Oracles & Price Feeds

01 · Concept — what problem does it solve?

A smart contract can't see the outside world. But lending, liquidations, perps, and stablecoins all need to know what is ETH worth right now? An oracle brings that answer on chain. It is also the single most attacked surface in DeFi: if you can lie to the oracle for one transaction, you can borrow against worthless collateral, trigger false liquidations, or drain a pool. Most "DeFi hacks" are really oracle-manipulation attacks.

02 · Mechanics

• Push oracles (Chainlink): a decentralized network of node operators aggregates prices from many sources and writes the median on chain on a schedule or deviation threshold. Manipulating it means corrupting a majority of independent nodes — expensive.
• Pull oracles (Pyth): prices are signed off-chain and pulled on chain by the user at the moment of use — lower latency, fresher, the consumer pays the update gas.
• TWAP (on-chain): derive a price from a DEX's own time-weighted average (Uniswap v3 accumulators). No external trust, but only as deep as the pool.
• Aggregation & deviation: good feeds take a median across many sources and only update on meaningful moves, filtering single-venue spikes.

03 · Formulas

// decentralized median feed
price_onchain = median(node₁, node₂, … nodeₙ)   // corrupt > n/2 to move it

// time-weighted average price (manipulation cost ∝ duration × depth)
TWAP = (cumulativePrice_now − cumulativePrice_then) / (t_now − t_then)

// attack math
single-block spike → cheap for spot, but TWAP needs sustained capital

04 · Edge cases & risks

• Flash-loan spot manipulation — borrow a fortune, slam a thin pool to skew its spot price, exploit a protocol reading that spot, repay — all in one transaction. The classic oracle attack; the reason naive getPrice() from a single DEX is fatal.
• TWAP ≠ safe — TWAP resists one-block spikes but is still beatable by multi-block manipulation, and it lags fast crashes (liquidations fire late → bad debt).
• Stale / frozen feeds — if a feed stops updating in a crash (or an L2 sequencer halts), protocols act on a stale price; good integrations check feed freshness.
• Centralization of "decentralized" feeds — if too many protocols read the same oracle, that oracle is a systemic single point of failure for the whole ecosystem.