Isolated, metered serverless compute

Compute that picks the right shape.

Submit a bounded Python workload. Dyson estimates the ceiling, routes it to the smallest compatible CPU or NVIDIA L4 profile, isolates execution, and settles only measured compute.

Open Dyson console Read the API guide

Python + NumPy CPU PyTorch + CUDA Firebase or API key

dyson / route preview metered

workload.py

auto route

memory1.6 GiB

parallel2 workers

timeout300 sec

policy + estimate

selected route

Balanced CPU

NumPy and parallel CPU workloads

0.05 crbase rate / minute

execution profiles

0.02

credits/min from

30+ days

minimum artifact protection

1 min

billing minimum

The control loop

One request. Four guarded stages.

Routing, isolation, billing, and reconciliation are one server-owned transaction—not a chain of client-side promises.

Describe

Send source, JSON arguments, timeout, memory, parallelism, and GPU intent.

Estimate

Dyson picks the smallest compatible profile and returns a credit ceiling before launch.

Isolate

A credential-free Cloud Run Job receives only short-lived signed artifact URLs.

Settle

Started compute settles from complete Cloud Run timing; an exact no-start cancellation releases at zero duration, while completed and failed workloads require canonical output before capture.

Source-first SDK

From function to isolated job.

Source jobs pair Python code with JSON arguments for stable, inspectable payloads. Existing Python applications can also submit callable payloads with the SDK.

One shared CrossGL key across Dyson and Credora APIs
Idempotent submissions and bounded polling
Lifecycle protection is guaranteed for at least 30 elapsed days; deletion is eligible only at the age-31 UTC boundary

Follow the quickstart

source-job.py

import dyson

SOURCE = """
def workload(values):
    return {
        "count": len(values),
        "sum": sum(values),
    }
"""

job = dyson.submit_source_job(
    SOURCE,
    "workload",
    [1, 2, 3, 4, 5],
    profile="auto",
    timeout_seconds=300,
)

terminal = dyson.wait_for_job(job["job_id"], timeout=600)
result = dyson.get_job_result(terminal["job_id"])
print(result["stdout"])

Execution catalog

Scale only to what the job needs.

Base rates are transparent. Growth applies a 0.85 multiplier to estimates and settlement.

0.02 cr/min

Small CPU

Bounded scripts and light transforms

compute: 1 vCPU
memory: 512 MiB
max runtime: 15 min

0.05 cr/min

Balanced CPU

NumPy and parallel CPU workloads

compute: 2 vCPU
memory: 2 GiB
max runtime: 30 min

0.12 cr/min

High-memory CPU

Larger arrays and simulations

compute: 4 vCPU
memory: 8 GiB
max runtime: 60 min

0.30 cr/min

NVIDIA L4

PyTorch inference on CUDA

compute: 4 vCPU + L4
memory: 16 GiB
max runtime: 55 min

CPU workers support Python and NumPy. The GPU worker supports PyTorch on CUDA. C++, TensorFlow, JAX, managed training, and multi-device execution are not advertised as GA runtimes.

Production boundaries

Built so workload code cannot become control-plane code.

Route from requirements

Choose a profile explicitly or let memory, parallelism, workload type, and GPU intent select the lowest compatible shape.

Deterministic policy

Keep secrets out of workloads

Worker images receive no billing secret, database access, cloud credentials, or general internet route—only expiring artifact URLs.

Credential-free workers

Reserve, then reconcile

A timeout ceiling is reserved before launch. Cancellation before compute starts releases it in full; started cancellation is canonically metered, and every other capture requires exact bounded output.

No silent double charge

Observe every job

Inspect state, routing reason, duration, logs, result metadata, included usage, paid overage, and final billing. Temporary storage faults retain the exact job and honor the server retry delay; lifecycle expiry is permanent and explicit.

One control plane

Plans