This page lists the metrics Arion emits when built with the metrics feature, grouped by family.
Metrics are exposed three ways: in Prometheus text format at the admin endpoint /stats/prometheus (requires the prometheus build feature and the admin listener), in Envoy-style text at admin /stats, and pushed over OTLP when envoy_bootstrap.stats_sinks and stats_flush_interval are configured. Prometheus names join the family and metric name with an underscore (http_downstream_rq_total); the /stats output prints family.<label values>.name: value. Histograms use millisecond buckets 5, 10, 50, 100, 500, 1000, 5000, 10000, +Inf (bytes buckets for mcp_tool_response_bytes) and export _bucket, _sum and _count series in Prometheus format. Any built-in metric can be renamed with the top-level metrics.rename map; see metrics.
| Name | Type | Labels | Description |
|---|
downstream_cx_total | counter | listener | Connections accepted. |
downstream_cx_destroy | counter | listener | Connections destroyed. |
downstream_cx_active | gauge | listener | Active connections. |
downstream_cx_length_ms | histogram | listener | Connection lifetime in milliseconds. |
no_filter_chain_match | counter | listener | Connections that matched no filter chain. |
Emitted by the HTTP connection manager.
| Name | Type | Labels | Description |
|---|
downstream_cx_total | counter | listener | HTTP connections. |
downstream_cx_ssl_total | counter | listener | HTTP connections over TLS. |
downstream_cx_ssl_active | gauge | listener | Active TLS HTTP connections. |
downstream_cx_destroy | counter | listener | HTTP connections destroyed. |
downstream_cx_active | gauge | listener | Active HTTP connections. |
downstream_cx_length_ms | histogram | listener | HTTP connection lifetime in milliseconds. |
downstream_cx_rx_bytes_total | counter | listener | Bytes received on HTTP connections. |
downstream_cx_tx_bytes_total | counter | listener | Bytes sent on HTTP connections. |
downstream_rq_total | counter | listener | Requests received. |
downstream_rq_active | gauge | listener | Requests in flight. |
downstream_rq_1xx … downstream_rq_5xx | counter | listener | Responses by status class (five metrics). |
downstream_cx_ws_upgrades_total | counter | listener | Successful WebSocket upgrades. |
downstream_cx_ws_upgrades_active | gauge | listener | Active upgraded connections. |
downstream_rq_ws_on_non_ws_route | counter | listener | Upgrade requests rejected by non-upgrade routes. |
| Name | Type | Labels | Description |
|---|
downstream_cx_total | counter | listener | TCP connections. |
downstream_cx_destroy | counter | listener | TCP connections destroyed. |
downstream_cx_active | gauge | listener | Active TCP connections. |
downstream_cx_length_ms | histogram | listener | TCP connection lifetime in milliseconds. |
cx_rx_bytes_received | counter | listener | Bytes received on proxied TCP connections. |
cx_tx_bytes_sent | counter | listener | Bytes sent on proxied TCP connections. |
| Name | Type | Labels | Description |
|---|
handshake | counter | listener | Successful TLS handshakes. |
| Name | Type | Labels | Description |
|---|
upstream_rq_total | counter | cluster | Upstream requests. |
upstream_rq_active | gauge | cluster | Upstream requests in flight. |
upstream_rq_time | histogram | cluster | Upstream request time in milliseconds. |
upstream_rq_timeout | counter | cluster | Requests that hit the route timeout. |
upstream_rq_per_try_timeout | counter | cluster | Attempts that hit the per-try timeout. |
upstream_rq_retry | counter | cluster | Request retries. |
upstream_cx_total | counter | cluster | Upstream connections. |
upstream_client_exec_total | counter | cluster | HTTP client executor spawns (connection dispatchers and HTTP/1 idle waiters). |
upstream_cx_active | gauge | cluster | Active upstream connections. |
upstream_cx_destroy | counter | cluster | Upstream connections destroyed. |
upstream_cx_idle_timeout | counter | cluster | Upstream connections closed by idle timeout. |
upstream_cx_connect_fail | counter | cluster | Failed upstream connection attempts. |
upstream_cx_connect_timeout | counter | cluster | Upstream connection attempts that timed out. |
upstream_cx_rx_bytes_total | counter | cluster | Bytes received from upstreams. |
upstream_cx_tx_bytes_total | counter | cluster | Bytes sent to upstreams. |
upstream_cx_overflow | counter | cluster | Connections rejected by the circuit breaker. |
upstream_rq_overflow | counter | cluster | Requests rejected by the circuit breaker. |
upstream_rq_retry_overflow | counter | cluster | Retries rejected by the circuit breaker. |
Unlabelled gauges, refreshed when the stats endpoints are read.
| Name | Type | Labels | Description |
|---|
uptime | gauge | — | Server uptime in seconds. |
concurrency | gauge | — | Number of worker threads. |
memory_heap_size | gauge | — | Reserved heap size in bytes (requires the jemalloc build feature; falls back to physical memory). |
memory_physical_size | gauge | — | Physical memory of the process in bytes. |
memory_allocated | gauge | — | Allocated heap bytes (requires jemalloc; falls back to physical memory). |
The result label is ok, rate_limited or not_applicable; filter is the filter's stat_prefix.
| Name | Type | Labels | Description |
|---|
connection_rate_limit | counter | filter, result | Listener local rate limit decisions. |
local_rate_limit | counter | filter, result | HTTP local rate limit decisions. |
user_rate_limit | counter | filter, user, result | User rate limit decisions. |
Per-caller metrics, partitioned by the identity configured in metrics.user_key (a request header or the TLS SNI). The label name is the configured attribute_name — user in the examples below.
| Name | Type | Labels | Description |
|---|
invocations | counter | user | API calls. |
throttles | counter | user | API calls that were rate limited. |
system_errors | counter | user | API calls that ended in a server-side error. |
user_errors | counter | user | API calls that ended in a client-side error. |
total_errors | counter | user | API calls that ended in any error. |
latency | histogram | user | API call latency in milliseconds. |
upstream_rq_time | histogram | user | Upstream request time in milliseconds. |
bytes_tx | counter | user, listener | Bytes sent to the caller. |
bytes_rx | counter | user, listener | Bytes received from the caller. |
inbound_streaming_bytes_processed | counter | user | Inbound WebSocket bytes. |
outbound_streaming_bytes_processed | counter | user | Outbound WebSocket bytes. |
connections | counter | user | Connections established. |
connections_active | gauge | user | Active connections. |
http_1xx_response … http_5xx_response | counter | user | Responses by status class (five metrics). |
http_404_response | counter | user | 404 responses. |
http_502_response | counter | user | 502 responses. |
http_504_response | counter | user | 504 responses. |
Emitted by the MCP gateway filter and exported through the OTLP pipeline; they are not listed by the admin /stats endpoints. Tools are identified by their qualified name, namespace.name. In code mode, every tool call a script makes is counted under that tool, the same as a direct tools/call.
| Name | Type | Labels | Description |
|---|
tool_rq_total | counter | mcp.tool.name, mcp.backend, mcp.upstream | Tool invocations dispatched to a backend. |
tool_rq_failures_total | counter | mcp.tool.name, mcp.backend, mcp.upstream, mcp.error.code | Tool invocations that produced a tool error. |
tool_rq_time | histogram | mcp.tool.name, mcp.backend, mcp.upstream | Tool invocation time in milliseconds. |
tool_response_bytes | histogram | mcp.tool.name, mcp.backend, mcp.upstream | Upstream tool response payload size in bytes. |
embedding_failures_total | counter | — | Failed embeddings requests for tool search, counted once per failed batch. |
User-defined counters, histograms and gauges declared under the top-level metrics.custom_metrics map extract values from request or response headers at six hook points and appear alongside the built-in families. Their names and labels are whatever the configuration declares; see metrics for the configuration.
