What kyma is

A single data engine that turns the whole production picture into something agents can query.

Point every OTLP emitter in your stack — services, Kubernetes, CI, queues, frontend RUM, your agents themselves — at one engine. Then let an agent ask it anything, in KQL or SQL, at sub-second latency over a decade of history. No dashboards to scrape. No vendor APIs to juggle. No rate limits.

What it does

• Ingests every signal your stack already emits — logs, traces, metrics, spans, tool calls, prompt and response bodies, deploy events, config diffs, audit trails — through one OTLP pipe, plus REST, Kafka, and file-drop for non-OTLP sources.
• Stores it as columnar Arrow on object storage you own (S3, MinIO, any object_store impl), with per-extent column statistics and token indices that make 99 %+ of queries skip 99 %+ of data.
• Answers in KQL, SQL, or PromQL over Arrow Flight gRPC — exact rows, streamed zero-copy — so an agent can ask twenty exploratory questions per user prompt without melting a credit card.
• Federates with your operational databases — Postgres, MySQL, MongoDB register as catalogs, and DataFusion joins them with kyma's own tables in a single query. Live or synced; you pick per source.
• Scales from one binary to many nodes without a rewrite. The catalog is externalized from byte one, compute is stateless, object storage is the source of truth.

What it is not

• Not a metrics database. Metrics are one signal among many; treating them specially fragments the query surface that agents need to reason across.
• Not a log search appliance. Full-text on the dynamic column works, but the engine is column-aware first; a search-shaped query plans like a filter, not like a Lucene query.
• Not an OLAP warehouse for business data. The data model is built around append-only, time-partitioned, token-indexed columnar storage. Use it for observability and operational state. Use it for joining production telemetry with your customer table. Don't use it as your primary BI cube.
• Not a managed service. kyma is a binary you run. It expects an object store and a Postgres catalog you own.

Who it's for

The agent that needs production awareness across the whole stack.

Concretely: anyone whose tools answer questions like "what changed?", "where did this break?", "which customer is affected?", "is this regression real or last-week's-rollout's tail?" — and who's tired of stitching three vendors together to answer them.

Where to go next

• The mental model: The five invariants.
• How a query stays fast: The pruning cascade.
• Multi-source data: Multi-source data.
• Hands-on in five minutes: Quickstart.