forgesight-clickhouse

The ClickHouse exporter for ForgeSight. Writes each ForgeSight record as one row in a denormalized single-table MergeTree, batched columnar — so a platform team can query its agent telemetry in SQL: spend by team, p99 LLM latency by model, tokens per workflow over 90 days.

pip install forgesight-clickhouse

import forgesight
from forgesight_clickhouse import ClickHouseExporter

forgesight.configure(exporters=[
    ClickHouseExporter(
        dsn="clickhouse://user:pass@ch-host:8443/agents?secure=true",
        table="forgesight_records",
        create_table=True,   # dev convenience; production runs the shipped migration
    ),
])

Or by name: exporters: [{name: clickhouse, config: {dsn: "${FORGESIGHT_CLICKHOUSE_DSN}"}}].

Why a columnar exporter

A Record is written once and never updated (OTel's immutable span model), so trace-level business metadata is denormalized onto every row — no join table, so analytical queries never pay a join:

SELECT agent_name, quantile(0.99)(duration_ms)
FROM forgesight_records WHERE kind = 'llm' GROUP BY agent_name;

SELECT metadata.team, sum(cost_usd)
FROM forgesight_records
WHERE kind = 'llm' AND start_time >= now() - INTERVAL 30 DAY
GROUP BY metadata.team;

Inserts are batched by the SDK's export pipeline (ClickHouse hates row-at-a-time): one columnar INSERT per batch, on the export worker, fault-isolated. A ClickHouse outage makes export() return FAILURE (counted, never raised — P6); it never blocks the agent.

Schema & migrations

The DDL ships in the package (migrations/0001_init.sql). create_table=True runs CREATE TABLE IF NOT EXISTS on first export for dev; production applies the migration out-of-band. New domain fields arrive as new nullable columns via numbered migrations — old queries keep working.

Configuration

Constructor kwargs win over env (FR-12). Prompt/response content is never stored in the base table; it is gated SDK-wide by capture_content (off by default, P7).

License

Apache-2.0