v0.13.3

Features

• Jupyter widget UI on JupyterLab - Progress bars and DataFrame views now use ipywidgets (>= 7), with prebuilt JupyterLab and classic Notebook extensions. options.display.notebook_widget defaults to None, which auto-enables widgets except on DSW and VS Code; set it to True or False to override.
• Cross-process MaxStorage session descriptors - Sessions can be exported with to_descriptor() and restored in another process through descriptor=, without re-creating them on the server. Write descriptors support only WriteMode.BATCH_COMPATIBLE.
• Extended Arrow IPC reads and v2 capability fields - Read sessions can use extended Arrow IPC, which delivers data and metrics together; extended_arrow_ipc_enabled reports it and open_arrow_reader handles it transparently. record_count returns -1 when unknown, and latest_txn_id supports exactly-once incremental reads.
• Vector type support - Vector values are now handled by the CSV and Arrow readers and by the MaxStorage schema parser, including validation.
• Opt-in write timeout - Adds the write_timeout option and WriteTimeout exception for streaming uploads. Behavior is unchanged unless the option is set.
• Batch-compatible commit wait timeout - TableWriteSession.commit() accepts wait_flying_writers_timeout_seconds (1–600) to bound the server's wait for in-flight writers.

Enhancements

• Small-file merge - Adds Table.merge_small_files() and Partition.merge_small_files().
• Partition iteration limits and Table.is_partitioned - iterate_partitions() accepts max_items (partitions per request) and limit (total partitions), and Table.is_partitioned reports whether a table is partitioned.
• Unknown instance and task statuses - Unrecognized statuses now map to UNKNOWN instead of raising ValueError.
• Single-request upload for small resources - Small file resources are uploaded in one request, while non-regular files such as pipes and sockets use the chunked path.
• Faster import odps - Importing odps no longer imports SQLAlchemy, IPython or the widget stack; optional dependencies load only when needed.
• Widget bundles built outside cibuildwheel - Wheel builds now build the widget bundles in a separate Node 22 step, keeping JS tooling out of the manylinux images.

Bug Fixes

• HTTP sessions rebuilt after a fork - RestClient and SignServerAccount now rebuild their HTTP sessions in forked children instead of reusing the parent's connections.
• Error responses keep status and request id - Errors from empty or malformed responses now retain status_code, request_id and response headers.
• Half-read row batches release their stream - Row batches that fail midway now close their decoder and release the HTTP connection immediately.
• Resource pack loading with environment credentials - load_packages_in_subprocess now re-enters the original script, so environment credentials are inherited.
• Older pyarrow compatibility - Temporal validation now works with older pyarrow and pandas versions and reports out-of-range values clearly.

Compatibility Notes

• options.display.notebook_widget defaults to None - Auto-detection replaces the previous True default; set it explicitly to force widgets on or off.
• odps.ui.common.html_notify removed - The notifier and ProgressUI.notify are gone, and setup/teardown magics no longer show notifications.
• MaxStorage descriptor restrictions - descriptor cannot be combined with table or session_id, and to_descriptor() supports batch-compatible write sessions only.
• Lazy SQLAlchemy registration - The odps dialect is registered only when SQLAlchemy is already imported; in an uninstalled source tree, import odps.sqlalchemy_odps before create_engine().

Docs

• Added an FAQ on preserving exact decimal values with pandas.read_csv. Thanks @dingxin-tech
• Documented lazy SQLAlchemy dialect registration and added examples for Table.is_partitioned and iterate_partitions(max_items=..., limit=...).