Methodology

How the RPC selector ranks providers

The selector gives the answer first, then exposes the measurement and scoring logic needed to trust or reject that answer. Schema version remains v3.

Collection

InfraBench collects Solana mainnet-beta RPC observations with fixed methodology. Raw events, runner internals, and private endpoint values remain private.

Aggregation

The public website reads static v3 artifacts from /snapshots. Provider summaries expose p50, p95, p99, success rate, sample count, freshness, and public observations.

Selector Scoring

The selector computes a transparent weighted score from latency, reliability, freshness, sample quality, and nearest-region match. Workload changes adjust those weights.

Confidence Limits

Stale, distant, low-confidence, or insufficient evidence does not disappear. It forces abstention or isolation from the primary ranking.

Workload weights

Reads

35% latency, 30% reliability, 15% freshness, 10% sample quality, 10% region match.

Trading

45% tail latency, 25% reliability, 10% freshness, 10% sample quality, 10% region match.

Indexing

15% latency, 35% reliability, 20% freshness, 20% sample quality, 10% region match.

Wallet

25% median latency, 40% reliability, 15% freshness, 10% sample quality, 10% region match.

Release gates

Any critical gate failure blocks a trusted recommendation and leaves the selector in preview or abstention mode.

Freshness

Public snapshot must be within the 24 hour recommendation SLA.

Coverage

Nearest measured region must be within the 250 km fallback boundary.

Confidence

Overall and coverage confidence must be above low for a prominent recommendation.

Eligibility

Providers need aggregate latency, successful observations, and sufficient samples before entering the primary table.

Measurement pipeline

Step 1

Collection

Step 2

Aggregation

Step 3

Selector Scoring

Step 4

Confidence Limits