issue and crash reports to Sentry.

- **Takes effect immediately when off** (both frontend and backend): the backend stops reporting (events are dropped before sending, performance sampling set to 0); the frontend no longer initializes Sentry for new browser sessions, and already-open sessions are forcibly rejected by the tunnel endpoint.
- **Persistence**: the setting is written to the database (the `issue_reporting` key in the `config_settings` table) and survives restarts.
- **Multi-instance sync**: in Redis mode, synced to other nodes via a config reload broadcast.
- This switch only controls issue reporting and does not affect the gateway's own functionality; changes on this page are admin-only.

> The Sentry DSN baked into the image is determined by environment variables such as `SENTRY_DSN` / `SENTRY_FRONTEND_DSN` (see [Environment variable reference](/en/reference/configuration.md)); this switch is a one-click master switch on top of that, without needing to change environment variables or restart.

## FAQ

**Q: Can monitor change the telemetry switch?**
No. The Settings page is visible to monitor (read-only); modification operations (including the telemetry switch) are admin-only.

**Q: What do I lose by turning off issue reporting?**
Only stops sending issue/crash reports to Sentry; gateway operation, logs, and statistics are unaffected.

**Q: Upstream health shows an upstream is "saturated" — what happens to requests?**
If the model has other available upstreams, requests automatically route to a healthy upstream; if all candidate upstreams are saturated, requests immediately return 503 + Retry-After rather than queue indefinitely. Saturation judgment is purely real-time in-flight counting — the moment an upstream recovers it is immediately fully usable, with no manual un-banning.

**Q: Why is the "Diagnostic download" button grayed out?**
Two cases: non-admin roles cannot export; or the current instance has no jemalloc heap profiler (only Linux container builds have it), in which case `/debug/diagnostic-bundle` returns 501 anyway, and the button is disabled to avoid misleading.

**Next**: [Console login and roles](/en/console/login-and-roles.md) for page permissions; [Environment variable reference](/en/reference/configuration.md) for Sentry-related variables; [Audit and security configuration](/en/reference/audit-and-security-config.md) for audit.