scientific-workflow

scientific-workflow provides Rust primitives for representing scientific system states and building reproducible simulation workflows.

The crate provides SystemState, a fixed-layout heterogeneous state container, and StateSeries, an ordered growable collection of complete states for in-memory analysis. Concrete payloads move through both layers without cloning, making them suitable for large arrays and tensors.

Features

• Standard config/{fixed,sweep,paths,state}.json scientific-project definition.
• Deterministic Cartesian and correlated explicit-case task expansion.
• Lazy complete TaskConfig handles combining parameters and shared paths.
• Exact sweep-value filtering and ambiguity-safe unique task selection.
• Clone-free dict-like resolved task views over shared JSON values.
• Named project-root-relative path resolution and byte-exact source export.
• JSON-defined state fields with deterministic order and optional descriptions.
• Dictionary-like typed access to heterogeneous Rust payloads.
• Coordinated immutable and mutable tuple borrowing for coupled kernels.
• Assembly-established field types retained across extraction and blank-state derivation.
• Clone-free payload insertion, in-place mutation, and owned extraction.
• Explicit deep cloning of complete states.
• Shared immutable state specifications.
• Integer and optional finite physical time coordinates.
• Strict template validation and semantic JSON round trips.
• Compatibility with owned scientific payloads such as physics_in_parallel tensors.
• Ordered state-series collection with strict shared-layout identity.
• Lightweight copyable series views and field-level analysis mutation.
• Borrowed JSON encoding without payload cloning.
• Writer-owned typed sampling intervals with no payload access for skipped states.
• Automatic exactly-once final-state sampling across sampling-interval boundaries.
• Exact finite-f64 JSON reconstruction through Serde JSON's round-trip parser.
• Finite byte- and record-bounded asynchronous writers.
• Exact-byte automatic chunking with indivisible JSONL records.
• Durable chunk publication through open-file sync, incremental descriptor preparation, atomic lifecycle rename, and stream-directory sync.
• Explicit interrupted-run append and complete typed checkpoint recovery.
• SHA-256-verified eager reconstruction through per-key payload decoders.
• Efficient latest-state reconstruction without loading earlier chunks.
• Automatic UTC lifecycle timestamps and monotonic active durations.
• Collision-resistant generated or caller-named execution scopes.
• Structurally separate terminal metadata and immutable completed-recording handles.
• Parameter-identified, parallel-safe centralized progress reporting.
• One exclusive terminal renderer with interactive, CI, and hidden modes.

Parallel Progress Reporting

ProgressReporter derives human-facing identity from task parameters and uses the automatically assigned task ordinal only for stable ordering. With no explicit identity selection, all sweep keys form the identity. Applications may choose any smaller parameter combination that remains unique:

use scientific_workflow::prelude::*;

# fn main() -> Result<(), Box<dyn std::error::Error>> {
let project = ScientificProject::load("project-root")?;
let reporter = ProgressReporter::for_project(&project)
    .identify_tasks_by(["temperature", "seed"])
    .start()?;

for task in project.task_configs() {
    let progress = reporter.start_task(&task, 0, Some(1_000))?;
    // After each successful scientific transition:
    progress.set_iteration(1_000)?;
    progress.complete()?;
}

let summary = reporter.complete("all scientific tasks completed")?;
assert!(summary.is_success());
# Ok(())
# }

Iteration updates are atomic and allocation-free. One renderer thread polls all tasks at a bounded frequency and is the only component permitted to write human-facing terminal output during the session. Interactive stderr is cleared once at renderer startup and then receives one row for every configured task, including tasks still waiting for a worker. Known-target rows show elapsed task execution time and ETA; redirected stderr receives line-oriented lifecycle events without being cleared. Dropping an unfinished TaskProgress marks that task failed.

Progress is not scientific state. Callers set it from the authoritative SystemState::simulation_time() after a successful transition. Known targets use absolute iterations, while None supports convergence-driven or otherwise open-ended work.

Mandatory Chunk Integrity

Every sealed JSONL chunk is described by an exact byte count and SHA-256 digest in metadata.json. Verification is mandatory whenever a chunk is validated or reconstructed. The public reader has no unchecked mode, checksum opt-out, feature switch, or performance flag: corruption produces StorageError rather than partially trusted scientific data.

Parsing alone is not validation. Skipping a chunk because an operation does not need its contents is permitted, but that chunk is then unexamined—not verified. Any chunk actually used to reconstruct scientific state must cross the checksum boundary first. This integrity guarantee detects accidental corruption; it is not a substitute for provenance, signatures, or validation of the scientific model itself.

The public SystemStateWriter facade owns multi-stream metadata, one bounded queue and worker, and the recording's completion or failure lifecycle. Workflow dispatch remains a later development stage.

Installation

Add the crate to a Rust project:

[dependencies]
scientific-workflow = "0.1"

The crate uses Rust edition 2024 and requires Rust 1.85 or newer.

Complete Project Example

The source repository includes examples/attractor_2d, a standalone downstream application that exercises configuration loading, Cartesian task expansion, directly owned mutable states, tuple payload borrowing, independent sample streams, bounded asynchronous recording, automatic chunking, and explicit completion. Its lazy TaskConfig iterator feeds Rayon's bounded work-stealing pool, while stable task indices keep recording paths deterministic regardless of completion order. It then reads the complete checkpoint's latest state with typed payload decoders and verifies the final live-to-stored round trip exactly. From the repository root, run:

cargo run --manifest-path examples/attractor_2d/Cargo.toml

The example is intentionally outside this crate directory and therefore is not part of the crates.io package. Its generated recordings remain under its ignored target/recordings directory.

Project Configuration

A standard scientific project keeps four files together:

project-root/
└── config/
    ├── fixed.json
    ├── sweep.json
    ├── paths.json
    └── state.json

fixed.json contains values shared by every task:

{
  "physical_time_increment": 0.125,
  "lattice_shape": [4, 8]
}

sweep.json supports ordered Cartesian axes:

{
  "mode": "cartesian",
  "axes": [
    {"name": "temperature", "values": [280.0, 300.0]},
    {"name": "seed", "values": [7, 11, 13]}
  ]
}

or correlated explicit cases:

{
  "mode": "cases",
  "cases": [
    {"temperature": 280.0, "physical_time_increment": 0.1},
    {"temperature": 300.0, "physical_time_increment": 0.05}
  ]
}

paths.json contains shared path strings resolved relative to the project root:

{
  "input_data": "data/input.json",
  "output_root": "results"
}

Load the project and consume exact JSON names through each resolved task's read-only dictionary:

use scientific_workflow::prelude::*;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let project = ScientificProject::load("project-root")?;
    for task in project.task_configs() {
        let physical_time_increment =
            task.decode_value::<f64>("physical_time_increment")?;
        let temperature = task.decode_value::<f64>("temperature")?;
        let seed = task.decode_value::<u64>("seed")?;
        let output_root = task.resolve_path("output_root")?;
        println!(
            "task={} dt={physical_time_increment} temperature={temperature} seed={seed} output={}",
            task.task_ordinal(),
            output_root.display()
        );
    }
    Ok(())
}

task_configs() lazily emits the complete Cartesian product—or exactly the declared correlated cases—in stable task-ordinal order. Each item is a cheap owned handle over shared fixed, sweep, and path storage, so it can move into a worker queue without cloning merged JSON dictionaries:

# use scientific_workflow::prelude::*;
# fn submit(_: TaskConfig) -> Result<(), Box<dyn std::error::Error>> { Ok(()) }
# fn main() -> Result<(), Box<dyn std::error::Error>> {
let project = ScientificProject::load("project-root")?;

for task in project.task_configs_matching("temperature", 300.0)? {
    submit(task)?;
}
# Ok(())
# }

Matching constrains only the named sweep dimension; every combination of the remaining axes is retained. Unique selection returns an error when no task or more than one task matches, rather than silently choosing the first. Fixed keys and path keys cannot be used as sweep selectors. Use unique_task_config_matching(key, value) only when that one sweep dimension is known to identify exactly one task, as is common for an explicit case ID.

Task handles share the parsed source allocations and do not clone values or construct merged maps. value and require_value borrow raw JSON; decode_value explicitly constructs one requested Rust value. The final sweep axis changes fastest. Fixed and swept names must be disjoint.

ProjectConfig::write_source_config(destination) reproduces the three parameter/path files byte for byte beneath a new destination project. It never overwrites an existing config/ directory. TaskParameters::to_json instead serializes one deterministic derived fixed-plus-sweep dictionary.

ScientificProject::load additionally requires config/state.json and exposes its shared schema through state_schema(). The lower-level ProjectConfig remains available when an application intentionally needs only parameter and path configuration.

State Template

A program begins with a JSON template that declares every state key and may document its payload in natural language:

{
  "fields": [
    {
      "name": "population",
      "description": "Population count at each modeled location"
    },
    {
      "name": "space"
    }
  ]
}

Field order defines the compact runtime slot order. The template contains no Rust type or storage codec information. The first payload inserted into a field establishes its concrete runtime type; that contract remains after take_payload or clear_payload and is copied into blank states derived with SystemState::clone_structure_without_payloads. Descriptions remain documentation only.

Basic Usage

use scientific_workflow::prelude::*;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let spec = SystemStateSchema::load_json_template("state.json")?;
    std::fs::create_dir_all("output")?;
    let mut state = spec.create_empty_state(SimulationTime::from_iteration(0));

    drop(state.insert_payload("population", vec![10_u64, 20, 30])?);

    state
        .payload_mut::<Vec<u64>>("population")?
        .push(40);

    let population = state.take_payload::<Vec<u64>>("population")?;
    assert_eq!(population, vec![10, 20, 30, 40]);
    assert!(state.has_no_payloads());

    Ok(())
}

insert_payload consumes the supplied payload, and take_payload returns that same owned payload. Neither operation calls Clone. Calling SystemState::clone creates a new erased box and calls Clone for every populated payload; the semantic depth is defined by each concrete type's Clone implementation.

Coupled Payload Access

Scientific kernels can borrow several distinct fields without payload copies, temporary extraction, locks, or application-side selector structures. Supply the expected concrete types and field names in matching tuple order:

use scientific_workflow::prelude::*;

# fn evolve(position: &mut Vec<f64>, velocity: &mut Vec<f64>) {
#     position[0] += velocity[0];
# }
# fn main() -> Result<(), Box<dyn std::error::Error>> {
let spec = SystemStateSchema::load_json_template("state.json")?;
let mut state = spec.create_empty_state(SimulationTime::from_iteration(0));
drop(state.insert_payload("position", vec![0.0_f64])?);
drop(state.insert_payload("velocity", vec![1.0_f64])?);

let (position, velocity) = state
    .borrow_payloads_mut::<(Vec<f64>, Vec<f64>)>(("position", "velocity"))?;
evolve(position, velocity);
# Ok(())
# }

Supported tuple arities are two through eight. The complete request is validated before any reference is returned, and repeating a field is rejected. Use payload or payload_mut for one field. Name lookup and type validation occur once per tuple borrow, so the returned references should normally surround the full kernel or simulation sweep.

In-Memory Time Series

StateSeries owns complete states for analysis. Appending validates that every state shares the series' exact specification allocation and that simulation indices increase strictly. Index gaps are allowed.

use scientific_workflow::prelude::*;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let spec = SystemStateSchema::load_json_template("state.json")?;
    let mut state = spec.create_empty_state(SimulationTime::from_iteration(0));
    drop(state.insert_payload("population", vec![10_u64, 20, 30])?);

    let mut series = StateSeries::new(spec);
    series.push_state(state)?;
    series
        .payload_mut_at::<Vec<u64>>(0, "population")?
        .push(40);

    let view = series.as_view();
    assert_eq!(view.len(), 1);
    Ok(())
}

The collection never returns &mut SystemState, because changing a stored state's time would invalidate ordering. payload_mut_at permits one typed payload mutation at a time. push_state, pop_state, and into_states move ownership without cloning. Explicit StateSeries::clone deep-clones all populated payloads; use as_view or Arc<StateSeries> for lightweight sharing.

StateSeries performs no serialization, chunking, queueing, or disk IO. Those responsibilities belong to the separate storage layer.

Tensor Payloads

Any concrete type satisfying Serialize + Clone + Send + 'static can be stored. For example, an application can use a dense physics_in_parallel tensor:

use physics_in_parallel::math::{Dense, Tensor};
use scientific_workflow::prelude::*;

let spec = SystemStateSchema::load_json_template("state.json")?;
let mut state = spec.create_empty_state(SimulationTime::from_iteration(0));

let mut population = Tensor::<u64, Dense>::zeros(&[3]);
population.set(&[0], 10);
population.set(&[1], 20);
population.set(&[2], 30);

drop(state.insert_payload("population", population)?);
let population = state.take_payload::<Tensor<u64, Dense>>("population")?;

The tensor crate is not a required runtime dependency of scientific-workflow; applications use their own concrete serializable scientific payload types without registering codecs.

The PiP 3.0.4 integration uses versioned Serde schemas for dense and sparse tensors, matrices, vector lists, square lattices, and heterogeneous PhysObj values. They reconstruct through the same generic registry path:

let decoders = JsonPayloadDecoderRegistry::new()
    .with_json_field::<Tensor<f64, Dense>>("population")?
    .with_json_field::<PhysObj>("particles")?;

Sparse PiP records contain only sorted nonzero indices and values; Scientific Workflow does not densify them during encoding or reconstruction.

Persistent State Recording

One import brings the complete supported state, analysis, storage, reader, and decoder API into scope:

use std::num::NonZeroU64;
use scientific_workflow::prelude::*;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let spec = SystemStateSchema::load_json_template("state.json")?;
    let mut writer = SystemStateWriter::builder("output/recording-001", &spec)
        .with_time_axis_metadata(
            TimeAxisMetadata::new("iteration")
                .with_iteration_unit("iteration")
                .with_physical_axis("physical_time", "s"),
        )
        .with_shared_stream_limits(
            NonZeroU64::new(64 * 1024 * 1024).unwrap(),
            NonZeroU64::new(256 * 1024 * 1024).unwrap(),
        )
        .add_sampled_state_stream(
            "signal",
            ["population"],
            SamplingInterval::iterations(1).unwrap(),
        )
        .create_new_recording()?;

    let mut state = spec.create_empty_state(
        SimulationTime::from_iteration_and_physical_time(0, 0.0).unwrap(),
    );
    drop(state.insert_payload("population", vec![10.0_f64, 20.0, 30.0])?);
    writer.observe_state(&state)?;
    writer.complete_recording_with_final_state(&state)?;

    let decoders = JsonPayloadDecoderRegistry::new()
        .with_json_field::<Vec<f64>>("population")?;
    let series = StoredStateSeriesReader::open_completed_recording("output/recording-001", decoders)?
        .read_stream_as_state_series("signal")?;
    assert_eq!(series.len(), 1);
    Ok(())
}

observe_state checks each stream's typed sampling interval before accessing any payload. Non-due streams perform no serialization or queue work. Due streams resolve each selected key once and borrow payloads only while producing owned encoded bytes, after which bounded blocking backpressure applies through the recording's single queue and worker. complete_recording_with_final_state records a non-aligned endpoint exactly once per stream before completion. Each chunk is synchronized, described in the sole metadata file, atomically renamed from .jsonl.tmp to .jsonl, and followed by a stream-directory sync. flush_stream_to_storage(stream) exposes this as an ordered durability barrier. continue_existing_recording recovers append position without reconstructing state, while continue_recording_from_latest_checkpoint also returns a complete typed checkpoint through registered decoders.

Custom Payload Decoders

Each decoder is registered for one exact state key and returns that key's concrete payload type. A closure is sufficient for stateless conversion; a named decoder can carry configuration or shared resources:

use scientific_workflow::prelude::*;
use serde::{Deserialize, Serialize};

#[derive(Clone, Deserialize, Serialize)]
struct ParticleBlock {
    positions: Vec<[f64; 3]>,
}

struct ParticleBlockDecoder;

impl JsonPayloadDecoder<ParticleBlock> for ParticleBlockDecoder {
    type Error = serde_json::Error;

    fn decode_json_payload(&self, raw_json: &str) -> Result<ParticleBlock, Self::Error> {
        serde_json::from_str(raw_json)
    }
}

fn configure() -> Result<JsonPayloadDecoderRegistry, StorageError> {
    let mut decoders = JsonPayloadDecoderRegistry::new();
    decoders.register_for_field("particles", ParticleBlockDecoder)?;
    decoders.register_for_field::<Vec<u64>, _>("counts", |raw_json: &str| {
        serde_json::from_str(raw_json)
    })?;
    Ok(decoders)
}

The reader performs record parsing and key lookup, passes only the matching raw JSON value to each decoder, and moves the returned payload into the reconstructed state. Custom decoders do not handle chunks, metadata, sibling fields, or state assembly.

Testing

From the package directory:

cargo test --all-targets --no-fail-fast --locked

The permanent suite contains seven logged integration workflows. Run each with --nocapture to display its stable semantic report.

Project configuration and task expansion:

cargo test --test configuration_workflow -- --nocapture

Simulation-owned state:

cargo test --test state_workflow -- --nocapture

In-memory analysis series:

cargo test --test analysis_workflow -- --nocapture

Successful storage and typed reconstruction:

cargo test --test storage_workflow -- --nocapture

Storage failure and corruption handling:

cargo test --test storage_resilience -- --nocapture

Interrupted-run recovery, checkpoint reconstruction, and append:

cargo test --test resume_workflow -- --nocapture

Doctests and lint gate:

cargo test --doc --locked
cargo clippy --all-targets --all-features --locked -- -D warnings

The repository-level tests.md documents complete method allocation, indirect private coverage, logging rules, and completion criteria.

License

Licensed under the MIT License.