Skip to content

Benchmark

scripts/benchmark.py syncs one small public account per chain into an empty database, so each run reads the account's whole history, one chain at a time, at each source's default rate. Requests include retries.

python scripts/benchmark.py --runs 3 --output results.md

Measured 2026-11-15 15:24 UTC from one home connection, 3 runs per account, median kept:

Chain Network Records Requests Seconds Records per second Requests per record
algorand testnet 108 3 1.2 89.8 0.03
solana devnet 6 9 17.9 0.3 1.50
near testnet 20 7 9.8 2.0 0.35
ethereum testnet 20 65 14.8 1.4 3.25
base testnet 7 145 65.8 0.1 20.71
bitcoin testnet4 4 5 4.7 0.9 1.25
bitcoin testnet4 1689 82 125.5 13.5 0.05

What sets the pace is the public endpoints' limits, not this code:

  • Algorand's Indexer returns up to 1,000 transactions a page, at 10 requests a second.
  • Solana's public RPC allows about one request every two seconds, and each transaction is read on its own.
  • NEAR reads history from FastNEAR at one request every two seconds.
  • Ethereum lists history from Blockscout, then reads each transaction and its receipt from the RPC, so requests grow with the records.
  • Base lists history from Alchemy's transfers API in windows of a million blocks, with one request for what the account sent and one for what it received in each window, then reads each transaction and receipt; a quiet account on Base Sepolia's long chain costs requests in every window, which is why 7 records took 145.
  • Bitcoin reads 25 transactions a page from mempool.space at one request a second. The larger account is a busy testnet4 miner; mempool.space refused or dropped some requests, which were retried (26 retries over its three runs).

A dedicated endpoint with a higher MCI_REQUESTS_PER_SECOND changes these numbers far more than anything in the code does.