Financial API Failover Strategy: Timeouts, 429s, Staleness and Fallbacks

A pricing page that crashes when the data provider hiccups is bad. A pricing page that silently shows a wrong number is worse. A solid financial API failover strategy covers both: availability problems and data-quality problems. This guide lays out the patterns with EZRate’s error model as the example.

Part of the Iranian Rial exchange rate API guide.

Error model

{ "error": { "code": "rate_limited", "message": "…" }, "meta": { "reset_at": "2026-10-20T00:00:00Z" } }

Response headers X-RateLimit-Remaining-Day, X-RateLimit-Remaining-Month and Retry-After let you act before you hit a wall. Full details: JSON reference.

Retry with backoff

async function getJson(url: string, key: string, tries = 3): Promise<any> {
  for (let i = 0; i < tries; i++) {
    try {
      const r = await fetch(url, { headers: { "X-API-Key": key }, signal: AbortSignal.timeout(10_000) });
      if (r.ok) return r.json();
      if (r.status === 429) {
        const wait = Number(r.headers.get("retry-after") ?? 60);
        throw Object.assign(new Error("rate_limited"), { retryAfter: wait, final: true });
      }
      if (r.status < 500) throw Object.assign(new Error(`client ${r.status}`), { final: true });
    } catch (e: any) {
      if (e.final || i === tries - 1) throw e;
    }
    await new Promise(res => setTimeout(res, 500 * 2 ** i + Math.random() * 250)); // jitter
  }
}

Cache with a last-good copy

• Fresh key with TTL = freshness_sec.
• Last-good key with a long TTL (e.g. 24 h).
• Background refresh from a scheduler, so user requests never wait on the upstream.
• Stale flag in your own payload so the UI can show “last updated 14 minutes ago”.

Implementations: Laravel, Python/Django, Node.js.

Data-quality guards

EZRate already applies median filtering and a jump guard upstream (see outlier detection); your own guards protect against integration bugs and edge cases.

Secondary sources

For flows where downtime costs money (checkout, trading), keep a second source. Use it only when the primary is stale beyond a threshold, log every switch, and never average two sources with different methodologies without understanding them.

Monitoring

• Alert on cache age, not just HTTP errors.
• Track 429s: they usually mean caching broke.
• Track the remaining quota from headers or /usage.
• Synthetic check: request one symbol every minute from outside your network.

Designing for degraded modes

Decide in advance how each part of your product behaves when data is old: product pages keep showing prices with a notice; checkout may refuse to finalise a price older than a threshold; automated repricing pauses. Writing these rules down turns an outage from an emergency into a documented, expected state.

Practise these modes too: occasionally block the upstream host in staging and confirm that pages, checkout and jobs behave exactly as documented.

Related live rates: US dollar rate, 18K gold price, Emami coin price.