kyma

Appearance

Query ​

Three query languages, two transports, one agent endpoint, and the performance rules that apply to all of them. They share more than they differ: every surface produces the same kyma logical plan, runs through the same pruning cascade, executes on the same DataFusion engine, and returns the same Arrow result.

Pick the surface that matches how you're calling kyma — not what the data looks like, because that's the same on the other side.

KQL ​

Pipe-style language — read left-to-right, one operator at a time. The default for ad-hoc operator queries and what the agent generates under the hood. Content-Type: application/x-kql.

SQL ​

Standard SQL via DataFusion. Same logical plan as KQL underneath, just a different surface. Use it for cross-source joins, complex sub-queries, or when SQL is what your tool already speaks. Content-Type: application/sql.

PromQL 🚧 ​

Reserved for Content-Type: application/promql; frontend lands in a later milestone. When it ships, existing Prometheus dashboards point at kyma's endpoint with no other change.

Arrow Flight ​

gRPC at :9090 — zero-copy Arrow batches over the wire. Use it when the NDJSON HTTP response is the bottleneck: streaming millions of rows to a Polars DataFrame, a Spark connector, an Arrow-native client.

Agent endpoint ​

POST /v1/agent/ask. Natural-language question in, Server-Sent Events out. The agent picks the right table, writes the KQL, runs it, streams back tokens. See The agent loop for the conceptual model.

Pruning and performance ​

The practical companion to The pruning cascade — what makes a query fast or slow, how to read pushdown_summary and the metrics, and a short cookbook of the patterns that prune well.

What's the same across all surfaces ​

  • One logical plan. Every frontend parses to the unified IR in kyma-plan. KQL and SQL are peer trait implementations; the planner doesn't know which one fed it.
  • One execution engine. DataFusion runs vectorized over Arrow. Cross-source joins are the same exec path as native joins.
  • One pruning cascade. Catalog → extent footer → block index → decode. A query without a time bound plans like a full scan regardless of which language wrote it.
  • One result shape. Arrow RecordBatches, served as NDJSON over HTTP or as Flight RecordBatch streams over gRPC.