Configuration

Both the gateway and the relay load a YAML config file at startup. The path defaults to config.yaml in the working directory and can be overridden with the CONFIG_PATH environment variable.

Environment variable expansion

All string values in config.yaml support env var expansion before YAML parsing:

s3:
  endpoint: "${S3_ENDPOINT:-https://s3.fr-par.scw.cloud}"  # with default
  access_key: "${S3_ACCESS_KEY}"                            # required

Syntax:

Form	Behaviour
`${VAR}`	Replaced with the value of `VAR`. Empty string if unset
`${VAR:-default}`	Replaced with `VAR` if set and non-empty, otherwise `default`

Hot reload

POST /-/reload

Reloads config.yaml from disk without restarting the gateway. If the new config is invalid, the old config stays active and the endpoint returns 500.

Service registry, rate limits, LLM proxy settings, and guardrails are all reloaded. Active connections are not interrupted.

Gateway config reference

`server`

server:
  addr: ":8080"               # default
  read_timeout: 120s          # default
  write_timeout: 0            # no timeout by default (safe for long inference)
  idle_timeout: 120s          # default
  consumer_header: "X-Consumer-Username"  # set by APISIX after auth
  user_type_header: "X-User-Type"         # set by OPA (sa | user | ...)
  priority_header: "X-Priority"           # triggers LPUSH instead of RPUSH

consumer_header enables per-consumer job ownership, metrics, and job listing. user_type_header enables per-user-type rate limits. Both are optional.

`s3`

s3:
  endpoint: "https://s3.fr-par.scw.cloud"
  region: "fr-par"
  access_key: "${S3_ACCESS_KEY}"
  secret_key: "${S3_SECRET_KEY}"
  bucket: "my-gatewai-jobs"
  use_path_style: true   # required for most S3-compatible providers
  ca_bundle: ""          # path to PEM file for private CA
  ssl_insecure: false    # disable TLS verification (dev only)

`redis`

redis:
  addr: "redis:6379"
  password: ""
  db: 0
  pending_max_age: "2h"   # jobs stuck in pending longer than this are swept by GC

`lifecycle`

lifecycle:
  persists_result: false    # true = keep result in S3/Redis until TTL; false = delete on first read
  job_ttl:
    global: "72h"           # fallback TTL for all statuses
    completed: "24h"        # override per status
    pending: ""
    failed: "48h"
    cancelled: ""
  gc:
    enabled: true
    interval: "15m"         # how often the GC runs
    orphan_min_age: "5m"    # min age before an S3 object is considered orphaned

`encryption`

encryption:
  key: "${ENCRYPTION_KEY}"  # hex-encoded 32-byte AES-256 key; empty = disabled

See Encryption for details.

`metrics`

metrics_config:
  top_consumers: 20      # expose top-N consumers via gauge; 0 = disabled
  consumer_labels: false # per-consumer labels on counter (< 50 consumers only)

`rate_limits`

rate_limits:
  llm:
    sa:    { rate: 100, period: 1m }
    user:  { rate: 20,  period: 1m, token_rate: 100000, token_period: 1h }
    "*":   { rate: 10,  period: 1m }   # fallback

See Rate limiting for details.

Relay config reference

The relay config file (relay/config.yaml by default, overridden by CONFIG_PATH) covers connectivity and service registration only.

redis:
  addr: "redis:6379"
  password: ""

s3:
  endpoint: "https://s3.fr-par.scw.cloud"
  region: "fr-par"
  access_key: "${S3_ACCESS_KEY}"
  secret_key: "${S3_SECRET_KEY}"
  bucket: "my-gatewai-jobs"
  use_path_style: true

encryption:
  key: "${ENCRYPTION_KEY}"   # must match the gateway key

services:
  - type: transcription
    model: whisper-large-v3
    inference_url: "http://127.0.0.1:9000/v1/audio/transcriptions"

queue:
  pop_timeout: 5m   # BLMOVE timeout; relay exits cleanly on timeout

The encryption.key must be identical in the gateway and every relay sidecar.