In This Article
Polymarket has opened a test environment for Nex, a trading system rewritten in Rust that, according to its development team, can handle 100,000 commands per second, ahead of a tentative launch in mid-December.
Polymarket’s vice president of DeFi engineering, Josh Stevens, announced the launch of the test environment on X on November 13, saying that Nex had been running in parallel with the existing exchange during the previous week. It duplicates every production order, allowing engineers to compare response times, stability under load, and matching accuracy against real trading activity.
The platform has not yet switched to Nex. Stevens said the team plans four weeks of “shadow” monitoring before making a decision on deploying the system in mid-December.
Polymarket Nex plans to increase system performance 10-fold over the current level
According to the performance figures published by Stevens, the current central limit order book (CLOB) processes 10,000 commands per second under sustained load and 15,000 during short bursts. Nex supports 100,000 and 150,000 commands per second, respectively.
Commands describe actions such as placing and canceling orders; the figures cited do not reflect the number of completed trades per second.
According to Stevens, with further optimization the system will be able to process 200,000 commands per second, with peaks of 250,000 commands. Larger changes to the underlying code could increase throughput to 1 million commands per second within the next year, although, he said, such a level is not yet required.
According to the announcement, startup time has been reduced from five minutes to 15 seconds. Stevens described the old CLOB as having data race issues and inconsistent behavior, saying that Nex was designed to deliver stable results and eliminate these problems.
The team began reworking the system in August, concluding that the existing system needed to be rebuilt from scratch.
Order performance metrics measure internal exchange speed
At the 99th percentile (p99), placing an order took 3.76 milliseconds on Nex versus 1.55 seconds on the existing system, roughly 412 times better, according to the published comparison. The p99 metric measures the time within which 99% of measured requests are completed.
Order cancellation time was reduced from 627.92 milliseconds to 3.20 milliseconds, and market order cancellation time from 810.16 milliseconds to 5.14 milliseconds.
“The figures below reflect internal exchange times, not end-to-end performance,” Stevens wrote, referring to start-to-finish performance. He added that the corresponding measurements will be published later.
This distinction means that the published figures do not reflect the entire path from a trader submitting an instruction to receiving the result.
At the p95 level, order placement time was reduced from 341.84 milliseconds to 1.59 milliseconds. The corresponding figures for order cancellations were from 78.22 milliseconds to 1.58 milliseconds, and for market order cancellations from 84.89 milliseconds to 2.27 milliseconds.
The changes affect order cancellations and market data, but they are being introduced separately
The announcement described a change in how the exchange handles requests to cancel each open order. Currently, the CLOB confirms the request before completing the cancellation in the background, which typically takes another 600 milliseconds to two seconds.
According to Stevens, Nex responds only after each order has actually been canceled. The p99 request execution time was 3.14 milliseconds, compared to roughly 0.64–2.04 seconds for the full cancellation process in the existing system.
The Rust-rewritten WebSocket service, which transmits changes in the order book, prices, and executions, showed a p99 write latency of 0.099 milliseconds versus 44.5–88.1 milliseconds previously. Stevens noted that delays in data transmission can cause traders to quote on stale prices.
The WebSocket improvements can be deployed separately from the exchange, so the team may release them ahead of the Nex launch.
