Latency asymmetry in perp DEXs

Every perpetuals venue advertises its end-to-end latency. Almost none of them talk about the shape of that latency — who experiences it, when, and what it costs the people on the slow side. This note looks at where latency actually hides in an onchain perp stack, and why the asymmetry matters more than the average.

The three clocks

An onchain perp trade passes through three clocks that rarely agree: the oracle's clock, the sequencer's clock, and the liquidator's clock. Each is fast in isolation. The risk lives in the gaps between them.

When the oracle updates before the sequencer includes your cancellation, you are quoting a stale price with live size. When the liquidation engine reads margin against a price the market has already left, positions are closed at levels no counterparty would agree to. Neither event shows up in a latency histogram.

Where the asymmetry comes from

• Colocation by another name. Whoever runs infrastructure closest to the sequencer sees state first. The advantage is structural, not episodic.
• Cancel priority. Venues that don't prioritize cancels over takes convert every oracle update into a race the maker loses by default.
• Batch boundaries. Auction or block boundaries create predictable moments where stale quotes are harvestable at scale.

What we measured

Replaying a week of production order flow through an instrumented fork of the matching pipeline, the median oracle-to-cancel window was unremarkable. The tail was not: the slowest decile of cancels absorbed a large majority of adverse fills. Averages describe the venue; tails describe who pays for it.

Takeaways

Latency asymmetry is a market-structure decision that gets made implicitly, by default, in the plumbing. Venues that treat cancel priority, oracle sequencing, and liquidation timing as first-class design surfaces will keep their makers. Venues that don't will rent them.

This is a placeholder article. Real research from the team will be published here.