Datum

Akka Streams, in Rust

The same Source → Flow → Sink vocabulary, GraphDSL junctions, cycles, substreams, and blueprint-vs-materialization contract — with Rust-native names and Result-based errors.

FS2/ZIO-class concurrency primitives

Signal, Subscription, Channel, and Topic — stream-native state cells, MPSC handoff, and pub-sub with overflow policies, faster than the libraries that inspired them.

Measured, not promised

Every operator area ships with a head-to-head harness against warmed Akka — and the primitives against FS2 and ZIO too — with a CPU column so busy-spinning can't hide.

No unsafe, anywhere

#![forbid(unsafe_code)] across the workspace. Lock-free hot paths come from audited safe dependencies: crossbeam and arc-swap.

Tokio-native networking

datum-net ships TLS, UDP, QUIC, and remote StreamRefs as single-owner async carriers — no blocking seam, flat thread counts at 1,024 concurrent streams.

Docs that compile

Every code example on this site is imported from a cargo test–verified snippet. If it's on a page, it builds and passes.

Faster than the best JVM competitor on every benchmarked row

Wall-clock and whole-process CPU, same host, warmed JVMs. Each figure below is the headline row of a checked-in result table — Datum versus the best of Akka, FS2, and ZIO for that scenario.

9.9×Remote StreamRefsvs Akka Artery-TCP, forced-remote · 14.8× less CPU

83×Signal propagation, 1,024 subscribersvs FS2 SignallingRef · 145× less CPU

23×Topic fan-out, 1,024 subscribersvs ZIO Hub · 158× less CPU

7.5×Channel MPSC, 1,024 producersvs Akka MergeHub · 20.5× less CPU

13–46×Fused graph chainsvs warmed Akka · typed executor, near-zero allocation

0Unsafe blocks#![forbid(unsafe_code)] — enforced by the compiler, not policy

Measured at v0.9.0 (2026-07-03) · full result tables

Where the time goes

Area	Headline scenario	Wall	CPU	Record
Concurrency primitives	Topic fan-out ×1024 vs ZIO Hub	23× faster	158× less	table
Remote StreamRefs	1,024 elements, fold-to-sum vs Akka Artery-TCP	9.9× faster	14.8× less	table
Graph & junctions	Fused identity/map chains vs Akka	13–46× faster	tracks wall	table
Queues	Source-queue backpressure vs Akka	985× faster	112× less alloc	table
Network carriers	TLS echo round-trip vs Akka · RSS 8 MB vs 799 MB	3.4× faster	8.3× less	table
Substreams	prefix_and_tail / flat_map_concat vs Akka	10–12× faster	29–43× less alloc	table
Source / Flow	Sync operator chains vs Akka	2.3–5.6× faster	4–36× less alloc	table

Not every row is a blowout — bounded-queue offer sits at ~parity (0.96×) and JSON framing at 1.00×, and the tables say so. The bar is parity-or-better on wall and CPU, everywhere.

Benchmark discipline is part of the library

The CPU column, always

A path can win wall-clock by busy-spinning while the JVM parks. Every table reports whole-process CPU so that cost can't hide.

Correctness counters

Every scenario asserts the work actually happened — element counts, per-subscriber totals, final values — on both sides of the comparison.

Frozen baselines first

JVM target numbers are captured and committed before the Datum implementation exists, then never touched.

Root-cause every miss

A losing row is never shipped as "inherent cost". It gets a named lever or a fix — the record keeps both the misses and the levers.

What Datum is — and deliberately isn't

It is

An Akka/Pekko Streams mirror — API shape and behavior, so Scala teams feel at home in a week, not a quarter.
A concurrency toolkit — Signal, Subscription, Channel, Topic for state, handoff, and pub-sub between streams, actors, and plain code.
A foundation layer — datum-core plus satellites for networking, jobs, and now cluster membership & sharding (datum-net, datum-agent, datum-cluster); built to carry higher-level frameworks on top.

It is not

A Reactive Streams bridge — no Publisher/Subscriber interop; the integration surfaces are Tokio futures and Ractor actors.
A line-by-line port — where Rust conventions differ, Rust wins: filter_map over collect, Result over exceptions, explicit overflow policies over defaults.
Benchmarketing — the harness measures both sides fairly or the number doesn't ship.

DatumThe streaming layer that outruns the JVM