Skip to content

feat: inline Arrow IPC support with format fallback#273

Closed
jamesbroadhead wants to merge 7 commits intodatabricks:mainfrom
jamesbroadhead:fix/arrow-stream-ts
Closed

feat: inline Arrow IPC support with format fallback#273
jamesbroadhead wants to merge 7 commits intodatabricks:mainfrom
jamesbroadhead:fix/arrow-stream-ts

Conversation

@jamesbroadhead
Copy link
Copy Markdown
Contributor

Summary

Adds support for serverless SQL warehouses that return inline Arrow IPC data in result.attachment instead of data_array, plus automatic format fallback for warehouse compatibility.

  • Inline Arrow IPC decoding: _transformArrowAttachment decodes base64 Arrow IPC from result.attachment into row objects via tableFromIPC
  • Format fallback: When default ARROW_STREAM is rejected, automatically falls back through JSON → ARROW (matching warehouse capabilities)
  • 586 lines of new tests: connector tests for Arrow IPC attachment handling + analytics plugin format fallback tests

Changes

  • packages/appkit/src/connectors/sql-warehouse/client.ts — Arrow IPC attachment detection and decoding
  • packages/appkit/src/plugins/analytics/analytics.ts_executeWithFormatFallback with ARROW_STREAM → JSON → ARROW chain
  • packages/appkit-ui/src/react/hooks/types.ts — ARROW_STREAM format type support
  • packages/appkit/src/type-generator/query-registry.ts — typegen compatibility

Test plan

  • 1566 TS tests pass (pnpm test)
  • New connector tests cover inline attachment decoding, external links, format fallback
  • New analytics tests cover format parameter handling and fallback chain
  • Manual test against serverless warehouse returning Arrow IPC

Related

This pull request was AI-assisted by Isaac.

Some serverless warehouses only support ARROW_STREAM with INLINE
disposition, but the analytics plugin only offered JSON_ARRAY (INLINE)
and ARROW_STREAM (EXTERNAL_LINKS). This adds a new "ARROW_STREAM"
format option that uses INLINE disposition, making the plugin
compatible with these warehouses.

Fixes databricks#242
Tests verify:
- ARROW_STREAM format passes INLINE disposition + ARROW_STREAM format
- ARROW format passes EXTERNAL_LINKS disposition + ARROW_STREAM format
- Default JSON format does not pass disposition or format overrides
The server-side ARROW_STREAM format added in the previous commit was
not exposed to the frontend or typegen:

- Add "ARROW_STREAM" to AnalyticsFormat in appkit-ui hooks
- Add "arrow_stream" to DataFormat in chart types
- Handle "arrow_stream" in useChartData's resolveFormat()
- Make typegen resilient to ARROW_STREAM-only warehouses by
  retrying DESCRIBE QUERY without format when JSON_ARRAY is rejected

Co-authored-by: Isaac
Signed-off-by: James Broadhead <jamesbroadhead@gmail.com>
…compatibility

ARROW_STREAM with INLINE disposition is the only format that works
across all warehouse types, including serverless warehouses that reject
JSON_ARRAY. Change the default from JSON to ARROW_STREAM throughout:

- Server: defaults.ts, analytics plugin request handler
- Client: useAnalyticsQuery, UseAnalyticsQueryOptions, useChartData
- Tests: update assertions for new default

JSON and ARROW formats remain available via explicit format parameter.

Co-authored-by: Isaac
Signed-off-by: James Broadhead <jamesbroadhead@gmail.com>
When using the default ARROW_STREAM format, the analytics plugin now
automatically falls back through formats if the warehouse rejects one:
ARROW_STREAM → JSON → ARROW.

This handles warehouses that only support a subset of format/disposition
combinations without requiring users to know their warehouse's
capabilities. Explicit format requests (JSON, ARROW) are respected
without fallback.

Co-authored-by: Isaac
Signed-off-by: James Broadhead <jamesbroadhead@gmail.com>
Previously, _transformDataArray unconditionally called updateWithArrowStatus
for any ARROW_STREAM response, which discards inline data and returns only
statement_id + status. This was designed for EXTERNAL_LINKS (where data is
fetched separately) but broke INLINE disposition where data is in data_array.

Changes:
- _transformDataArray now checks for data_array before routing to the
  EXTERNAL_LINKS path: if data_array is present, it falls through to the
  standard row-to-object transform.
- JSON format now explicitly sends JSON_ARRAY + INLINE rather than relying
  on connector defaults. This prevents the connector default format from
  leaking into explicit JSON requests.
- Connector defaults reverted to JSON_ARRAY for backward compatibility with
  classic warehouses (the analytics plugin sets formats explicitly).
- Added connector-level tests for _transformDataArray covering ARROW_STREAM
  + INLINE, ARROW_STREAM + EXTERNAL_LINKS, and JSON_ARRAY paths.

Co-authored-by: Isaac
Signed-off-by: James Broadhead <jamesbroadhead@gmail.com>
Some serverless warehouses return ARROW_STREAM + INLINE results as base64
Arrow IPC in `result.attachment` rather than `result.data_array`. This adds
server-side decoding using apache-arrow's tableFromIPC to convert the
attachment into row objects, producing the same response shape as JSON_ARRAY
regardless of warehouse backend.

This abstracts a Databricks internal implementation detail (different
warehouses returning different response formats) so app developers get a
consistent `type: "result"` response with named row objects.

Changes:
- Add apache-arrow@21.1.0 as a server dependency (already used client-side)
- _transformDataArray detects `attachment` field and decodes via tableFromIPC
- Connector tests use real base64 Arrow IPC captured from a live serverless
  warehouse, covering: classic JSON_ARRAY, classic EXTERNAL_LINKS,
  serverless INLINE attachment, data_array fallback, and edge cases

Co-authored-by: Isaac
Signed-off-by: James Broadhead <jamesbroadhead@gmail.com>
@jamesbroadhead
Copy link
Copy Markdown
Contributor Author

Closing in favour of #256, which addresses Mario's review feedback (simplified to 2 formats matching API enums: JSON_ARRAY and ARROW_STREAM).

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant