Stargate development benchmark

This is a standalone Linux Rust 2024 workspace. Spark remains the load generator. The CLI owns plans, workload generation, detached execution in a Spark Pod, acceptance, reconciliation, and offline summaries. Deployment provisioning stays in ../scripts/.

Build and test

Run from this directory. rust-toolchain.toml pins the toolchain:

cargo build --locked --release
cargo fmt --check
cargo test --locked --all-targets
cargo clippy --locked --all-targets -- -D warnings

Tests use a native stand-in for kubectl. They need Linux process APIs, Bash, coreutils, setsid, flock, grep, and tar. The test support compiles the stand-in with the rustc on PATH; through rustup, that is the pinned toolchain. Tests do not require cluster access. CLI fixtures retain their own executable copy for resume. The repository's .github/workflows/build-test.yml runs these Rust checks.

Meaningful changes must also satisfy the Stargate mock integration and performance gates. The kubectl stand-in does not satisfy the runtime integration gate. Run a campaign against real Stargate with mock inference backends and retain accepted artifacts. Measurement or workload changes need a fixed workload and independent raw measurements so changed reporting cannot hide a performance regression.

Local end-to-end check

To run the CLI against real clusters without AWS, bring up the local-a and local-b k3d regions from README.md, then run from this directory:

cargo run --locked --release -- verify --region local-a --peer-region local-b

cluster.rs embeds each environment file with include_str!. When you add a region file under ../environments/, add it there as well. The helmfile reads every file in that directory as an environment.

Contracts

• Keep suite parsing and workload recipes deterministic. Preserve checked numeric bounds, workload ordering, and disjoint cold-capacity prompt ranges.
• Keep effective inputs immutable. Resume must validate them and reconcile owned work before checking current controls or starting traffic.
• A lost acknowledgement is not evidence that a launch failed. Never relaunch it.
• Require process ownership and stop proof for cleanup. Leave an uncertain stop blocking rather than marking it complete.
• Publish acceptance only after every required report and environment check. Preserve raw evidence; rendering reads accepted artifacts and launches no work.
• Keep credentials out of arguments, logs, and persisted command records.

Use artifact.rs for shared atomic JSON, hashing, and managed-path operations. Keep external effects in campaign.rs, cluster.rs, pod.rs, and process.rs. Follow the existing typed plan structure. The crate forbids unsafe Rust and denies Clippy warnings.

../loadtest/suite.yaml is the canonical input. ../loadtest/reduced.yaml selects the reduced pairs. Operator commands and artifact semantics are documented in README.md. Update that section when the CLI or acceptance contract changes.