Skip to main content

Module update_reader

Module update_reader 

Source
Expand description

Per-file updates (Insert and Delete) for PK-index sink output. The planner, not the reader’s local history, determines whether a file is new or retained. This does not schedule phases or commit progress: callers must checkpoint cursors with output and enforce Delete-before-Insert.

Structs§

IcebergUpdateBatch
An empty chunk still carries read progress. EOF, not the last nonempty chunk, completes a task.

Enums§

IcebergUpdateReadMode
The planner must retain this classification and the parent descriptors across replay.

Functions§

newly_deleted_positions
read_file_updates
Read one whole immutable file with bounded row batches. SDK _pos is used internally and removed before creating logical chunks. No shredded-VARIANT-to-NULL fallback is allowed here. Resume rereads from the beginning; seeking is an optimization, not part of the cursor contract. The caller supplies a reader for this table’s FileIO and controls its cache lifetime. Sequential tasks can reuse it after dropping the returned stream; parallel tasks need separate readers.