Staff/Sr Software Engineer, Compute Capacity

What This Team OwnsThe teams work spans three functional areas. Depending on your background and interests, youll focus primarily in one, but the boundaries are fluid and the problems overlap:- Data infrastructure collecting, normalizing, and serving the fleet-wide data that powers everything else. This means building pipelines that ingest occupancy and utilization telemetry from Kubernetes clusters, normalizing billing and usage data across cloud providers, and maintaining the BigQuery layer that the rest of the org queries against. Correctness, completeness, and latency matter here.```{=html}```- Fleet observability making the state of the accelerator fleet legible and actionable in real time. This means building cluster health tooling, capacity planning platforms, alerting on occupancy drops and allocation problems, and driving systemic improvements to scheduling and fragmentation. The work sits at the intersection of Kubernetes operations and cross-team coordination.```{=html}```- Compute efficiency measuring and improving how effectively every major workload uses the hardware its running on. This means instrumenting utilization metrics across training, inference, and eval systems, building benchmarking infrastructure, establishing per-config baselines, and collaborating directly with system-owning teams to close efficiency gaps.```{=html}```- Internal compute tooling building the platforms and interfaces that make capacity data usable across the org. This includes capacity planning tools, workload attribution systems, cost dashboards, and self-service APIs. The consumers are research engineers, infrastructure teams, finance, and leadership each with different needs and different levels of technical depth. The work involves product thinking as much as engineering: figuring out what people actually need, defining schema contracts, and making the data discoverable.You will be placed on a pod based on your background and interests. We are especially focused on hiring for Data Platform, but strong candidates for any of the three active pods will move forward. What Youll Do- Build and operate data pipelinesthat ingest accelerator occupancy, utilization, and cost data from multiple cloud providers into BigQuery. Own data completeness, latency SLOs, gap detection, and backfill automation.- Develop and maintain observability infrastructure Prometheus recording rules, Grafana dashboards, and alerting systems that surface actionable signals about fleet health, occupancy, and efficiency.- Instrument and analyze compute efficiency metricsacross training, inference, and eval workloads. Build benchmarking infrastructure, establish per-config baselines, and work with system-owning teams to improve utilization.- Build internal tooling and platformsthat enable capacity planning, workload attribution, and cluster debugging. The consumers are other engineering teams, finance, and leadership not external users.- Operate Kubernetes-native systems at scale deploying data collection agents, managing workload labeling infrastructure, and understanding how taints, reservations, and scheduling affect capacity.- Normalize and reconcile data across heterogeneous sources including AWS, GCP, and Azure billing exports, vendor-specific telemetry formats, and internal systems with different schemas and billing arrangements.- Collaborate across organizational boundarieswith research engineering, infrastructure, inference, and finance teams. Gather requirements from technical stakeholders, translate them into useful systems, and communicate trade-offs to non-technical audiences. You May Be a Good Fit If You Have- 5+ years of software engineering experiencewith a strong track record building and operating production systems. You write code every day this is a hands-on engineering r le, not a planning or coordination role.- Kubernetes fluency at operational depth youve operated production K8s at meaningful scale, not just written manifests. Comfort with scheduling, taints, labels, node management, and debugging cluster-level issues.- Data pipeline engineering experience designing, building, and owning the full lifecycle of production data pipelines. Experience with data warehouses (BigQuery preferred), schema management, streaming ingestion, SLOs for latency and completeness, and a strong instinct for correctness.- Observability tooling experience Prometheus, PromQL, and Grafana are in the critical path for this team. Experience writing recording rules, understanding metric semantics, and building monitoring systems that engineering teams actually rely on.- Python and SQL at production quality.Most pipeline code is Python; the presentation layer is BigQuery SQL including table-valued functions and views. Both need to be idiomatic, well-tested, and maintainable.- Familiarity with at least one major cloud provider (AWS, GCP, or Azure)at the infrastructure level compute, billing, usage APIs, cost management tooling. Multi-cloud experience is a strong plus.- High autonomy and strong cross-team communication.You can gather your own requirements, navigate ambiguity, and work across organizational boundaries. Scrappiness and ownership matter more than polish. Strong Candidates May Also Have- Multi-cloud data ingestion experience especially working with AWS and GCP APIs, billing exports, or vendor-specific telemetry formats. Experience normalizing data from external providers with different billing arrangements is directly applicable.- Accelerator infrastructure familiarity GPU metrics (DCGM), TPU utilization, Trainium power and utilization metrics, or experience working with ML training/inference systems at the hardware level.- Performance engineering and benchmarking experience building benchmark harnesses, establishing baselines, reasoning about compute efficiency (FLOPs utilization, memory bandwidth, interconnect throughput), and working with system teams to diagnose and improve performance.- Data-as-product thinking experience building internal data products with self-service access, schema contracts, API serving, documentation, and discoverability. Not just building pipelines, but thinking about how platform data gets consumed.- Experience with capacity planning, resource management, or cost attribution systemsat a hyperscaler or large-scale ML environment. FinOps, chargeback systems, or infrastructure cost modeling.Familiarity with ClickHouse, Terraform, or Rust.ClickHouse is the teams current streaming store; Terraform for infrastructure-as-code; Rust for high-performance data collection agents.::: {s=

Created: 2026-02-25