Open Standards
Prometheus, Grafana, Loki, Jaeger, OpenTelemetry. If your team already knows these tools, they’ll feel at home on day one.
/ OBSERVABILITY
Metrics, logs & traces purpose-built for AI
Tscale Observability delivers Prometheus metrics, Grafana dashboards, Loki log aggregation, and Jaeger distributed tracing — pre-wired for GPU workloads, training jobs, and inference endpoints. 90-day retention, zero infra to manage.
GPU-Native Telemetry
DCGM exporters, NVML metrics, NCCL debug logs, CUDA traces. Deep GPU observability that off-the-shelf stacks don’t ship.
90-Day Retention
Hot metrics for 30 days, warm for 90, cold in S3 forever. Long enough to debug that production regression from three months ago.
/ THE THREE PILLARS
Metrics, logs, traces all in one stack
Open-source observability, pre-wired and production-ready. No operators to install, no Helm charts to debug, no Prometheus HA architecture to maintain — Tscale handles the platform, you handle the dashboards.
- Prometheus metrics — 10K+ series per cluster, sub-second scrape intervals, long-term storage in Thanos.
- Loki log aggregation — structured logs from every service, queryable with LogQL, no log volume limits.
- Jaeger distributed tracing — trace requests across services, find the slow span, debug in production.
- Pre-built dashboards — 50+ Grafana dashboards for GPU, training, inference, networking, and storage — ready to use.
/ AI TELEMETRY
GPU telemetry that actually tells you something
Stock Prometheus exporters miss the metrics that matter for AI. Tscale ships GPU-aware collectors out of the box — SM occupancy, tensor core utilization, HBM bandwidth, NVLink health, and NCCL all-reduce timing.
-
DCGM exporter built-in
SM occupancy, tensor core util, HBM throughput, NVLink errors, PCIe bandwidth — every metric, no setup.
-
NCCL profiling
All-reduce, all-gather, and broadcast timing per-step. Find the communication bottleneck in your training job.
-
Anomaly detection
ML-based alerts on drift, hang detection, and silent failures — catch issues before they cost you a training run.
/ PLATFORM
Built on the open-source stack
Tscale Observability is the production-hardened version of the CNCF observability ecosystem — same APIs, same query languages, same dashboards — just managed for you at AI scale.
Metrics
- Prometheus 2.x
- Thanos long-term storage
- OpenMetrics ingestion
- Recording rules engine
- Push gateway support
Logs
- Loki 3.x
- LogQL query language
- Multi-tenant storage
- S3-backed chunk store
- Live tail & filter
Traces
- Jaeger 1.x
- OpenTelemetry SDK
- OTLP / Zipkin / Jaeger
- Tail-based sampling
- Service dependency maps
Visualisation
- Grafana 11.x
- 50+ pre-built dashboards
- Custom dashboard builder
- Public dashboard links
- Embeddable panels
Alerting
- Alertmanager integration
- PagerDuty & Opsgenie
- Slack & Microsoft Teams
- Webhooks (any destination)
- ML-based anomaly alerts
Integrations
- Datadog / New Relic import
- Snowflake / BigQuery export
- Splunk forwarder
- Custom OTLP receivers
- Terraform provider
Performance
Per cluster scrape
Sub-second scrape intervals, multi-tenant, no rate limits. Every metric, every second.
Hot + warm + cold
30 days hot (fast queries), 60 days warm (compressed), forever in S3 (forensic).
Real-time visibility
Sub-second Prometheus scrapes, sub-100ms Grafana queries. No 30-second polling lag.
No per-volume limits
Ship every log without thinking about retention tiers or per-stream caps.
For teams that need answers, not just data
ML Engineers
GPU telemetry, training job metrics, model serving traces. The observability you actually need for ML workloads — not generic host metrics.
Learn MoreSRE & Platform Teams
Open standards, long retention, alerting integrations, RBAC. Built for the team that owns uptime.
Learn MoreObservability ties the platform together
Every Tscale service ships with metrics, logs, and traces. Pair Observability with the rest of the stack to get full visibility into your AI fleet.