Portable by construction
Load MLeap bundles from local files, Amazon S3, or Google Cloud Storage.
file:// · s3:// · gs://
OPEN SOURCE / APACHE-2.0 / SCALA 2.12
LOAD. TRANSFORM. RANK.
Skynet is a production-shaped runtime for portable MLeap bundles. It turns model artifacts into inspectable HTTP services—with explicit lifecycle control, typed frame transforms, ranking, health checks, graphs, and metrics.
RUNTIME / 0.3.3
REFERENCE
EXAMPLE / RECOMMENDER
portable_bundle.zip
s3://models/production
METRICSON
EXECpar-12
FRAMES04
ORIGINSFILE / S3 / GCS
FRAME SHAPES04
HTTP OPERATIONS10
RUNTIMEZIO + HTTP4S
01 / SYSTEM
A control plane for the model after training.
Training produces an artifact. Skynet handles what comes next: acquiring the bundle, loading its transformer graph, understanding its schema, executing it safely, and exposing enough evidence to operate it.
Frames with structure
Serve Leap, Cartesian, Context, and Prefixed frame shapes through one API.
Frame → DefaultLeapFrame
Ranking as a first-class path
Transform, select, group, average, sort, and return top-k identifiers.
rank · group · avg · k
Inspect before guessing
Generate sample frames, inspect schemas, render model graphs, and warm models.
sample · graph · health
Parallelism per request
Choose sequential, unconstrained parallel, or bounded parallel execution.
seq · par · par-n
Operational surface included
Swagger, service and model health, latency histograms, errors, and pool metrics.
/docs · /health · /metrics
02 / LIFECYCLE
The entire serving path is visible.
Each stage has a clear responsibility. Change the artifact source without changing inference. Change request parallelism without rebuilding the service.
-
01
Register
Name a model and provide its bundle URI.
PUT /models/{model} -
02
Acquire
Resolve the URI through the matching repository adapter.
file · S3 · GCS -
03
Load
Deserialize the MLeap bundle and cache its transformer.
Bundle[Transformer] -
04
Warm
Generate schema-correct sample frames and execute predictions.
GET /models/{model}/health -
05
Serve
Transform or rank frames with request-level execution control.
POST /transform · /rank -
06
Observe
Read response, failure, latency, and executor-pool metrics.
GET /metrics
03 / HTTP API
A small surface with deep control.
Tapir defines the endpoints once and derives the OpenAPI document used by the bundled Swagger UI. Filter the operations below by intent.
REQUEST LAB
From artifact to ranked output.
PUT /models/recommender
04 / QUICK START
Serve a bundle locally.
Skynet is an open-source runtime, not a hosted account. Clone it, choose a port, and register the model artifact you want the process to serve.
JDK 11sbt 1.xMLeap bundle
Clone and run
git clone https://github.com/AdrielC/skynet.git
cd skynet && HTTP_PORT=8080 sbt run
Verify the runtime
curl http://localhost:8080/health
Register a bundle
curl -X PUT http://localhost:8080/models/recommender --data 's3://models/recommender.zip'
Inspect and test
Open http://localhost:8080/docs for Swagger, or ask the model for a
schema-correct sample at /models/recommender/sample.
05 / OPERATE
Know what the runtime is doing.
Skynet’s operational surface covers the service, each loaded model, every named endpoint, and the executors doing the work.
PROMETHEUS / LIVE
- http_requests_duration_millis
- endpoint × status
- http_requests_total
- service × error
- workers_count
- blocking × async
- pool_tasks_enqueued
- executor pressure
GET /healthConfirms the model registry is available.
GET /models/{model}/healthRuns generated sample data through the loaded transformer.
GET /models/{model}/sampleProduces input shaped to the model’s declared schema.
GET /models/{model}/graphRenders the model’s internal computation graph as SVG.
OPEN SOURCE / BUILD WITH IT
The model is portable.