Wirewiki's June post explains how it prefetches 240 million domains
In a June 22 engineering breakdown, founder Ruurtjan Pul describes prefetching the next character, compressing the index to 2.5 GB, and accepting 100-200 ms of added US latency.
By RuntimeWire Staff · Published
Primary source: Ruurtjan Pul
Why it matters
Pul turned the physical act of typing into a network latency budget. The pattern gives small software teams a practical way to make remote data feel local without immediately building a global server fleet.
In a June 22 engineering breakdown, Ruurtjan Pul (@ruurtjan), the solo founder behind Wirewiki, described an autocomplete system designed to feel instantaneous while searching roughly 240 million domain names.
The "p99 0 ms" headline came with a narrow and useful definition. Pul measured latency from the moment a user releases a key until suggestions are ready to render. Wirewiki starts fetching earlier, when the key goes down, and uses the time occupied by the user's next keystroke to complete the network request. If the response arrives before the next key comes up, the interface records zero milliseconds of perceived latency.
Pul found that this window gave him a p99 budget of 121 milliseconds while he typed 100 domain names at what he described as a reasonably fast pace. His server still has to receive, process and return each request. Wirewiki simply does that work before the user expects to see the answer.
The distinction captures Pul's larger bet on Wirewiki. DNS lookup tools are plentiful, and generating another utility page has become cheaper. Pul chose interface speed and connected navigation as the points where a small, founder-run product could still separate itself.
Spending the user's typing time
Wirewiki's autocomplete is the main route into its collection of DNS records, IP addresses, delegation data and email configuration tools. Pul made it speculative: each request returns suggestions for the current prefix and precomputed results for every valid next character.
A request for "wi," for example, can return the leading results for that prefix alongside separate lists for "wia," "wik," "wiz" and the other possible continuations. When the user presses the next key, the likely result is already in the browser cache.
The approach sends extra data because each request can include suggestions for the current prefix and precomputed results for possible next characters.
Pul split the search index into a popular "head" and a much larger "tail." The head uses an in-memory character trie containing eight precomputed suggestions for each prefix. It draws from Tranco, a research-oriented ranking of one million popular domains.
The tail is a sorted, delta-compressed index stored on SSD and accessed through memory mapping. Pul said a 27 MB in-memory directory points to fixed blocks of 256 names, limiting each lookup to a binary search followed by a short linear scan. He reported that the full set of roughly 240 million names occupies about 2.5 GB on disk.
Most of those additional domains came from zone files accessed through ICANN's Centralized Zone Data Service. CZDS provides access to files from participating generic top-level domains such as .com, .net and .org. Country-code domains do not carry the same contractual access requirement, leaving gaps that Wirewiki partly fills through the Tranco ranking. Pul has considered certificate transparency logs and Archive.org as additional sources but had not integrated them when he published the article.
A benchmark with one author and one server
Pul said an LLM-generated stress test produced 720,000 keystroke queries by simulating 60,000 typed domain names. The test replayed those requests at fixed rates against the API, the Nginx origin path and the complete route through Cloudflare.
Most API-only requests finished within 2 milliseconds, according to Pul's results. At 1,600 requests per second, he reported a p99 of 15 milliseconds for Nginx and the API together. He also said the system remained within his local latency budget while simulating 1,000 people typing simultaneously.
Those figures are Pul's production measurements, rather than an independent benchmark. The larger limitation is geography. Wirewiki runs on one server in Europe, and Pul said US traffic can add 100 to 200 milliseconds. That delay is enough to break the p99 zero-millisecond result for users across the Atlantic.
Pul could place servers in multiple regions and route each user to a nearby instance. In a June 22 X post, he wrote: "That would give me the p99 0 ms* latency. But that's a bit much. Even for me."
That restraint is part of the engineering story. Pul optimized the code path until the network became the dominant cost, then stopped before a free autocomplete feature turned into a distributed infrastructure project. He said the end-to-end delay in practice is roughly the browser-to-server round trip through Cloudflare plus 10 milliseconds.
The second DNS product
Wirewiki launched on February 11, 2026 as Pul's attempt to build a browsable map of internet infrastructure. The site connects domains, IP addresses and DNS servers, allowing users to move through related records instead of opening a separate lookup tool for every question.
Pul arrived at that design after building Nslookup.io while working as a software and infrastructure consultant at BigData Republic. Wirewiki's account of his background says he later sold Nslookup.io and moved full time onto Wirewiki. Wirewiki's press materials say Nslookup.io had surpassed 600,000 monthly users, a figure supplied by Pul.
In his February launch post, Pul said the DNS-focused name had placed a ceiling on how far Nslookup.io could expand. Wirewiki gave him room to connect DNS records with broader infrastructure layers, including hosting providers, content delivery networks and autonomous system numbers.
Wirewiki remains free and requires no account. Its published business model calls for brand integrations with hosting providers, registrars and infrastructure vendors. Pul also left the door open to selling access to the autocomplete API, though he described that opportunity as too niche to pursue without clear customer demand.
The autocomplete work shows how Pul plans to compete before those commercial pieces mature. The product's current edge is a fast workflow: Pul makes a common lookup experience unusually quick, exposes the implementation and lets the interface carry the argument.