Skip to main content

Goals

The goals of smelt are to provide:

  • Expressive test generation
  • Easy failure reproduction
  • Simplicity -- users should be able to run a testlist interactively without knowledge of smelt internals
  • Modular runtimes -- running locally vs on a slurm should feel the same
  • Expose coarse grained parallelism

Non-goals​

  • Being a build system
  • Sacrificing simplicity for performance

Roadmap

Smelt is still a new tool -- we're actively adding features. Priority for feature work is as follows:

  1. Slurm executor -- execute commands on slurm the same way you would locally
  2. Passing arguments to procedural test lists -- this would allow end users to pass "seeds" or other such input to testlists, and would help with the ergonomics of random test generation
  3. Command Replay from Invocation object -- invocation messages should contain all of the information to replay a command
  4. Restructure command interface -- currently we have this as two mirrored structures in Python and rust, it would be best to move this to protobuf
  5. Support for async test running -- workflows can be very timeconsuming, especially in EDA
  6. REAPI Remote Caching of tests -- Add optional fingerprinting and remote caching for commands that declare inputs and outputs
  7. Starlark rules -- Add a starlark interface to generate commands