about summary refs log tree commit diff
path: root/tvix/docs/src/TODO.md

TODO

This contains a rough collection of ideas on the TODO list, trying to keep track of it somewhere.

Of course, there's no guarantee these things will get addressed, but it helps dumping the backlog somewhere.

Feel free to add new ideas. Before picking something, ask in #tvix-dev to make sure noone is working on this, or has some specific design in mind already.

Cleanups

Nix language test suite

  • Think about how to merge, but "categorize" tvix_tests in glue and eval. We currently only have this split as they need a different feature set / builtins.
  • move some of the rstest cases in tvix-glue to the .nix/.exp mechanism. Some of them need test fixtures, which cannot be represented in git (special file types in the import tests for example). Needs some support from the test suite to create these fixtures on demand.
  • extend verify-lang-tests/default.nix mechanism to validate tvix-eval and tvix-glue test cases (or the common structure above).
  • absorb eval/tests/nix_oracle.rs into tvix_tests, or figure out why it's not possible (and document) it. It looks like it's only as nix is invoked with a different level of --strict, but the toplevel doc-comment suggests its generic?

Correctness > Performance

A lot of the Nix behaviour isn't well documented out, and before going too deep into performance optimizations, we need to ensure we properly grasped all hidden features. This is to avoid doing a lot of "overall architecture perf-related work" and increased code complexity based on a mental model that might get disproved later on, as we work towards correctness.

We do this by evaluating more and more parts of the official Nix test suite, as well as our own Tvix test suite, and compare it with Nix' output.

Additionally, we evaluate attributes from nixpkgs, compare calculated output paths (to determine equivalence of evaluated A-Terms) and fix differences as we encounter them.

This currently is a very manual and time-consuming process, both in terms of setup, as well as spotting the source of the differences (and "compensating" for the resulting diff noise on resulting mismtaches).

  • We could use some better tooling that periodically evaluates nixpkgs, and compares the output paths with the ones produced by Nix
  • We could use some better tooling that can spot the (real) differences between two (graphs of) derivations, while removing all resulting noise from the diff in resulting store paths.

Performance

Even while keeping in mind some of the above caveats, there's some obvious low-langing fruits that could have a good impact on performance, with somewhat limited risk of becoming obsolete in case of behaviorial changes due to correctness:

  • String Contexts currently do a lot of indirections (edef) (NixString -> NixStringInner -> HashSet[element] -> NixContextElement -> String -> data) to get to the actual data. We should improve this. There's various ideas, one of it is globally interning all Nix context elements, and only keeping indices into that. We might need to have different representations for small amount of context elements or larger ones, and need tooling to reason about the amount of contexts we have.
  • To calculate NAR size and digest (used for output path calculation of FODs), our current SimpleRenderer NarCalculationService sequentially asks for one blob after another (and internally these might consists out of multiple chunks too). That's a lot of roundtrips, adding up to a lot of useless waiting. While we cannot avoid having to feed all bytes sequentially through sha256, we already know what blobs to fetch and in which order. There should be a way to buffer some "amount of upcoming bytes" in memory, and not requesting these seqentially. This is somewhat the "spiritual counterpart" to our sequential ingestion code (ConcurrentBlobUploader, used by ingest_nar), which keeps "some amount of outgoing bytes" in memory. Our seekable NAR AsyncRead implementation already removes most complexity in rendering everything between blobs. It should be possible to extend / write a wrapped version of it that prefetches a configurable sliding window of blobs. Per-blob prefetching itself is somewhat blocked until the {Chunk/Blob}Service split is done, as then prefetching there would only be a matter of adding it into the one BlobReader.

Error cleanup

  • Currently, all services use tvix_castore::Error, which only has two kinds (invalid request, storage error), containing an (owned) string. This is quite primitive. We should have individual error types for BS, DS, PS. Maybe these should have some generics to still be able to carry errors from the underlying backend, similar to IngestionError. There was an attempt to give PS separate error types (cl/11695), but this ended up very verbose. Every error had to be boxed, and a possible additional message be added. Some errors that didn't wrap another underlying errors were hard to construct, too (requiring the addition of errors). All of this without even having added proper backtrace support, which would be quite helpful in store hierarchies. anyhow's .context() gives us most of this out of the box. Maybe we can use that, using enums rather than &'static str as context in some cases?

Fixes towards correctness

  • rnix only supports string source files, but NixString uses bytes (and Nix source code might be no valid UTF-8).

Documentation

Extend the other pages in here. Some ideas on what should be tackled:

  • Document what Tvix is, and what it is not yet. What it is now, what it is not (yet), explaining some of the architectural choices (castore, more hermetic Build repr), while still being compatible. Explain how it's possible to plug in other frontends, and use tvix-{[ca]store,build} without Nixlang even. And how nix-compat is a useful crate for all sorts of formats and data types of Nix.
  • Update the Architecture diagram to model the current state of things. There's no gRPC between Coordinator and Evaluator.
  • Add a dedicated section/page explaining the separation between tvix-glue and tvix-eval, and how more annoying builtins get injected into tvix-eval through tvix-glue. Maybe restructure to only explain the component structure potentially crossing process boundaries (those with gRPC), and make the rest more crate and trait-focused?
  • Restructure docs on castore vs store, this seems to be duplicated a bit and is probably still not too clear.
  • Describe store composition(s) in more detail. There's some notes on granular fetching which probably can be repurposed.
  • Absorb the rest of //tvix/website into this.

Features

Fetchers

Some more fetcher-related builtins need work:

  • fetchGit
  • fetchTree (hairy, seems there's no proper spec and the URL syntax seems subject to change/underdocumented)

Derivation -> Build

While we have some support for structuredAttrs and fetchClosure (at least enough to calculate output hashes, aka produce identical ATerm), the code populating the Build struct doesn't exist it yet.

Similarly, we also don't properly populate the build environment for fetchClosure yet. (Note there already is ExportedPathInfo, so once structuredAttrs is there this should be easy.

Builders

Once builds are proven to work with real-world builds, and the corner cases there are ruled out, adding other types of builders might be interesting.

  • bwrap
  • gVisor
  • Cloud Hypervisor (using similar technique as //tvix//boot).

Long-term, we want to extend traits and gRPC protocol. This requires some more designing. Some goals:

  • (more granular) control while a build is happening

  • expose more telemetry and logs

  • Add pre-flight checks in the OCI builder:

    • ensure fusermount suid binary exists
    • ensure allow_other is set
    • ensure runc exists in $PATH

Store composition

  • Combinators: list-by-priority, first-come-first-serve, cache
  • Store composition hierarchies (@yuka).
    • URL format too one-dimensional.
    • We want to have nice and simple user-facing substituter config, including sensible default wrappers for caching, retries, fallbacks, as well as granular control for power-users.
    • Current design idea:
      • Have a concept similar to rclone config (map with store aliases as keys, allowing to refer to stores by their alias from other parts of the config). It allows both referring to by name, as well as ad-hoc definition: https://rclone.org/docs/#syntax-of-remote-paths
      • Each store needs to be aware of its "instance name", so it can be included in logs, metrics, …
      • Have a "instantiation function" traversing such a config data structure, creating store instances and plugging them together, ultimately returning a dyn …Service interface.
      • No reconfiguration/reconcilation for now
      • Making URLs the primary data format would get ugly quite easily (hello multiple layers of escaping!), so best to convert the existing URL syntax to our new config format on the fly and then use one codepath to instantiate/assemble. Similarly, something like the "user-facing substituter config" mentioned above could aalso be converted to such a config format under the hood.
      • Maybe add a ?cache=$other_url parameter support to the URL syntax, to easily wrap a store with a caching frontend, using $other_url as the "near" store URL.
    • Each store should get its instance name passed down, and add this as a field in the instrumentation calls. This causes log messages and per-instance store metrics to be traceable back to the specific instance (if multiple backends of the same type are present).

Store Config

There's already serde for some store options (bigtable uses serde_qs). We might also have common options global over all backends, like chunking parameters for chunking blobservices. Think where this would fit in.

  • Rework the URL syntax for object_store. We should support the default s3/gcs URLs at least.

BlobService

  • On the trait side, currently there's no way to distinguish reading a known-chunk vs blob, so we might be calling .chunks() unnecessarily often. At least for the object_store backend, this might be a problem, causing a lot of round-trips. It also doesn't compose well - every implementation of BlobService needs to both solve the "holding metadata about chunking info" as well as "storing chunks" questions. Design idea (@flokli): split these two concerns into two separate traits:

    • a ChunkService dealing with retrieving individual chunks, by their content digests. Chunks are small enough to keep around in contiguous memory.
    • a BlobService storing metadata about blobs.

    Individual stores would not need to implement BlobReader anymore, but that could be a global thing with access to the whole store composition layer, which should make it easier to reuse chunks from other backends. Unclear if the write path should be structured the same way. At least for some backends, we want the remote end to be able to decide about chunking.

  • While object_store recently got support for Content-Type (https://github.com/apache/arrow-rs/pull/5650), there's no support on the local filesystem yet. We'd need to add support to this (through xattrs).

PathInfoService

  • sqlite backend (different schema than the Nix one, we need the root nodes data!)

Nix Daemon protocol

  • Some work ongoing on the worker operation parsing (griff, picnoir)

O11Y

generated by cgit-pink 1.4.1 (git 2.44.2) at 2024-11-21T21·59+0000