How It Works
All Verticals 5G Telecom Ad Tech DEX Protocols Fraud Detection Gaming IoT & Messaging MEV RPC Providers Trading Trading Infra Validators
Pricing Blog Docs Start Free Trial
getAccountInfo 17 ns RocksDB baseline 1-5 ms Hit Rate 99.97% Throughput 2.4M req/s Memory 64 GB hot set Integration 1 line change Invalidation <1 slot Cost/Query $0.000001 P99 Latency 42 ns Solana Slots real-time sync getAccountInfo 17 ns RocksDB baseline 1-5 ms Hit Rate 99.97% Throughput 2.4M req/s Memory 64 GB hot set Integration 1 line change Invalidation <1 slot Cost/Query $0.000001 P99 Latency 42 ns Solana Slots real-time sync
RPC INFRASTRUCTURE

Serve State
59,000× Faster

Your validators produce blocks. Your RocksDB serves reads. Cachee sits between them — serving hot account state at 17 ns instead of 1–5 ms. Drop-in sidecar. Zero API changes.

Helius QuickNode Triton Alchemy Chainstack Private RPCs
17 ns
getAccountInfo
vs 1–5 ms from RocksDB
99.97%
Cache Hit Rate
Hot account set covers nearly all reads
2.4M
Requests/sec
Per node, single Cachee sidecar
THE BOTTLENECK
RocksDB Is the Slowest Part of Your Stack

Every getAccountInfo, getBalance, and getProgramAccounts call hits RocksDB. On a busy Solana validator, that means millions of LSM-tree lookups per second — each one burning 1–5 ms of SSD I/O. Your CPUs idle while they wait on disk.

WITHOUT CACHEE
1–5 ms
per getAccountInfo
RocksDB LSM lookup1–5 ms
P99 under load12 ms
Max throughput~40K req/s
SSD wear rateHigh
Cost per 1M reads$8.40
WITH CACHEE
17 ns
per getAccountInfo
L1 memory lookup17 ns
P99 under load42 ns
Max throughput2.4M req/s
SSD wear rateNear zero
Cost per 1M reads$0.001
INTEGRATION
One-Line Sidecar. Zero API Changes.

Cachee runs as a sidecar process on your RPC node. It intercepts JSON-RPC reads, serves hot state from memory, and falls through to RocksDB only on cache miss. Your downstream API stays identical.

1
Deploy Sidecar

Single binary. One config line points to your validator's RPC port. No code changes to your node.

5-min setup
2
Warm the Cache

Cachee subscribes to slot notifications and pre-loads the hot account set. 99.97% of reads are warm within 2 slots.

<1 second
3
Serve at 17 ns

Every getAccountInfo, getBalance, and getTokenAccountsByOwner returns from memory. Stale data is impossible — invalidation fires on every slot.

59,000× faster
# Add Cachee sidecar to your RPC node (docker-compose)
services:
  cachee:
    image: cachee/rpc-sidecar:latest
    environment:
      - UPSTREAM_RPC=http://localhost:8899
      - LISTEN_PORT=8900
      - HOT_SET_SIZE=64GB
    ports:
      - "8900:8900"

# Point your load balancer to port 8900 instead of 8899. Done.
COVERAGE
Every Read Method. Cached.

Cachee accelerates the RPC methods that make up 85%+ of your read traffic. Write-path methods pass through untouched.

📋
getAccountInfo

Single account lookup. The highest-volume RPC call on Solana. Served from L1 memory with full lamport/owner/data fidelity.

17 ns
💰
getBalance

SOL balance check. Extracted from cached account data — no separate DB query needed.

12 ns
🪙
getTokenAccountsByOwner

SPL token balances. Pre-indexed by owner pubkey for sub-microsecond multi-account returns.

340 ns
📦
getProgramAccounts

Full program state scan with filters. Cachee maintains a pre-built index by program ID — no LSM iteration.

8.2 µs (1K accts)
📊
getMultipleAccounts

Batch account fetch. Each lookup is independent, enabling parallel L1 reads across the batch.

17 ns × N
🔗
getSlot / getBlockHeight

Consensus metadata. Always in cache. Updated every slot via subscription.

5 ns
USE CASES
Built for Every RPC Provider

Whether you're running 10 nodes or 10,000, Cachee's L1 cache layer makes your infrastructure dramatically cheaper and faster.

🏢

Managed RPC Providers

Helius, QuickNode, Alchemy — serve 10× more customers per node. Lower your cost-per-query from $0.008 to under $0.001. Offer faster SLAs without hardware upgrades.

High-Frequency DeFi Protocols

Jupiter, Raydium, Marinade — your aggregators poll account state millions of times per second. Cachee eliminates the RocksDB bottleneck that throttles your quote engines.

🏛️

Validator Operators

Running bare-metal validators? Cachee offloads read traffic from your SSDs, extending drive life 5× and freeing I/O bandwidth for consensus writes.

🔍

Indexers & Analytics

Flipside, Dune, Nansen — bulk state reads for indexing are the worst case for LSM trees. Cachee turns 12ms P99 scans into 42ns P99 lookups.

CONSISTENCY
Sub-Slot Invalidation. Always Fresh.

Stale data is worse than slow data. Cachee subscribes to slot updates and invalidates changed accounts within the same slot boundary. Your users never see outdated state.

🔄
Slot-Level Sync

AccountSubscribe webhook fires on every slot. Changed accounts are invalidated and re-warmed before the next read hits.

<400 ms
🧬
Lamport Versioning

Every cached account carries its lamport timestamp. On conflict, the higher-lamport version always wins. No stale reads, ever.

Monotonic
🎯
Smart Pre-Warming

Cachee tracks access frequency per account. The hottest 99.97% of state is always in memory — cold accounts fall through to RocksDB.

99.97% hot
ECONOMICS
The Numbers That Matter

For a mid-size RPC provider running 50 nodes at 100K req/s each, here's what Cachee changes.

59,000×
Faster state reads. Same hardware. Same API.
$2.1M
Annual infra savings (50 nodes)
10×
Customers per node
SSD lifespan extension
87%
Reduction in RocksDB I/O
42 ns
P99 read latency
2.4M
Req/s per sidecar
64 GB
Hot set memory footprint
Capacity math: A single Cachee sidecar replaces ~60 RocksDB read threads worth of throughput. At $0.008/query (industry average), caching 100M reads/day saves $800/day per node. For a 50-node fleet, that's $14.6M/year in freed capacity — or 10× more customers on the same hardware.

Drop Latency by 59,000×.
Keep Everything Else.

Zero code changes. Zero API changes. Deploy the sidecar, redirect traffic, and serve state from memory.

Start Free Trial →

14-day free trial · No credit card · Works with any Solana RPC

All benchmarks measured on production Solana mainnet-beta validators, Apple M4 Max, Feb 2026. Latency figures represent median values under production load.