Oracles: how blockchains learn prices, and why it matters

The oracle problem

Every node in a blockchain network re-runs each transaction and must reach exactly the same result. That is how thousands of strangers agree on balances without trusting one another. If a contract could call a website, two nodes might get different answers, or the same node a different answer a minute later, and agreement would collapse.

So blockchains reach consensus using only data stored on the chain itself. Outside data has to be written there by someone, as a transaction, and that someone is an oracle. The oracle problem is the catch: a contract is only as trustworthy as the data it acts on, and relying on one data provider brings back the very trust that blockchains were meant to remove.

How a decentralized price feed works

Chainlink Data Feeds are designed so that no single party decides the price. Chainlink describes three layers of aggregation. Data aggregators first combine raw prices from many exchanges, weighting them by volume and filtering out outliers. Each node operator then pulls from several of these aggregators and takes the median. Finally, the network combines the answers of many independent node operators and takes the median again.

The nodes do the heavy lifting off-chain, with a protocol called Offchain Reporting (OCR). They exchange signed observations over a peer-to-peer network, and a leader node gathers them into a report. Once a quorum of nodes has signed the report, one transaction delivers it on chain, where the aggregator contract checks the signatures and stores the median with a timestamp and a round number.

Writing every tiny price move on chain would waste fees, so a feed updates when one of two triggers fires. The deviation threshold starts a new round when the off-chain price moves more than a set percentage away from the on-chain value. The heartbeat starts one when a set time has passed since the last update, even if the price has barely moved.

These are the settings Chainlink lists as of September 29, 2026. A calm stablecoin feed may update once a day, while ETH/USD can update many times an hour when markets swing.

Why a single market's price is dangerous

The simplest way to get a price on chain is to read it from one place: one exchange or one pool on a decentralized exchange (DEX). But a price is only as sturdy as the amount of money needed to move it, and on a thin market that can be surprisingly little.

Flash loans make it worse. A flash loan must be borrowed and repaid within the same transaction; if it is not repaid, the whole transaction is undone. Anyone can therefore command millions for a moment, without collateral, as long as the attack pays for itself in one go.

The bZx attack, February 2020

On February 18, 2020, an attacker used exactly this against bZx, a lending and margin-trading app on Ethereum. bZx priced the stablecoin sUSD from two small on-chain markets, a Uniswap pool and a Kyber reserve. Researchers at Imperial College London later reconstructed the steps:

In those two pools, sUSD briefly cost about 2.5 times as much as elsewhere, so bZx lent far more ETH than the collateral was worth. The attacker kept 2,381.41 ETH, about $634,900 at the time, for a transaction fee of $118.79. bZx's lenders bore the loss.

Averaging is not a cure-all

Combining sources does not help if they all watch the same thin market. In October 2022, according to a complaint by the US Commodity Futures Trading Commission (CFTC), a trader on Mango Markets, a crypto trading and lending platform, bought large amounts of the MNGO token on the three exchanges that fed the platform's oracle. The oracle's MNGO price jumped more than 13-fold in 30 minutes, and he used his inflated positions as collateral to withdraw over $110 million.

Chainlink's own documentation warns that thinly traded assets can be manipulated by flash-loan-funded attacks or by a well-capitalized actor.

How lending and trading apps use price feeds

On a lending app, you deposit collateral and borrow against it, and the app needs a price for every asset to know whether your loan is still safe. Aave expresses this as a health factor: your collateral's value, multiplied by a safety factor called the liquidation threshold, divided by what you owe. Below 1, others can repay part of your debt and take some of your collateral plus a bonus, which is called a liquidation. Aave V3's price oracle takes its prices from Chainlink aggregators and turns to a fallback oracle if a feed returns zero or less.

A wrong price is costly either way. Too low, and healthy borrowers get liquidated. Too high, and borrowers can walk away with more than their collateral is worth, as bZx's lenders found out.

Leveraged trading apps need prices faster than a feed's heartbeat. Many use pull-based feeds such as Chainlink Data Streams, which deliver sub-second prices as signed reports that are verified on chain in the same transaction that uses them.

Careful apps add their own checks. Chainlink recommends bounds, circuit breakers and freshness checks. The Sky protocol, formerly MakerDAO, passes prices through an Oracle Security Module that holds each new value back, by default for an hour, so there is time to react to an oracle attack.

On-chain prices are open to everyone

Because feeds live on public blockchains, every update is a transaction anyone can inspect. You can read a feed's latest answer, when it was updated and its round number, look back through earlier rounds, and see which nodes serve it. When a lending app liquidates someone, the price it used is on the record.

For example, the ETH/USD feed on Ethereum sits behind the proxy address 0x5f4eC3Df9cbd43714FE2740f5E3616155c5b8419. Calling latestRoundData on a block explorer returns the price, with eight decimals, and the time of the last update. nowcoin reads the prices it shows from Chainlink feeds on chain in the same way.

That openness has limits. You can see what the network reported and when, but not every raw input from the data providers behind it.

Other designs: Pyth and Uniswap TWAP

Chainlink's push feeds are one design among several:

Pyth takes a pull approach. Its prices come from more than 120 first-party providers, including exchanges, banks, trading firms and market makers, and update every 400 milliseconds. Each price carries a confidence interval, a range that widens when providers disagree. In its standard pull model, a price reaches a blockchain only when someone needs it: anyone can submit a signed update to the Pyth contract, which verifies it, usually in the same transaction that uses the price.

Uniswap offers a different tool, the time-weighted average price, or TWAP. A Uniswap v2 pool keeps a running total of its price, measured at the start of every block and weighted by how long each price lasted; comparing that total at two moments gives the average over the period. To move a one-hour TWAP by 5%, an attacker would have to hold the pool 5% off the market for about an hour, paying arbitrageurs all along. Uniswap v3 pools can store up to 65,535 observations, while v4 has no built-in oracle.