Parquet exports to blob storage
Niklas Semmler, PhDScheduled blob storage exports can now be written as Apache Parquet files. Parquet is the default for new integrations and configurable via the public API.
Scheduled blob storage exports to S3, GCS, and Azure can now be written as Apache Parquet files — available to all projects and the new default for new integrations. Parquet is a columnar binary format that loads directly into warehouses and query engines like BigQuery, Snowflake, ClickHouse, or DuckDB, with no CSV parsing or JSON casting step:
- Query exports in place. Point DuckDB or Athena at your bucket and query the files directly — column pruning and predicate pushdown work out of the box.
- Skip the staging schema. Parquet files carry typed columns, so warehouse loads don't need per-field cast logic like text formats do.
- Smaller files without gzip. Parquet is compressed by the storage engine itself; the gzip option doesn't apply.
Select Parquet under Project Settings → Integrations → Blob Storage, or set it programmatically: the fileType field on GET/PUT /api/public/integrations/blob-storage now accepts PARQUET alongside CSV, JSON, and JSONL. Existing integrations keep their configured format.
Parquet observation exports do not include the per-unit model price columns (input_price, output_price, total_price). These columns snapshot model definition prices at export time and may be deprecated in the future — use cost_details and total_cost for cost data, which are included in every file type. Trace and score exports are identical across all file types. See the Export Field Reference for details.