Drava is an event-driven runtime for edge scientific streaming pipelines. At facilities like the Advanced Photon Source, detectors produce data faster than it can be shipped to a central site, so reconstruction and inference run near the instrument under latency and throughput constraints. Drava runs these multi-stage pipelines and manages the system concerns that are usually left to manual tuning, such as batching, scheduling, and resource placement.
A developer writes each stage as a reusable callback. The runtime manages microbatching, dispatch, data movement, and observability, which separates the scientific logic from the system concerns. Drava is built on the XKRT tasking runtime.
A Drava workflow is a directed acyclic graph. Nodes are stages and edges are transport routes. Each stage has an ingress route, a callback that performs the computation, an egress route for publishing results, and runtime parameters that control batching, scheduling, and threads.
The application defines the ingress, callback, and egress. The runtime parameters come from a YAML file, so the same callback runs under different execution policies without code changes. Stages are stateless functions triggered by data events.
- Application interface. Callbacks are registered through a C or Python API.
- Pipeline configuration. A YAML file binds each stage to its transport, batching policy, and thread team.
- Execution engine. Built on XKRT. One I/O thread pulls events, compute threads run callback tasks with work-stealing, and a microbatching layer flushes on a size threshold, an end-of-stream event, or a timeout. Lock-free atomic counters record metrics.
- Transport. Stages exchange events over a publish/subscribe model. A stage subscribes to its ingress subject and publishes to its egress subject. Drava uses two interchangeable backends, a Unix socket for intra-node communication and NATS JetStream for durable pub/sub across nodes.
Each stage records throughput, end-to-end latency, and GPU and CPU energy, and exposes them as per-run CSV and JSON. This observability drives an agentic ytopt search that tunes runtime parameters automatically. The metrics schema is documented on Read the Docs.
Guides and the generated C and Python API reference are on Read the Docs: https://drava.readthedocs.io.
| Example | Description |
|---|---|
| PtychoNN | Two-stage ptychographic inference and stitching |
| TomoGAN | Single-stage tomographic denoising with energy reporting |
| Iris KNN | Minimal single-stage inference |
| Bare runtime | Runtime message-rate ceiling, no model |
| Dataflow | Minimal transport demonstration |
- XKRT, the tasking runtime Drava is built on
- NATS JetStream, the publish/subscribe transport
- ytopt, the Bayesian optimization framework used for tuning
- Advanced Photon Source (APS), Argonne National Laboratory
- JLSE, the evaluation cluster at Argonne National Laboratory
Distributed under the terms in LICENSE. Contributors are listed in AUTHORS.