Deterministic Routing Strategies

Ce sujet n’est pas encore traduit en français ; la version anglaise est affichée. Lire l’original anglais

Deterministic routing strategies

OmniRoute’s self-hosted gateway entry (/v1/chat/completions, RIC-738) routes to a single provider by default. When you configure a strategy: block, the route decision becomes a deterministic, explainable policy — the OmniRoute differentiator vs NotDiamond/Martian “predict” black-box routing.

Every decision is:

• Deterministic — the same config + same runtime state yields the same provider.
• Config-expressed — one rule per line; no “一体化智能体”, no learned model.
• Explainable — every response carries x-omniroute-route-decision, a one-line answer to “why this model?” (and “why NOT that one”).

This is M2 of the RIC-697 differentiation (D3 可审计路由). There is no predictive / ML dependency — the strategies are pure rules over observable signals (consecutive failures, declared cost, measured latency).

Enable

Add a strategy: block to the same YAML document as providers:, or point OMNIROUTE_SELF_HOSTED_STRATEGY (inline YAML) / OMNIROUTE_SELF_HOSTED_STRATEGY_FILE (a file) at a standalone strategy: document. An explicit env strategy merges over the inline block per-key.

providers.yaml

providers:
  - id: cheap
    kind: openai
    baseUrl: http://127.0.0.1:11434/v1
    model: llama3
    costPer1MInput: 0.2 # USD per 1M input tokens — used by cost-priority
  - id: fast
    kind: openai
    baseUrl: http://127.0.0.1:8080/v1
    model: gpt-4o-mini
    costPer1MInput: 0.6
  - id: premium
    kind: anthropic
    baseUrl: http://127.0.0.1:8081
    model: claude-sonnet
    costPer1MInput: 3.0

strategy:
  blacklist: [] # provider ids never used
  whitelist: [cheap, fast, premium] # when non-empty, ONLY these are used
  costPriority: true # cheapest eligible candidate first
  latencyAware:
    enabled: true # fastest recent-average candidate first
  cooldown:
    consecutiveFailures: 2 # breaker trips after this many in a row
    cooldownMs: 30000 # …and the provider stays excluded this long
  fallbackChain: [cheap, fast, premium] # explicit fallback order

The five strategies

Strategy	Config	Effect
Blacklist	blacklist: [id, ...]	Listed providers are never candidates.
Whitelist	whitelist: [id, ...]	Non-empty → only listed providers are candidates.
Cooldown/breaker	cooldown: {consecutiveFailures, cooldownMs}	After N consecutive failures a provider is excluded for the window. A success resets the counter.
Cost-priority	costPriority: true	Eligible candidates sorted by costPer1MInput ascending (declared, never fabricated).
Latency-aware	latencyAware.enabled: true	Eligible candidates sorted by recent average request latency ascending. Unsampled providers sort last.
Fallback chain	fallbackChain: [id, ...]	Explicit primary→backup order. Wins over cost/latency ordering.

Filters (blacklist / whitelist / cooldown) run first and remove candidates. Then the ordering stage sorts the survivors: explicit fallbackChain wins; otherwise costPriority then latencyAware apply in that order.

Pinned providers

The x-omniroute-provider header and the provider/model model-prefix are explicit pins, not preferences. If a pinned provider is excluded by a hard filter (blacklist / whitelist / active cooldown), the request fails with 400 and the pin reason — the gateway never silently re-routes a client who asked for a specific provider. If the pin survives the filters it is the first candidate, and the remaining candidates serve as its fallback chain.

Fallback on failure

The ordered candidate list is the fallback chain. When the first provider fails (network unreachable, connection refused, DNS/TLS, or any non-2xx), the gateway walks the next candidate, and so on. Each failure is recorded into the cooldown breaker and each attempt is latency-sampled — so a broken primary also cools down for subsequent requests. When every candidate fails, the last normalized error (OpenAI shape) is returned. When no candidate is eligible at all, a 503 with the full explainable decision is returned.

Explainability

Every response routed through the strategy engine carries:

x-omniroute-route-decision: cheap(cost-priority: cheap #1) -> fast(cost-priority: fast #2); premium excluded: blacklist: premium forbidden

The value is the ordered candidate list plus every exclusion reason — the audit trail for “why this model, why not that one”.

Failure contract (unchanged from RIC-738)

• Upstream non-2xx body is normalized through parseUpstreamError + buildErrorBody into the standard OpenAI error shape (Hard Rule #12).
• Network-level failures return a normalized 502; after the whole chain is exhausted, the last normalized error is surfaced.
• Credential / session headers echoed upstream are stripped from every response.
• Provider config with a malformed strategy: block returns 500 — a misconfigured policy never silently becomes a no-op.

Tests

• tests/unit/routing-strategies.test.ts — 25 unit tests over the pure strategy engine: parse, blacklist/whitelist, cooldown breaker, cost-priority, latency-aware, fallback chain, combined pipeline, explainability.
• tests/unit/self-hosted-entry.test.ts — fault injection over real HTTP: primary down → fallback succeeds; all-down → last error; pinned-blacklisted → 400; cooldown excludes a failing provider on the next request; no-eligible → 503 with explainable decision.

---

Code source d’OmniRoute (a58000c7685f)