Datum

🪼 - Just "Stream" anything :]

Datum is a Rust stream processing library scaffolded around Ractor actors and a push-based stream abstraction. The compatibility target is Akka/Pekko Streams Typed API shape and behavior, with Rust-native ownership, async, and benchmarking constraints.

Install

Add Datum to your project from crates.io:

cargo add datum-core

Or in Cargo.toml:

[dependencies]
datum-core = "0.9"

Note: The package is published as datum-core (the name datum is taken on crates.io), but the import path stays use datum::… — the crate's library name is datum.

To track an unreleased commit instead:

[dependencies]
datum-core = { git = "https://github.com/Aethergrids/Datum", tag = "v0.9.0" }

Upstream References

Ractor crate: 0.15.13
Optional Ractor cluster crate: 0.15.13, enabled with --features cluster
Akka source submodule: third_party/akka, tracking upstream main and pinned in this repository at commit 58f1f6db2e505e87f5dc115ee9476833872e7ae0
Latest stable Akka tag observed during setup: v2.10.18

Feature highlights (v0.9.0)

Concurrency primitives (datum::concurrent) — Signal<T> (latest-value state cell with a lock-free get() and a coalesced change feed), Subscription<T> (the same state cell with a lossless every-change feed), Channel<T> (closeable bounded MPSC), and Topic<T> (pub-sub broadcast with backpressure/sliding/dropping overflow policies). Benchmarked 4-way against FS2, ZIO, and Akka: every row beats the best JVM competitor on both wall-clock and CPU.
API ergonomics (additive, zero semantic change) — use datum::prelude::*;, terminal shortcuts (run_fold/run_foreach/run_reduce), (0..n).into_source() / Source::from(vec), the graph wiring DSL builder.wire(a.to(&b)) (with connect() still first-class), and idiomatic try_* fallible operators (try_map, try_filter, … — *_result names deprecated).
Complete Akka Streams surface — the linear DSL, GraphDSL with typed junctions and cycles, substreams, dynamic streams (kill switches, hubs), StreamRefs (same-process splice + remote transport via datum-net), streaming IO, and IO adapters bridging to std::io::Read/Write.
#![forbid(unsafe_code)] on the datum crate itself; lock-free internals come from safe deps (crossbeam-queue, arc-swap).

Development

cargo test
cargo check --benches
cargo bench --bench push_baseline
cargo bench --bench source_flow
cargo bench --bench materialization
cargo bench --bench graph
cargo bench --bench actor_ask
benches/actor_ask_compare/run.sh

use datum::prelude::*;

let sum = (0_u64..1_000)
    .into_source()
    .map(|item| item + 1)
    .filter(|item| item % 2 == 0)
    .run_fold(0_u64, |acc, item| acc + item)
    .unwrap()
    .wait()
    .unwrap();

assert_eq!(sum, 250_500);

Benchmarks

Datum is benchmarked head-to-head against warmed Akka/Pekko Streams across eight core areas — Source/Flow, materialization, graph/junctions, actor ask, dynamic streams, streaming IO, substreams, and queues — plus network carriers (datum-net) and a 4-way concurrency-primitives comparison against FS2, ZIO, and Akka. Honest per-path numbers (including any at-parity rows) are in the result tables under roadmap/benchmarks/. The harness adds a Datum CPU us/op column on purpose: some wins come partly from busy-spinning while a JVM parks — a real cost the wall-clock number hides.

Qualitative summary of the current state:

Fused linear path: 13–46x Akka (typed plan; all common sink shapes).
Junctions: 2.3–340x; typed kernels cover all major shapes; graph cycles ~29x.
Concurrency primitives (vs best of FS2/ZIO/Akka per row): Channel MPSC ×1024 7.5x wall / 20.5x CPU; Signal reads ~ns-class (3.6–7.8x ZIO); Signal propagation ×1024 83x; Subscription lossless ×256 43x; Topic fanout ×1024 23x.
Remote StreamRefs (forced-remote Akka Artery-TCP, fair): ~9.9x faster wall, ~14.8x lower CPU, ~35x lower allocation; same-process splice ~6.5x.
Dynamic hubs: MergeHub p16 ~16x (p1 ~0.46x — core materialization floor, documented); BroadcastHub 1.3–2x; PartitionHub 1.4–2.7x.
Queues: source-queue backpressure ~985x; bounded queue ~parity wall with ~112x less allocation.
Actor ask: ~2.4–3x at p2–p4; allocation dominated by upstream Ractor boxing (~848 B/element, upstream-blocked).
Streaming IO adapters: ~7.4x on the round-trip scenario; materialization: at parity.

Run a comparison (requires a JDK + sbt); each runner writes a rendered table under target/<area>-comparison/:

benches/source_flow_compare/run.sh
benches/materialization_compare/run.sh
benches/graph_compare/run.sh
benches/actor_ask_compare/run.sh
benches/dynamic_streams_compare/run.sh
benches/streaming_io_compare/run.sh
benches/substreams_compare/run.sh

Current result tables, the per-operator coverage matrix, and apples-to-apples caveats:

roadmap/benchmarks/source-flow.md
roadmap/benchmarks/materialization.md
roadmap/benchmarks/graph.md
roadmap/benchmarks/actor-ask.md
roadmap/benchmarks/dynamic-streams.md
roadmap/benchmarks/streaming-io.md
roadmap/benchmarks/substreams.md
roadmap/M1-v0.1.0-foundation.md — coverage matrix, optimization status & apples-to-apples caveats
roadmap/M4-v0.4.0-completeness-hardening.md — M4 work packages and exit criteria