Streaming xlsx reports

[KremnevDmitry]

Summary

Added streaming XLSX report generation for large datasets: a report band can be marked to render row-by-row straight from a database cursor, so memory stays bounded regardless of row count (verified at ~1M rows). The feature is opt-in per band and inert for existing reports.

What was done

• Introduced a per-band streaming flag (BandDefinition, exposed via ReportBand/ReportQuery) and a Streaming checkbox in the report editor, gated to a first-level band backed by a single SQL or JPQL dataset.
• Added a streaming XLSX formatter built on POI SXSSF that writes rows incrementally, plus supporting pieces (band feed, style cache, XLSX→CSV streaming writer, template-band model). The streaming formatter is selected automatically when a band has the flag set.
• Added forward-only DB-cursor streaming for the SQL and JPQL loaders (configurable JDBC fetch size, read-only transaction, periodic persistence-context clearing) behind a StreamingReportDataLoader abstraction.
• Added StreamingReportValidator that enforces the structural rules streaming requires (single dataset, first-level band, supported dataset type) and surfaces clear errors otherwise.
• Added configuration under jmix.reports.streaming.* (fetch size and related knobs, validated to be positive) and translations (default + ru) for the new UI.
• (jmix-premium) Added a spreadsheet-report preview size gate: results above a configurable threshold are downloaded instead of being loaded into the in-memory Vaadin Spreadsheet, avoiding RecordFormatException/OOM on huge outputs.

How it works

When a first-level band is marked streaming, the engine reads its rows from a forward-only database cursor and writes them directly into an SXSSF sheet one row at a time, so peak memory does not grow with the result size. Non-streaming bands and templates render through the existing engine unchanged. Formulas are resolved at write time — in-row formulas shift per band instance, and trailing aggregates placed below the data grow to cover all rendered rows. The render runs in a read-only transaction, and the persistence context is cleared periodically to bound managed-entity growth while lazy attributes stay resolvable through the open session. Output can be post-converted to CSV via a streaming writer.

How to use

In the report editor, check Streaming on a first-level band whose dataset is a single SQL or JPQL query (the checkbox is disabled for other configurations). Run the report with an XLSX or CSV template as usual — the streaming engine is chosen automatically. Adjust jmix.reports.streaming.fetch-size and related properties if the default cursor page size doesn't suit the data store.

Compatibility

Opt-in and backward compatible: bands are non-streaming by default, and any report that doesn't set the flag renders exactly as before.