Nix’s substituter model is one of those designs that is almost right, but isn’t exactly there. It is simple in itself: you list a few binary caches in nix.conf, the daemon walks them in order, and if the path you want is anywhere on the internet a build doesn’t have to happen on your poor laptop from 2001. The binary cache system is often listed as a strength of NixOS , but it’s actually a strength of Nix, and a bare minimum for something like NixOS to work for its users. Which is to say that it’s actually perfectly fine until you have a large enough multi-cache setup for your configuration’s dependencies, or in rarer cases, your projects dependencies. By which I mean multiple binary cache instances because you decided to fetch massive C++ and Rust projects from the internet that need to be built, on your system.

In the case of a rather complex setup, the first cache in your substituter list is almost always https://cache.nixos.org. It is fast, it is global, and it has the binary substitutions for your system to be built in a few minutes. It does not, however, have your overlay’s packages. The second tends to be something like the nix-community cache, because you usually pull something really useful 1 from the nix-community organization or just hope it has some overlay packages that you need in your obscure hardware setup. In less common cases you also add a third party project’s Cachix (or a similar 3rd party) and occasionally, if you’re technical (or dedicated) enough, your homelab’s private cache. In such a setup, every narinfo lookup walks that rather large list. Every nix-shell -p hello becomes a serialized scan across four hosts on three continents because Nix has no concept of which substituter is most likely to answer for this path. It just asks them all, in the order you wrote them down, one after the other.

The Shape of the Problem

To explain why a proxy is the right answer, you—or rather, I—have to be honest about what Nix’s substituter logic is:

  • A loop over substituters, in order.
  • For each: HEAD /<hash>.narinfo. If 200, fetch and use it. If 404, continue.
  • No concurrency. No latency tracking. No memory of which cache won last time you asked for this hash.

The substituter list is a preference, not a routing table. There is a priority field, but it is a static integer chosen at config time. It does not know that cache.nixos.org is 40 ms away on your home connection and 800 ms away from the office VPN. It does not know that your private cache is the only one that has the path. It does not know anything, because there is nothing to know. The daemon is stateless on every request. For a tool that prides itself on being a model of declarative purity, the network layer underneath is still static and request-local.

ncro, Briefly

ncro—Nix Cache Route Optimizer, pronounced Necro—is a small HTTP proxy that sits between nix-daemon and your substituters. It is about three thousand 2 lines of clean, performant Rust code. It does three things:

  1. On a narinfo lookup, it races all configured upstreams in parallel with HEAD and remembers which one won.
  2. On a NAR fetch, it streams the body straight through to the client. No disk. No buffer. No NAR ever lives on the proxy.
  3. It keeps a small, bounded SQLite table of route decisions so a restart doesn’t force it to relearn the entire world.

That is the whole product. It is deliberately not another ncps, which mirrors caches to disk and gives you all the cache-invalidation grief that comes with mirroring caches to disk. ncro does not retain payload data once it is streamed through.

What’s Actually In It

The interesting parts are the ones the architecture diagram (which you might or might not have paid attention to) doesn’t show:

  1. The race. ncro’s router groups candidates by priority, then for each priority tier spawns a FuturesUnordered of HEAD requests and breaks on the first success. The tier loop is what lets you say “prefer my private cache, but only if it answers—otherwise fall through to cache.nixos.org” without writing any of that logic into Nix itself. There’s a deadline pinned to the select loop so a single hung upstream can’t stall the entire lookup. Failures are classified as not found, network error, and timeout because “every upstream returned 404” and “every upstream’s TCP handshake died” deserve different answers to the client.

  2. The cache, in two layers. A moka Cache sits in front of SQLite, with a 1024-entry capacity and TTL bound to the route TTL. SQLite underneath, with narinfo_bytes stored alongside the route so a hot path doesn’t even need a second upstream fetch. Eviction is throttled to fire every hundred writes by abusing AtomicU64::fetch_add’s pre-increment semantics. This is a detail that bit me in review because count % 100 == 0 fires on the first write, when the counter is zero. The fix was one character, but the impact was real: latency metrics were skewed until this edge case was corrected.

  3. Health, with EMA. , . The first sample bypasses the smoothing. Otherwise, the first probe permanently anchors to whatever junk was in ema_latency at startup. Consecutive failures bin the upstream into Active / Degraded / Down with multiplicative backoff (x1, x4, x10) so a dead cache stops costing you probe traffic. Sorting falls back to priority only when two upstreams are within 10% of each other, which is the only honest interpretation of “ties” when you’re dealing with network latency that jitters by milliseconds.

  4. Inflight dedup. If two clients ask for the same narinfo at the same time, only one race happens; the other waits on the mutex and reads the LRU on the way out. The cleanup path uses remove_if with Arc::ptr_eq because a naïve remove will happily delete someone else’s goddamned Arc that landed in the slot between the time your guard read it and the time it dropped. This avoids a TOCTOU class bug in the dedup map.

  5. Signatures. ed25519 via ed25519-dalek, public keys parsed from Nix’s name:base64(key) format, the fingerprint reconstructed in the exact 1;StorePath;NarHash;NarSize;refs shape that nix store sign writes. If you configure a key for an upstream, narinfos that don’t verify are rejected before they hit the cache. This is the part you cannot skip: a proxy that strips signature checks is a proxy that turns one compromised upstream into every machine pulling through you.

What I Did Not Build

There is no NAR cache. There is no mirror. There is no “warm-up” job. ncro will happily serve the same 200 MB closure ten times in a row by streaming it from upstream ten times, and that is the correct behaviour for a router. The optimization is in which upstream you ask, not in avoiding the upstream entirely. The moment you start storing NARs you are signing up for free-disk-space alerts, content-addressed-but-not-really cache poisoning, and a maintenance surface that has nothing to do with what you actually wanted, which was for nix build to stop spending 400 ms on DNS for a cache that never has the path. There is also no mesh by default. There is an optional gossip layer with signed UDP packets for the trusted-peer case, but it is off, and it should be off unless you have a clear trust and threat model for peer exchange. I plan to expand upon this in the future.

Was It Worth Writing?

Yes.

The proxy is small enough that I can keep the whole thing in my head, which I cannot say about most software I actually depend on. It solves one problem, the problem it set out to solve, and it does not try to grow into a content cache or a CI artifact store or a peer-to-peer Nix mesh—even though all three are tempting and each would complicate the design substantially. Nix’s substituter logic could support this kind of behavior, but in practice it has remained a static ordered loop for years. A dedicated proxy is a practical way to add dynamic routing without patching the daemon itself.

Perhaps it interests you too. Give it a try!

Footnotes

  1. Cough nh cough. It’s useful, it’s large, and it’s annoying to build if you’re building from the master branch because Rust is just like that. I can also list Hyprland in the things I fetch from its own flake and hope the cache has it working. What? I like my software fresh. ↩

  2. A bit more than that now, but the tool itself is feature-complete and functional. Not sure if I’ll stop tinkering with the tool, but the scope is well-defined by itself. ↩