Platform
Live infrastructure, the seven research systems behind the product, and how they are measured.
Live system status
Research systems (each reproduced on real data)
B1DRL-PPO scheduler
Xu et al., 2024 (arXiv:2403.07905)
+112% reward vs best classical baseline; 0.57% SLA violations
PPO agent (stable-baselines3) trained in a custom Gymnasium cluster environment; learns placement from live telemetry instead of static heuristics.
B2eBPF telemetry
eHashPipe (HashPipe sketch)
Top-K syscall tracking at 100% precision (small k) in bounded memory
Tetragon in-kernel probes feed a 500ms state vector to the scheduler; first-party corpus of 171k syscall events captured on GCP.
B3LSTM prewarming
IEEE Transformer cold-start (LSTM baseline)
Median N-RMSE 0.085 on the Azure Functions 2019 production trace
A global LSTM trained across functions beats the per-function paper baseline; predictions drive the warm-pool size.
B4Adaptive sandboxing
Iacovazzi & Raza, IEEE CSR 2022 + 2 more
Policy gate recall 0.95 @ ~0 FPR; IDS 0.80 acc / 0.10 FPR on first-party eBPF data
Risk scorer picks runc / gVisor / Firecracker per workload; a syscall-graph IDS and a transactional policy gate block escapes.
B5Multi-cluster federation
arXiv:2512.24914 (AI multi-cluster)
Live Karmada failover verified: replicas reschedule to the surviving cluster
Divided 70/30 propagation across regions; the AI loop retunes weights from utilization + carbon; cluster-loss reschedule proven on GCP.
B6Carbon-aware scheduling
PCAPS (Lechowicz et al.)
25.8% CO2 reduction with 12h flexibility, 45% with 24h
Live grid intensity (electricityMaps / UK Carbon API) shifts deferrable work to greener windows and regions.
B7CRDT collaboration
Eg-walker (Kleppmann, EuroSys 2025)
Convergence verified on the authors’ published editing traces
Yjs CRDTs over WebSocket keep every cursor and keystroke consistent with no central lock; same family as Figma’s sync.
Tested to failure
We killed a member cluster live on GCP: Karmada detected it (NotReady, NoExecute taint) and rescheduled all replicas onto the surviving cluster in about 10 seconds. The executor policy gate blocks destructive code (rm -rf /, fork bombs) before it ever runs.