Chunduru Sri Abhijit, Zilinghan Li, Matthew Joel, Ravi Madduri · SoftwareX 2026 · 2026
DOI: 10.1016/j.softx.2026.103195
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
We present HiveWatch , an open-source, framework-agnostic observability toolkit for federated learning (FL). HiveWatch provides a minimal five-call training-lifecycle API and an FL-native schema for round summaries, client updates, communication volume, resource utilization, staleness measurements, failure events, and geographic metadata. A pluggable emitter interface routes a single normalized event stream simultaneously to replayable local JSONL artifacts and a geographic dashboard through GeoEmitter , to Weights & Biases through WandbEmitter , and to MLflow through MLflowEmitter . We demonstrate HiveWatch ’s portability across three architecturally distinct FL frameworks, APPFL, Flower, and NVIDIA FLARE, using the same API in each case. In a synthetic server-side microbenchmark of 50 rounds and 20 clients per round on an AWS EC2 t3.medium instance with 2 vCPUs and 4 GiB of RAM, internal bookkeeping required an average of 7.3 𝜇 s per client-update call and 102.3 𝜇 s per round-summary call, whereas the synchronous GeoEmitter required 3.02 ms and 3.52 ms, respectively. The software, framework integrations, examples, tests, and benchmark artifacts are available under the MIT License in the versioned v0.3.1 release.
No comments yet — start the discussion below.