Is Anyscale Down?

If Prismix shows Anyscale as "Operational" but your clusters or endpoints are failing:

• Ray cluster not scaling (autoscaler stuck) — the Anyscale autoscaler can stall when the cloud provider (AWS/GCP) has insufficient capacity for the requested instance type. Check the cluster event log in the Anyscale console for InsufficientInstanceCapacity errors and switch to a different instance family or region with available GPU stock.
• LLM endpoint cold start over 60 seconds — Anyscale LLM endpoints on serverless tiers may take 60–90 seconds to start from a cold state when no replicas are warm. Enable a minimum replica count of 1 in your endpoint configuration to keep at least one replica always warm, avoiding cold starts at the cost of idle compute.
• Out-of-memory on GPU instance (CUDA OOM) — large models like Llama-70B require multi-GPU tensor parallelism. If you deploy to a single A10G (24GB VRAM) instead of multiple GPUs, the model will OOM during loading. Verify the minimum recommended GPU configuration for your model size and set num_gpus accordingly in your serve config.
• Ray object store eviction during long training runs — Ray's in-memory object store evicts objects under memory pressure using LRU. During distributed training, large checkpoints or datasets pinned to the object store can be evicted mid-run. Increase the object_store_memory setting or pin critical objects explicitly with ray.put(obj) and hold a reference.
• API endpoint returning 503 during scheduled maintenance — Anyscale performs rolling maintenance windows for the control plane. During these windows, new cluster launches and endpoint deployments may fail with 503. Check status.anyscale.com for scheduled maintenance notices; existing running clusters are unaffected during control-plane maintenance.
• Region-specific GPU availability exhausted — certain GPU types (A100, H100) are in high demand in us-west-2 and eu-west-1. If your cluster config hard-codes a single region, you may be unable to scale when GPUs are sold out. Add fallback regions to your cluster config or use Anyscale's multi-cloud compute config to allow cross-region autoscaling.