Architecture

System architecture

Prism RPC sits between your application and every Robinhood Chain provider you would otherwise integrate one at a time. This is what happens inside.

Layers

Edge

TLS termination in every region you call from
API key validation and method allowlisting
Per-key token buckets and burst allowances
JSON-RPC envelope and batch parsing

Balancer core

Routing engine with four selectable strategies
Per-provider circuit breakers with half-open probing
Retry policy with exponential backoff and full jitter
In-flight deduplication and deterministic-result cache

Transport

Keep-alive HTTP pools per upstream
WebSocket manager with automatic reconnection
Subscription registry replayed after a reconnect
Response validation and block-height drift checks

Upstreams

Commercial providers: Alchemy, QuickNode, Ankr, Chainstack
Public Robinhood Chain nodes as overflow capacity
Optional private nodes you bring yourself
Independent health state per upstream and per region

Request flow

A cache miss on a read, routed to the fastest healthy upstream. Timings are representative of a warm pool.

1
Request received
Key validated, method allowlist checked
0.08 ms
2
Cache lookup
Deterministic results answered immediately
0.04 ms
3
Dedupe
Identical in-flight read joined instead of re-sent
0.02 ms
4
Provider selected
Healthy upstreams ranked by rolling p50
0.11 ms
5
Upstream call
JSON-RPC forwarded, response validated
38 ms
6
Metrics recorded
Latency, status and breaker state updated
0.03 ms

Failure isolation

Each upstream carries its own breaker. One provider degrading never becomes your outage.

Closed

Traffic flows. Prism counts failures in a rolling window.

5 failures in 30s →

Open

The provider is removed from the pool. Requests reroute instantly.

after 15s cooldown →

Half-open

A single probe request is allowed through to test recovery.

probe succeeds → Closed

Retries and breakers share one rule: a request is only replayed when replaying it is safe. Reads are retried across providers freely; anything that submits a transaction is sent exactly once, to one upstream, and its failure is returned to you verbatim rather than being silently duplicated.

policy defaults

{
  "retry": {
    "maxAttempts": 3,
    "baseDelayMs": 120,
    "maxDelayMs": 2000,
    "jitter": "full",
    "retryOn": ["timeout", "5xx", "429", "connection-reset"],
    "neverRetry": ["eth_sendRawTransaction"]
  },
  "breaker": {
    "failureThreshold": 5,
    "windowMs": 30000,
    "cooldownMs": 15000,
    "halfOpenProbes": 1
  }
}

Observability

Every routing decision is recorded. The same numbers Prism uses to pick a provider are the ones you see.

Provider health

live

Provider	p50	p99	Success	Share
Alchemy	41 ms	118 ms	99.98%	34%
QuickNode	38 ms	104 ms	99.99%	38%
Ankr	52 ms	191 ms	99.81%	19%
Chainstackbreaker open	-	-	87.40%	0%
Public node	96 ms	402 ms	99.20%	9%

Metrics reference