How we ingest 2.4 billion spans a month

Agents are chatty

A single agent run can emit hundreds of spans, each carrying full prompts and completions. At our current scale that’s 2.4 billion spans and roughly 90 TB of payload a month — and customers expect to search all of it in under a second.

Our pipeline buffers spans at the edge, batches them into columnar segments, and writes payloads to object storage with content-addressed deduplication. Repeated system prompts, which make up nearly 40% of raw bytes, are stored exactly once.

Sample everything except what matters

Successful, low-cost runs are sampled adaptively once you pass your plan’s volume. Errors, guardrail events and anything an eval flagged are always kept at full fidelity. You never lose the trace you’ll need during an incident.