Router Access Log Fields Reference

Kthena emits one access-log entry after each /v1/* inference request. Requests routed through Gateway API on other paths currently do not emit an access-log entry. Each recorded entry contains the HTTP result, routing selection, token counts, timing checkpoints, and any router error captured during processing.

Formats and defaults

Access logs support two formats:

• text is the current Helm default and preserves existing installations.
• json is recommended for log aggregation and field-based queries, but must be enabled explicitly.

The text representation is a single line with this shape:

[timestamp] "METHOD /path PROTOCOL" status_code [error=type:message] [model_name=name] [model_route=route] [model_server=server] [selected_pod=pod] [request_id=id] [gateway=gateway] [http_route=route] [inference_pool=pool] [tokens=input/output] timings=totalms(request+upstream+response)

Fields in square brackets are omitted when no value is available. JSON uses the names in the tables below.

HTTP fields

Field	Type	Presence	Description	Example
timestamp	RFC3339 timestamp	Always	Time the router received the request	2026-01-09T14:35:22.147Z
method	String	Always	HTTP method	POST
path	String	Always	URL path only; the query string is not included	/v1/chat/completions
protocol	String	Always	HTTP protocol reported by Go	HTTP/1.1
status_code	Integer	Always	Final HTTP response status	200

Error fields

The error object is omitted from JSON when the router did not record an error. In text logs it appears immediately after the status code.

Field	Type	Description	Example
error.type	String	Stable error category assigned by the router	pod_discovery
error.message	String	Diagnostic message for this failure	no available pods for model server: default/llama2-server

Current router error categories are:

Error type	Failure area
request_parsing	Request body or required model field could not be parsed
prompt_parsing	Prompt extraction failed
input_rate_limit	Input-token limit was exceeded
output_rate_limit	Output-token limit was exceeded
rate_limit	General request/token rate limit was exceeded
model_server_matching	No matching ModelServer could be selected
pod_discovery	Backend pod lookup failed or returned no available pods
inference_pool_selection	A matched HTTPRoute had no eligible InferencePool backend
inference_pool_discovery	The selected InferencePool could not be found
port_discovery	An InferencePool had no usable target port
route_not_found	Neither ModelRoute nor HTTPRoute routing matched
scheduling	Queue admission or backend scheduling failed
proxy	Forwarding to the selected backend failed

An HTTP status is not uniquely determined by error.type; consult status_code for the actual result.

Model and routing fields

Field	Type	Presence	Description	Example
model_name	String	Always in JSON; may be empty when parsing fails	Model requested by the client	llama2-7b
model_route	String	When selected	Namespaced ModelRoute	default/llama2-route-v1
model_server	String	When selected	Namespaced ModelServer	default/llama2-server
selected_pod	String	When selected	Backend pod name	llama2-deployment-5f7b8c9d-xk2p4
request_id	String	Normally present	Incoming x-request-id, or an ID generated by the access-log middleware	550e8400-e29b-41d4-a716-446655440000
gateway	String	Gateway API requests	Namespaced Gateway	default/inference-gateway
http_route	String	HTTPRoute requests	Namespaced HTTPRoute	default/inference-route
inference_pool	String	Inference Extension requests	Namespaced InferencePool	default/llama2-pool

Token fields

Field	Type	Presence	Description	Example
input_tokens	Integer	JSON omits zero	Prompt-token count calculated by the router	150
output_tokens	Integer	JSON omits zero	Completion-token count parsed from the backend response when available	75

Text output includes tokens=input/output when either count is nonzero.

Timing fields

All timing values are integer milliseconds.

Field	Type	Description	Example
duration_total	Integer	End-to-end time from receipt until access-log finalization	2350
duration_request_processing	Integer	Initial router parsing and tokenization phase	45
duration_upstream_processing	Integer	Marked backend-proxy phase	2180
duration_response_processing	Integer	Marked response-processing phase	5

There is no duration_queue field and no access-log field uses an _ms suffix. duration_total is authoritative. The phase fields are independently rounded checkpoints and may not sum to the total, including when rate limiting, routing, scheduling, or queue time falls between marked phases.

Examples

Successful JSON entry

{
  "timestamp": "2026-01-09T14:35:22.147Z",
  "method": "POST",
  "path": "/v1/chat/completions",
  "protocol": "HTTP/1.1",
  "status_code": 200,
  "model_name": "llama2-7b",
  "model_route": "default/llama2-route-v1",
  "model_server": "default/llama2-server",
  "selected_pod": "llama2-deployment-5f7b8c9d-xk2p4",
  "request_id": "550e8400-e29b-41d4-a716-446655440000",
  "input_tokens": 150,
  "output_tokens": 75,
  "duration_total": 2350,
  "duration_request_processing": 45,
  "duration_upstream_processing": 2180,
  "duration_response_processing": 5
}

Failed JSON entry

{
  "timestamp": "2026-01-09T14:35:28.456Z",
  "method": "POST",
  "path": "/v1/chat/completions",
  "protocol": "HTTP/1.1",
  "status_code": 503,
  "error": {
    "type": "pod_discovery",
    "message": "no available pods for model server: default/llama2-server"
  },
  "model_name": "llama2-7b",
  "request_id": "660e8400-e29b-41d4-a716-446655440001",
  "input_tokens": 200,
  "duration_total": 51,
  "duration_request_processing": 4,
  "duration_upstream_processing": 47,
  "duration_response_processing": 0
}

Text entry

[2026-01-09T14:35:22.147Z] "POST /v1/chat/completions HTTP/1.1" 200 model_name=llama2-7b model_route=default/llama2-route-v1 model_server=default/llama2-server selected_pod=llama2-deployment-5f7b8c9d-xk2p4 request_id=550e8400-e29b-41d4-a716-446655440000 tokens=150/75 timings=2350ms(45+2180+5)

Helm configuration

The root chart uses the following values. The defaults shown here preserve the current text-format behavior:

networking:
  kthenaRouter:
    accessLog:
      enabled: true
      format: text
      output: stdout

Set format: json explicitly when a structured log pipeline requires JSON.

For the standalone networking subchart, omit the networking prefix:

kthenaRouter:
  accessLog:
    enabled: true
    format: json
    output: stdout

Environment variables

The deployment maps Helm values to these router environment variables:

Variable	Description	Helm default	Accepted values
ACCESS_LOG_ENABLED	Enable access logging	true	Values accepted by Go's boolean parser
ACCESS_LOG_FORMAT	Output format	text	text, json
ACCESS_LOG_OUTPUT	Destination	stdout	stdout, stderr, or a file path

Prefer stdout for Kubernetes workloads. A file path writes inside the router container and requires a mounted volume plus a separate collection strategy.