Batch standalone multi-updates by write-set capacity

[Meowooh]

A standalone updateMany currently fails with TransactionTooLarge when its accumulated document and index writes exceed the 62 MiB transaction write-set limit. This change executes ordinary multi-updates in capacity-bounded transactions. A batch commits when the next complete document plus its index writes cannot fit, or when the scan ends; crossing a cursor prefetch window does not trigger a commit.

The executor owns transaction boundaries, saves and restores scan state, and re-reads pending documents after a transaction change. It confirms candidate IDs and statistics only after a successful commit, and retries only the uncommitted batch after a definite write-conflict abort. Index admission and execution share the same encoded-byte accounting, including multikey, partial unique, and creating-index writes.

Behavior to review

• A later error rolls back the current batch while preserving earlier committed batches. Error responses retain the existing write-error representation; a failed entry's count is not a partial-progress API.
• Explicit transactions, retryable writes, and existing outer atomic operations keep the original transaction path. Ordered/unordered update entries and legacy OP_UPDATE use the same batching rules where eligible.
• Unknown commit outcomes are surfaced without replaying a possibly committed non-idempotent update. A single document plus its index changes that exceeds the hard limit still fails.
• Cursor restore rejects external schema changes while accepting this operation's own committed multikey metadata.

The change is based directly on upstream main 6959f0c79abf5eda5347f7301bced8c7bff716e2. It uses the existing modifier-update interfaces and adds no aggregation pipeline-update support. Pipeline-dependent regression scenarios have been rewritten with ordinary modifiers while retaining schema and batch-failure coverage.

Dependency

Depends on eloqdata/tx_service#574 for the byte-accounting APIs. The submodule pin contains only that change on top of the current EloqDoc main pin; .gitmodules continues to use the upstream GitHub URL. If the dependency is squash-merged, update the pin to the accepted upstream commit before merging this PR.

Validation

• Six JS regression scripts pass syntax checks; clang-format 18.1.8 passes for the complete Eloq module; shell syntax and whitespace checks pass.
• The three adapted regression scripts (capacity/index movement, schema changes, and batch failures) pass against the existing downstream batching runtime. This is a test-adaptation check, not proof of the upstream build.
• Six new JS scripts are collected by eloq_basic / eloq_core. Two C++ accounting tests are explicitly built and executed by the existing CI build phase, before build-output cleanup.
• Full upstream-workflow validation is running at https://github.com/Meowooh/eloqdoc/actions/runs/34807034571; exact upstream build and runtime results are pending.

Reviewer focus: transaction ownership and partial completion, retained-ID replay and deduplication, exact document/index byte accounting, and schema lifetime across commits. Reverting this PR restores command-level transaction handling for standalone multi-updates.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}