§ 01 / heroperformance infrastructure · solana live
Stream shreds.Route the packetsthat land the block.
p50 18 ms ingressland 99.94 %tput 91k shreds/sregions ewr · fra · ams · sgp
ample.sh is the data plane for Solana teams who care about the 99th percentile.
Shred-level ingestion, decoded streams, and dual-path transaction delivery —
built on the same primitives validators use, exposed where you actually need them.
§ 02 · trusted by — teams who measure the 99th percentile14 of 32 top searchers · 6 of top 10 validators
mango/v4
drift
jupiter
phoenix
kamino
marginfi
tensor
orca
raydium
meteora
backpack
zeta
pyth
switchboard
§ 03 · manifesto
we ship at the layer below RPC.
Solana runs on microseconds. Most infra wraps RPC and calls
that a product. We work at the wire — shreds, streams, routing
— so the apps on top inherit the network's actual speed,
not a proxy of it.
P · 01
Latency is a product surface.
A 50 ms gap is the difference between a fill and a miss. We treat the p99 as a feature, not a footnote, and publish it with the SLA.
P · 02
Don't share memory with the validator.
Geyser plugins live in the validator's process. One bad plugin slows the whole network. We offload to a proxy so the validator stays cold-paged and predictable.
P · 03
Two paths, one promise.
Jito wins under congestion. Staked relay wins on average. We ship both, route automatically, and surface per-path land rate so you can audit every signature.
P · 04
Open source where it matters.
The proxy is Apache-licensed because the network benefits when validators run inspectable software. The decoder and router are commercial because they fund the work.
P · 05
Numbers we publish, not numbers we pick.
Rolling 90-day windows, raw CSVs, same node, same hour. If a benchmark only works in a press release, we don't put it in one.
§ 04 · architecture
the path a transaction actually takes
Five layers. No middleware tax. Drop in at the layer you care about — and only that one.
L0
Validator
turbine · leader schedule · QUIC ingress
92%util
L1
Geyser Proxy oss
offload · zero-plugin · fan-out
78%util
L2
Shred Ingress
udp · 4 regions · 12 nodes
88%util
L3
Decoder
structured · typed · indexed
66%util
L4
Dual-path Sender
jito + staked relay · auto
95%land
FIG·02 · global ingest map 4 regions live
ewrus-east · newark17ms3.4 Gb
fraeu-central · frankfurt19ms2.9 Gb
amseu-west · amsterdam18ms3.8 Gb
sgpap-south · singapore22ms2.3 Gb
§ 05 · products
five instruments, one rack
Each product is independently consumable. Use one. Use all five. Drop in at the layer that buys you the most time.
Drop-in config, zero plugin overhead on the validator host. Stream account, slot, and transaction events to as many consumers as you want — without touching the box that's earning your stake.
Shred Ingress / Egress. Closer to the leader than the competition.
Stream raw shreds in, publish them out, across four regions. The same firehose validators see — exposed to the apps that need it. No polling, no batching tax.
Skip the parsing. Subscribe to accounts, instructions, and reconstructed blocks as they emerge — typed, indexed, and filterable. Protobuf schemas, gRPC + WebSocket + REST.
Dual-path delivery across Jito and staked relays. Auto-fallback, retry with fresh blockhash, land-rate telemetry per program and per wallet. Every signature traced end-to-end.
Engineering teardowns, not marketing. We publish the CSVs.
post · engineering2026.03.28· 12 min
Routing shreds in 18ms: what the p99 actually costs
A teardown of ingest-to-decode latency across four regions, with flamegraphs from our UDP path and a frank accounting of where we still lose microseconds. Includes the bench we use internally and the CSVs you can reproduce.
by m. okonkwo2,418 readsdiscuss on lobste.rs
post · validators2026.03.14· 9 min
Why we gave up on Geyser plugins (and what replaces them)
by s. petrov · 3,901 reads
post · research2026.02.21· 14 min
Dual-path sending, measured: 12 weeks of land-rate data