Cross-trace cost dashboard SQL patterns

Applies to:

Plan - Pro and Enterprise for custom SQL charts
Deployment - Any
Use case - How to build accurate multi-project cost dashboards and SQL charts that filter by root-span metadata (user_id) while aggregating estimated_cost() across traces/projects

Summary

Use Cross-trace filtering for cost charts filtered by root-span metadata so contributing child spans remain included.

What is happening

Span filtering matches individual spans. A filter on a field that exists on a different span will exclude those measured spans. Cross-trace filtering includes all spans from traces where any span matches the filter.

Fix or suggestion

For unattributed cost across projects, use a per-trace aggregation pattern. Group by root_span_id, surface metadata with any_value(), sum estimated_cost() for the selected spans, then filter for missing user_id and aggregate in an outer query. Example per-trace rollup across projects:

SELECT coalesce(sum(trace_cost), 0) AS unattributed_cost
FROM (
  SELECT
    root_span_id,
    any_value(metadata.user_id) AS user_id,
    sum(CASE WHEN metadata.braintrust.integration_name = 'langchain-py' THEN estimated_cost() ELSE 0 END) AS trace_cost
  FROM project_logs('proj-id-1', 'proj-id-2', 'proj-id-3')
  GROUP BY root_span_id
) AS per_trace
WHERE user_id IS NULL

Option 1: Use `spans` shape and `WHERE is_root` for root-only inspection

Use this when you only need root-span rows. WHERE is_root = true is also valid. Example: count root spans and roots with user_id for one project.

SELECT
  metadata.braintrust.integration_name AS integration,
  count(*) AS root_spans,
  count(metadata.user_id) AS root_spans_with_user_id
FROM project_logs('99fa755f-3519-4f5b-9e36-e1dd23917dca')
WHERE is_root
GROUP BY integration

Option 2: Use `traces` or `summary` shape with `ANY_SPAN(...)`

Use this when you need the traces or summary shape but must restrict traces by a root-span condition. Example: match traces whose root span has a metadata.user_id value of 123.

SELECT *
FROM project_logs('proj-id', shape => 'traces')
WHERE ANY_SPAN(is_root AND metadata.user_id = '123')

How to confirm it worked

Use the root-count query to inspect where user_id is present. Roots without user_id will still have zero root_spans_with_user_id.
Compare summed estimated_cost() before and after switching to Cross-trace filtering. Totals can change because additional spans contribute. Each span’s estimated cost calculation remains unchanged.

Notes

Select projects by passing their IDs to project_logs() because project_id is not guaranteed as a row field.
For a per-project breakdown, query each project separately. Passing multiple project IDs combines their data.

​Summary

​What is happening

​Fix or suggestion

​Option 1: Use spans shape and WHERE is_root for root-only inspection

​Option 2: Use traces or summary shape with ANY_SPAN(...)

​How to confirm it worked

​Notes