Overview

Write a design note on staging writes in the RecoveryUnit instead of writing through to WiredTiger, and on merge cursors that present storage plus staged writes as a single view. This guides the staged-write prototype tickets in this epic.

Background

In the proposed architecture, writes on opted-in collections are buffered in the RecoveryUnit until commit and are not written to the storage engine on the primary. A transaction must still read its own writes, including through collection and index scans, so storage reads must merge the snapshot with the staged writes and produce the same results as a write-through transaction would.

Scope of Work

• Staged buffer structure: per-ident ordered map keyed by KeyString/RecordId with tombstones.
• Interaction with savepoints and statement-level rollback in multi-document transactions.
• RecordId allocation for inserts without writing to WiredTiger.
• Count and data-size delta tracking.
• Cursor semantics: seek, next/prev, save/restore, bounds, prepare conflicts.
• Inventory of every RecordStore, SortedDataInterface and Container entry point, including non-cursor ones (findRecord, dataFor, seekExact, numRecords, dup-key checks, Container get/insert). Classify each as: merge needed / buffer only / unsupported in prototype.
• Memory limits (reuse existing transaction size limits where possible).

Acceptance Criteria

• Design note recorded on this ticket or committed alongside code in the project branch.
• Inventory table complete, with open questions listed.

Technical Notes

• Prototype quality; not intended for commit to master.