How Solana Network Pulse measures

Cohorts

reference-probe — independent VMs with full ContactInfo (gossip + TVU + TPU), optional observe-only repair, no mesh inject, no Path A egress, no tx-relay.

solanacdn-pop — serving POPs (fra-1, sgp-1, ams-1, lax-2). Published only as “SolanaCDN observed paths.” Never mixed into “the network.”

What is measured

• First unique shred arrival per FEC set, classified by source IP.
• Time to recoverable FEC (32-of-64 / num_data unique shreds).
• Share of sets that still need repair at the deadline.
• Edge TVU race: native vs SolanaCDN Path A vs repair, by region — omitted until 7 dual-path days exist. Never approximated from POP ingest tags.

What is not measured

• A composite network health score.
• Mesh loss_ppm as Turbine loss.
• Probe-miss as a skipped slot (skips come from RPC block production).
• Spy-node gossip (no TVU) as a Turbine sample.
• Serving-POP observation as an independent sensor.
• Named validator skip without consent.

Clocks

Race on one NIC needs no PTP. Cross-region percentiles are t_first relative to the earliest probe for that (slot, FEC set). Samples with NTP |offset| > 5 ms are dropped. Every figure discloses stratum, sample size, exclusions, and that this is not the entire Solana network.

Incident glossary

Cluster — RPC skip/halt, confirmed by getBlockProduction.

CDN — Path A silent, mesh drop, or replayable-stuck from network_incidents. Not cluster consensus failure.

Unstaked leaf

Probes join as unstaked nodes. Stake weighting places them at the leaves of the Turbine tree, so some shreds never arrive over native Turbine. That is expected. Completeness uses FEC recovery and observe-only repair, not raw shred counts.