about summary refs log tree commit diff
path: root/web/tvl/blog/2024-02-tvix-update.md

We've now been working on our rewrite of Nix, Tvix, for a little more than two years.

Our last written update was in September 2023, and although we did publish a couple of things in the meantime (flokli's talk on Tvix at NixCon 2023, our interview at the Nix Developer Dialogues, or tazjin's talk on tvix-eval (in Russian)), we never found the time to write something down.

In the meantime a lot of stuff has happened though, so it's time to change that :-)

Note: This blog post is intended for a technical audience that is already intimately familiar with Nix, and knows what things like derivations or store paths are. If you're new to Nix, this will not make a lot of sense to you!

Evaluation regression testing

Most of the evaluator work has been driven by evaluating nixpkgs, and ensuring that we produce the same derivations, and that their build results end up in the same store paths.

Builds are not hooked up all the way to the evaluator yet, but for Nix code without IFD (such as nixpkgs!) we can verify this property without building. An evaluated Nix derivation's outPath (and drvPath) can be compared with what C++ Nix produces for the same code, to determine whether we evaluated the package (and all of its dependencies!) correctly 1.

We added integration tests in CI that ensure that the paths we calculate match C++ Nix, and are successfully evaluating fairly complicated expressions in them. For example, we test against the Firefox derivation, which exercises some of the more hairy bits in nixpkgs (like WASM cross-compilation infrastructure). Yay!

Although we're avoiding fine-grained optimization until we're sure Tvix evaluates all of nixpkgs correctly, we still want to have an idea about evaluation performance and how our work affects it over time.

For this we extended our benchmark suite and integrated it with Windtunnel, which now regularly runs benchmarks and provides a view into how the timings change from commit to commit.

In the future, we plan to run this as a part of code review, before changes are applied to our canonical branch, to provide this as an additional signal to authors and reviewers without having to run the benchmarks manually.

ATerms, output path calculation, and builtins.derivation

We've implemented all of these features, which comprise the components needed to construct derivations in the Nix language, and to allow us to perform the path comparisons we mentioned before.

As an interesting side note, in C++ Nix builtins.derivation is not actually a builtin! It is a piece of bundled Nix code, that massages some parameters and then calls the actual builtin: derivationStrict. We've decided to keep this setup, and implemented support in Tvix to have builtins defined in .nix source code.

These builtins return attribute sets with the previously mentioned outPath and drvPath fields. Implementing them correctly meant that we needed to implement output path calculation exactly the same way as Nix does (bit-by-bit).

Very little of how this output path calculation works is documented anywhere in C++ Nix. It uses a subset of ATerm internally, produces "fingerprints" containing hashes of these ATerms, which are then hashed again. The intermediate hashes are not printed out anywhere (except if you patch Nix to do so).

We already did parts of this correctly while starting this work on go-nix some while ago, but found some more edge cases and ultimately came up with a nicer interface for Tvix.

All the Derivation internal data model, ATerm serialization and output path calculation have been sliced out into a more general-purpose nix-compat crate, alongside with more documentation unit tests and a Derivation ATerm parser, so hopefully this will now be more accessible for everyone now.

Note our builtin does not yet persist the Derivation anywhere "on disk" (though we have a debug CL that does write it to a temporary directory, in case we want to track down differences).

tvix-[ca]store

Tvix now has a store implementation!

The Nix model

Inside Nix, store path contents are normally hashed and communicated in NAR format, which is very coarse and often wasteful - a single bit of change in one file in a large store path causes a new NAR file to be uploaded to the binary cache, which then needs to be downloaded.

Additionally, identifying everything by the SHA256 digest of its NAR representation makes Nix store paths very incompatible with other content-addressed systems, as it's a very Nix-specific format.

The more granular Tvix model

After experimenting with some concepts and ideas in Golang, mostly around how to improve binary cache performance2, both on-disk as well as over the network, we settled on a more granular, content-addressed and general-purpose format.

Internally, it behaves very similar to how git handles tree objects, except blobs are identified by their raw BLAKE3 digests rather than some custom encoding, and similarly, tree/directory objects use the BLAKE3 digest of its canonical protobuf serialization as identifiers.

This provides some immediate benefits:

  • We only need to keep the same data once, even if it's used across different store paths.
  • Transfers can be more granular and only need to fetch the data that's needed. Due to everything being content-addressed, it can be fetched from anything supporting BLAKE3 digests, immediately making it compatible with other P2P systems (IPFS blake3 blobs, …), or general-purpose content-addressed caches (bazel-remote).

There's a lot more details about the data model, certain decisions etc. in the docs.

Compatibility

We however still want to stay compatible with Nix, as in calculating "NAR-addressed" store paths the same, support substituting from regular Nix binary caches, as well as storing all the other additional metadata about store paths.

We accomplished this by splitting the two different concerns into two separate tvix-store and tvix-castore crates, with the former one holding all Nix-specific metadata and functionality, and the latter being a general-purpose content-addressed blob and filesystem tree storage system, which is usable in a lot of contexts outside of Tvix too. For example, if you want to use tvix-castore to write your own git alternative, or provide granular and authenticated access into large scientific datasets, you could!

Backends

In addition to a gRPC API and client bindings, there's support for local filesystem-based backends, as well as for sled, an embedded K/V database.

We're also currently working on a backend supporting most common object storages, as well as on more granular seeking and content-defined chunking for blobs.

FUSE/virtiofs

A tvix-store can be mounted via FUSE, or exposed through virtiofs3. While doing the obvious thing - allowing mounting and browsing the contents of the store, this will allow lazy substitution of builds on remote builders, be in containerized or virtualized workloads.

We have an example in the repository seeding gnu hello into a throwaway store, then booting a MicroVM and executing it.

nar-bridge, bridging binary caches

nar-bridge and the NixHTTPPathInfoService bridge tvix-[ca]store with existing Nix binary caches and Nix.

The former exposes a tvix-[ca]store over the common Nix HTTP Binary Cache interface (both read and write).

The latter allows Tvix to substitute from regular Nix HTTP Binary caches, unpacking NARs and ingesting them on-the-fly into the castore model. The necessary parsers for NARInfo, signatures etc are also available in the nix-compat crate.

EvalIO / builtins interacting with the store more closely

tvix-eval itself is designed to be quite pure when it comes to IO - it doesn't do any IO directly on its own, but for the very little IO functionality it does as part of "basic interaction with paths" (like importing other .nix files), it goes through an EvalIO interface, which is provided to the Evaluator struct on instantiation.

This allows us to be a bit more flexible with how IO looks like in practice, which becomes interesting for specific store implementations that might not expose a POSIX filesystem directly, or targets where we don't have a filesystem at all (like WASM).

Using the EvalIO trait also lets tvix-eval avoid becoming too strongly coupled to a specific store implementation, hashing scheme etc4. As we can extend the set of builtins available to the evaluator with "foreign builtins", these can live in other crates.

Following this pattern, we started implementing some of the "basic" builtins that deal with path access in tvix-eval, like:

  • builtins.pathExists
  • builtins.readFile

We also recently started working on more complicated builtins like builtins.filterSource and builtins.path, which are also used in nixpkgs.

Both import a path into the store, and allow passing a Nix expression that's used as a filter function for each path. builtins.path can also ensuring the imported contents match a certain hash.

This required the builtin to interact with the store and evaluator in a very tight fashion, as the filter function (written in Nix) needs to be repeatedly executed for each path, and its return value is able to cause the store to skip over certain paths (which it previously couldn't).

Getting the abstractions right there required some back-and-forth, but the remaining changes should land quite soon.

Catchables / tryEval

Nix has a limited exception system for dealing with user-generated errors: builtins.tryEval can be used to detect if an expression fails (if builtins.throw or assert are used to generate it). This feature requires extra support in any Nix implementation, as errors may not necessarily cause the Nix program to abort.

The C++ Nix implementation reuses the C++ language-provided Exception system for builtins.tryEval which Tvix can't (even if Rust had an equivalent system):

In C++ Nix the runtime representation of the program in execution corresponds to the Nix expression tree of the relevant source files. This means that an exception raised in C++ code will automatically bubble up correctly since the C++ and Nix call stacks are equivalent to each other.

Tvix compiles the Nix expressions to a byte code program which may be mutated by extra optimization rules (for example, we hope to eliminate as many thunks as possible in the future). This means that such a correspondence between the state of the runtime and the original Nix code is not guaranteed.

Previously, builtins.tryEval (which is implemented in Rust and can access VM internals) just allowed the VM to recover from certain kinds of errors. This proved to be insufficient as it blew up as soon as a builtins.tryEval-ed thunk is forced again – extra bookkeeping was needed. As a solution, we now store recoverable errors as a separate runtime value type.

As you can imagine, storing evaluation failures as "normal" values quickly leads to all sorts of bugs because most VM/builtins code is written with only ordinary values like attribute sets, strings etc. in mind.

While ironing those out, we made sure to supplement those fixes with as many test cases for builtins.tryEval as possible. This will hopefully prevent any regressions if or rather when we touch this system again. We already have some ideas for replacing the Catchable value type with a cleaner representation, but first we want to pin down all the unspoken behaviour.

String contexts

For a long time, we had the working theory that we could get away with not implementing string contexts, and instead do reference scanning on a set of "known paths" (and not implement builtins.unsafeDiscardStringContext).

Unfortunately, we discovered that while this is conceptually true, due to a bug in Nix that's worked around in the stdenv.mkDerivation implementation, we can't currently do this and calculate the same hashes.

Because hash compatibility is important for us at this point, we bit the bullet and added support for string contexts into our NixString implementation, implemented the context-related builtins, and added more unit tests that verify string context behaviour of various builtins.

Strings as byte strings

C++ Nix uses C-style zero-terminated strings internally - however, until recently, Tvix has used standard Rust strings for string values. Since those are required to be valid UTF-8, we haven't been able to properly represent all the string values that Nix supports.

We recently converted our internal representation to byte strings, which allows us to treat a Vec<u8> as a "string-like" value.

JSON/TOML/XML

We added support for the toJSON, toXML, fromJSON and fromTOML builtins.

toXML is particularly exciting, as it's the only format that allows expressing (partially applied) functions. It's also used in some of Nix' own test suite, so we can now include these in our unit test suite (and pass, yay!).

Builder protocol, drv->builder

We've been working on the builder protocol, and Tvix's internal build representation.

Nix uses derivations (encoded in ATerm) as nodes in its build graph, but it refers to other store paths used in that build by these store paths only. As mentioned before, store paths only address the inputs - and not the content.

This poses a big problem in Nix as soon as builds are scheduled on remote builders: There is no guarantee that files at the same store path on the remote builder actually have the same contents as on the machine orchestrating the build. If a package is not binary reproducible, this can lead to so-called frankenbuilds.

This also introduces a dependency on the state that's present on the remote builder machine: Whatever is in its store and matches the paths will be used, even if it was maliciously placed there.

To eliminate this hermiticity problem and increase the integrity of builds, we've decided to use content-addressing in the builder protocol.

We're currently hacking on this at Thaigersprint and might have some more news to share soon!


That's it for now, try out Tvix and hit us up on IRC or on our mailing list if you run into any snags, or have any questions.

เจอกันนะ :)

  1. We know that we calculated all dependencies correctly because of how their hashes are included in the hashes of their dependents, and so on. More on path calculation and input-addressed paths in the next section!

  2. See nix-casync for one example - investing content-defined chunking (while still keeping the NAR format)

  3. Strictly speaking, not limited to tvix-store - literally anything providing a listing into tvix-castore nodes.

  4. That's the same reason why builtins.derivation[Strict] also lives in tvix-glue, not in tvix-eval.