Knowledge, paid for in mistakes, time, and money
Ten Solana rules I learned the expensive way — so you can skip the invoice.
For context: I learned Solana by poking at it. No structured plan, no mentor, no curriculum — just trial, error, and invoices. The tuition was paid in my own money, my own sleep, and a few servers that never should have been rented.
If the Solana from zero series is the map, this note is the list of places where I fell into a hole. Ten rules, each one bought with something — time, money, usually both. Over time each rule will grow into its own note with the full story and the numbers behind it; for now, every rule stands on its own.
Be more skeptical of these than I was of anything back then. That’s partly the point.
Infrastructure & data
1. Don’t buy a server for your RPC node. Just pay a provider.
In 99.9999% of cases renting access is cheaper than running your own — once you count what your time is worth. I know because I did the math wrong: I rage-quit Helius over pricing, rented an 800€/month Hetzner monster, and spent three weeks of sleepless nights getting it to sync — to serve a workload that never needed a dedicated node in the first place.
And even once it worked, my node was still seeing transactions later than Helius’s. They have serious stake behind their infrastructure, and stake-weighted connections mean their nodes hear about everything before an unstaked one ever will. You can rent the hardware; you can’t rent their position in the network.
Then I went back to Helius.
2. Don’t write your own parsers and SDKs from scratch.
You’ll learn a lot — I did — but maintaining them is a second job. Some DEXes ship breaking changes every two weeks, and your indexer and trade bot die at 2am until you fix them. Use maintained software: the fnzero repos cover 99% of what a Solana trader or parser-builder needs. Build from scratch to learn, never to depend on.
3. Your bot’s language doesn’t matter.
Rust, TypeScript, Python, Elixir — for a normal trade bot they all land in the same blocks. Unless you’re pushing hundreds of transactions per second, the network, not your runtime, is the bottleneck. Pick the language you’ll still enjoy debugging at 2am.
4. logsSubscribe is a toy, not a data pipeline.
Every beginner example uses it; your dataset shouldn’t. It drops, it lags, and it tells you fragments. For real data you want whole blocks — block subscriptions or Geyser.
5. Estimate your credit burn before you write the script.
I designed my pipeline first and looked at pricing second; the free tier died in an afternoon. Every wide scan multiplies. Price the queries per iteration, then multiply by your loop — before writing a line of code.
Trading
6. Shred Stream won’t make you a winner.
Even at shred level, every transaction you see has already been sequenced by the leader. You’re racing to react, not to front-run — and the people you’re racing are colocated. Don’t pay for latency you can’t convert into an edge.
7. Copytrading doesn’t work.
By the time a wallet is visible enough to be worth copying, you’re too late — and after latency, slippage, and priority fees, copying it is negative EV. I spent real time studying wallets that always printed, and it was never luck: they were scammers, bloggers trading against their own audience, or guys with insane infrastructure — own validators included. Every one of those edges is structural. You can’t copy it; you can only exit-liquidity it. The full story — my indexer, the counter-bots, the backtests — is in The casino has better odds.
Discipline
8. Backtest your theories on historical data. It’s sobering.
Every idea feels brilliant until you run it against what actually happened. Most of mine died there — which is the cheapest possible place for them to die.
9. A backtest that ignores fees, slippage, and latency is fiction.
Mine looked beautiful until I added fees, priority fees, slippage — and execution delay: building the transaction, the network round-trip, waiting to land in a block. Then it went to zero. Model the full cost and the full latency of every fill, or you’re testing a fantasy market where trading is free and instant.
10. Assume everything you see on-chain is manufactured until proven otherwise.
Volume gets washed, holders get sybiled, trending gets bought, PnL screenshots get cropped. The chain doesn’t lie, but almost everyone presenting chain data to you does. Verify from raw data or don’t believe it. What manufactured looks like from the inside — the master-wallet play, the coordinated “organic” buying — is in The casino has better odds too.
That’s the list. Each rule here will eventually link to its own full story — the server saga behind rule 1 alone is worth a note, and the backtest autopsy behind rule 9 has actual numbers attached.