Building a Bounded-Latency Quote Pipeline That Never Lies

Your users don’t care why a quote failed—they care that the number didn’t flash 0.00. Here’s a pragmatic, production-ready pipeline that returns the best price you can get within a strict time budget, exposes provenance, and never silently lies.

Audience: engineers shipping finance dashboards, brokers, or portfolio tools.

Deliverables: architecture, TypeScript code (fetch with timeout + backoff), and a reproducible test harness you can run locally.

The failure modes we must beat

1. Single-provider coupling. One upstream hiccup → you show 0.00 or stale data.
2. Unbounded waits. A slow API call blocks the UI thread or burns your edge budget.
3. Silent staleness. You cache yesterday’s price and don’t tell the user.
4. Inconsistent math. Different providers round/adjust differently, breaking P/L.

We’ll design for bounded latency, graceful degradation, and auditability.

Architecture: time-boxed fan-out + LKG + provenance

Imagine the following diagram (textual description):

[ Client ]  --(fetch / 1.5s budget)--> [ Edge Worker ]
                                        |-> Provider A (Yahoo)   [score 0]
                                        |-> Provider B (NBBO)    [score +1]
                                        |-> Provider C (CoinGecko/OHLC) [score +2]
                                        |
                             <best valid within budget>
                                        |
                           [ LKG cache (30s TTL) ]
                                        |
                           returns {price, ts, source, stale}

• The edge queries providers in parallel within a strict budget (e.g., 1500 ms).
• We prefer the first valid response using a simple health score (recent failures raise the score; ties go to lowest score).
• If nothing lands in time, we return LKG (last-known-good) marked stale: true.
• The client renders a Provenance chip: Yahoo • 12s • Live or Cache • 45s • Stale. No black boxes.

Core TypeScript: bounded fetch with backoff + jitter

We’ll start with a few helpers.

// utils/timeout.ts
export async function fetchWithTimeout(
  url: string,
  ms: number,
  init: RequestInit = {}
): Promise<Response> {
  const ctrl = new AbortController();
  const timer = setTimeout(() => ctrl.abort("time_budget_exceeded"), ms);
  try {
    const res = await fetch(url, { ...init, signal: ctrl.signal });
    return res;
  } finally {
    clearTimeout(timer);
  }
}

// utils/backoff.ts
export function backoffDelay(attempt: number, base = 150, cap = 1000) {
  // exponential backoff with full jitter
  const exp = Math.min(cap, base * 2 ** attempt);
  return Math.floor(Math.random() * exp);
}

Provider adapters (pure functions)

Each adapter returns a normalized quote or throws.

// providers/types.ts
export type Quote = {
  symbol: string;
  price: number;
  ts: number;       // epoch millis
  source: string;   // "yahoo", "nbbo", etc.
};

export type Provider = (s: string) => Promise<Quote>;
// providers/yahoo.ts
import { fetchWithTimeout } from "../utils/timeout";
import type { Quote, Provider } from "./types";

export const yahoo: Provider = async (symbol) => {
  const r = await fetchWithTimeout(`https://api.example/yahoo?symbol=${symbol}`, 700);
  if (!r.ok) throw new Error(`yahoo:${r.status}`);
  const j = await r.json();
  const price = Number(j?.price);
  if (!Number.isFinite(price)) throw new Error("yahoo:invalid");
  return { symbol, price, ts: Date.now(), source: "yahoo" };
};

Duplicate for other sources (nbbo, coingecko, etc.) with their own timeouts.

Circuit breaker + scoring

Providers accumulate penalties on failure (simple, sticky health heuristic).

// engine/quotes.ts
import type { Provider, Quote } from "../providers/types";
import { backoffDelay } from "../utils/backoff";

type Health = Record<string, number>; // lower is healthier

export type BudgetedOptions = {
  budgetMs?: number;          // total time budget
  attempts?: number;          // retries per provider
  lkg?: Quote | null;         // last-known-good
  health?: Health;            // provider health map
};

export async function getQuoteWithBudget(
  symbol: string,
  providers: Record<string, Provider>,
  opts: BudgetedOptions = {}
): Promise<Quote & { stale: boolean }> {
  const budgetMs = opts.budgetMs ?? 1500;
  const attempts = opts.attempts ?? 1;
  const health = opts.health ?? {};

  const start = Date.now();
  let winner: Quote | null = null;

  // run providers in order of health (lowest first)
  const entries = Object.entries(providers).sort(
    ([a], [b]) => (health[a] ?? 0) - (health[b] ?? 0)
  );

  // Launch all providers in parallel but respect total budget
  await Promise.any(
    entries.map(async ([name, p]) => {
      for (let i = 0; i < attempts; i++) {
        const elapsed = Date.now() - start;
        const remaining = budgetMs - elapsed;
        if (remaining <= 0) throw new Error("time_budget_exceeded");

        try {
          const q = await p(symbol); // each p uses its own per-call timeout
          if (!Number.isFinite(q.price)) throw new Error("invalid");
          if (!winner) winner = q;    // first valid wins
          return;
        } catch {
          health[name] = (health[name] ?? 0) + 1; // penalize on fail
          await new Promise((r) => setTimeout(r, backoffDelay(i)));
        }
      }
      throw new Error(`${name}_exhausted`);
    })
  ).catch(() => null);

  // If nothing landed, fall back to LKG
  const now = Date.now();
  if (!winner && opts.lkg) {
    return { ...opts.lkg, source: `${opts.lkg.source}+lkg`, stale: true };
  }

  if (!winner) {
    // honest sentinel; caller should render skeleton/“unavailable”
    return { symbol, price: NaN, ts: now, source: "unavailable", stale: true };
  }

  // Mark staleness based on age (e.g., 30s)
  const stale = now - winner.ts > 30_000;
  return { ...winner, stale };
}

Why this shape works: bounded total latency, first valid result preference, provider health nudges future calls, and a clean stale flag keeps the UI honest.

Client UX: provenance chip (text-described diagram)

Imagine a small pill next to the price:

[ Yahoo • 12s • Live ]    // green background
[ Cache • 47s • Stale ]    // amber background

When a user clicks it, show a popover with source, requestId, and the time budget used—handy for bug reports and audits.

// ui/ProvenanceBadge.tsx
export function ProvenanceBadge({ source, ts, stale }: { source: string; ts: number; stale: boolean }) {
  const ageSec = Math.max(0, Math.round((Date.now() - ts) / 1000));
  const label = `${source} • ${ageSec}s • ${stale ? "Stale" : "Live"}`;
  const cls = `px-2 py-1 rounded-full text-xs ${stale ? "bg-yellow-100 text-yellow-800" : "bg-green-100 text-green-800"}`;
  return <span className={cls} aria-live="polite">{label}</span>;
}

Reproducible test harness

We’ll simulate providers with deterministic latency/failure. You can paste this into a Node/TS project (no network required).

// harness/sim.ts
import type { Provider, Quote } from "../providers/types";
import { getQuoteWithBudget } from "../engine/quotes";

function fakeProvider(name: string, opts: { latency: number; fail?: boolean; price?: number }): Provider {
  return async (symbol: string) => {
    await new Promise((r) => setTimeout(r, opts.latency));
    if (opts.fail) throw new Error(`${name}_fail`);
    const price = opts.price ?? 100 + Math.random();
    return { symbol, price, ts: Date.now(), source: name } as Quote;
  };
}

async function main() {
  const providers = {
    yahoo: fakeProvider("yahoo", { latency: 180, price: 189.97 }),
    nbbo:  fakeProvider("nbbo",  { latency: 420, price: 189.95 }),
    alt:   fakeProvider("alt",   { latency: 1100, fail: true }),
  };
  const out = await getQuoteWithBudget("AAPL", providers, { budgetMs: 600, attempts: 1, lkg: null, health: {} });
  console.log(out);
}

main().catch(console.error);

Expected behavior: with a 600 ms budget, yahoo wins (180 ms). If you increase yahoo latency above 600, nbbo becomes the winner. If everything fails, the function returns { price: NaN, source: "unavailable", stale: true }—which your UI should render as a skeleton or an explicit “unavailable” state.

Unit tests (Vitest)

// harness/sim.test.ts
import { describe, it, expect } from "vitest";
import { getQuoteWithBudget } from "../engine/quotes";
import type { Provider } from "../providers/types";

function p(name: string, latency: number, ok = true, price = 101): Provider {
  return async (s) => {
    await new Promise((r) => setTimeout(r, latency));
    if (!ok) throw new Error(name);
    return { symbol: s, price, ts: Date.now(), source: name };
  };
}

describe("getQuoteWithBudget", () => {
  it("returns first valid within budget", async () => {
    const out = await getQuoteWithBudget("AAPL", { slow: p("slow", 500), fast: p("fast", 80, true, 190.12) }, { budgetMs: 300 });
    expect(out.source).toBe("fast");
    expect(out.stale).toBe(false);
    expect(Number.isFinite(out.price)).toBe(true);
  });

  it("falls back to LKG when none arrive in time", async () => {
    const lkg = { symbol: "AAPL", price: 188.5, ts: Date.now() - 40_000, source: "cache" };
    const out = await getQuoteWithBudget("AAPL", { slow: p("slow", 800) }, { budgetMs: 200, lkg });
    expect(out.source).toContain("lkg");
    expect(out.stale).toBe(true);
    expect(out.price).toBe(188.5);
  });

  it("returns sentinel when all fail & no LKG", async () => {
    const out = await getQuoteWithBudget("AAPL", { a: p("a", 50, false), b: p("b", 60, false) }, { budgetMs: 150 });
    expect(out.source).toBe("unavailable");
    expect(isNaN(out.price)).toBe(true);
  });
});

Run it

npm i -D vitest typescript tsx
npx vitest run

Production notes (learned the hard way)

• Time budgets must be end-to-end (edge → providers → marshaling). Don’t just set per-call timeouts; total elapsed matters.
• Never coerce NaN to 0.00. Return a sentinel and render it honestly.
• Record provenance (source, ts, requestId) for every price; it turns user bug reports into reproducible tickets.
• Cache with intent. LKG cache should be small (e.g., 30–60 s) and keyed by symbol + venue where applicable.
• Score providers. A sticky health score quickly pushes you away from flaky sources without a full breaker library.
• UI skeletons + chips > spinners. Show skeleton rows during refresh and clear “Stale” chips when using LKG.

What this buys you

• Predictable UX: Users get a number (or a truthful “stale/unavailable”) within your SLA.
• Operational safety: Slow or failing providers don’t stall the entire pipeline.
• Auditability: Every number is explainable—source + age + freshness.

Next steps you can ship today

1. Drop getQuoteWithBudget into your edge worker.
2. Add the Provenance chip.
3. Wire a 30–60 s LKG cache and emit “stale” when using it.
4. Expand the test harness with your real providers behind a feature flag.

If you want a code drop-in adapted to your stack (Vercel Edge, Cloudflare Workers, or Node), ping me your environment and symbols—you’ll get a tailored snippet.

Want the companion “OpenBrokerCSV v0.1” spec + validator to make imports boringly reliable? I’ve got a minimal template and test fixtures—DM me your broker and I’ll spin up a mapping PR.