Make SharedString snapshot format selection sticky - #28370

How contribute to this repo.

Guidelines for Pull Requests.

Description

Make SharedString's flat MergeTree snapshot format opt-in and sticky across summaries and loads. Select the next DDS summary format using explicit configuration/runtime flag, recorded flag, then legacy default. Flat summaries persist newMergeTreeSnapshotFormat: true in the DDS .attributes; legacy summaries use undefined, so the serialized field is omitted. An explicit false overrides recorded true and clears the remembered flat setting when the DDS next summarizes naturally.

Use the same effective setting for writer selection and internal bookkeeping, with per-instance attributes to prevent settings leaking between DDSs. Do not dirty an unchanged DDS or force a summary when the setting changes. Leave both snapshot formats, their readers, SharedMatrix, and snapshotFormatVersion unchanged.

Add option documentation, a changeset, attached/detached unit coverage, four SharedString fuzz workloads for inherited settings and explicit overrides, and focused local-server summary/reload tests. DDS fuzz summaries capture the real .attributes blob after summarization and load those snapshot-time attributes through joining, detached rehydration, and stashing. Fuzz assertions observe summaries already generated by the harness rather than introducing extra summaries that advance live client state. The conflict-farm additions have been removed because they bypass SharedString's DDS attributes and format-selection logic.

Omitting the default legacy flag preserves the existing serialized attributes shape. The snapshot-normalizer workaround and its tests have been removed; golden fixtures remain unchanged.

Breaking Changes

None. Both formats remain supported, public API signatures are unchanged, and SharedString defaults to legacy when neither an explicit nor recorded setting exists. Previously recorded false values are still accepted when loading.

Reviewer Guidance

The review process is outlined in the pull request guidelines.

Please focus on explicit-false precedence, clearing versus preserving the serialized flag, per-instance attribute isolation, snapshot-time metadata preservation through fuzz client loads, and agreement between the emitted format and recorded setting. Format changes take effect only when the DDS is naturally summarized, including normal summary-handle reuse for unchanged channels.

Validation

The latest simplification was published before running checks, following the author's requested review workflow. The following checks passed on the current revision. Test commands used FLUID_TEST_MODULE_SYSTEM=ESM; fuzz commands also used FUZZ_STRESS_RUN=short.

• ESM build: from the repository root, pnpm exec fluid-build --root . --task build:test:esm '^@fluidframework/sequence$' '^@fluidframework/merge-tree$' '^@fluidframework/tool-utils$' '^@fluid-private/test-end-to-end-tests$' '^@fluid-internal/test-snapshots$' — passed.
• Precedence and persistence units: from packages\dds\sequence, with MOCHA_SPEC=lib\test\snapshotFormat.spec.js, npm run test:mocha:esm -- --reporter dot — 65 passed, including serialized omission of the legacy flag and clearing a recorded flat setting.
• SharedString snapshot-format fuzz: from packages\dds\sequence, with MOCHA_SPEC=lib\test\fuzz\sharedString.snapshotFormat.fuzz.spec.js and FUZZ_TEST_COUNT=100, npm run test:mocha:esm -- --reporter dot — 400 passed (100 seeds per workload).
• Local server: from packages\test\test-end-to-end-tests, with MOCHA_SPEC=lib\test\sharedStringSnapshotFormat.spec.js, npm run test:realsvc:local -- --compatKind=None --reporter dot — 2 passed, checking the actual attributes blob and summary content through reloads.
• Golden snapshot comparison: from packages\test\snapshots, npm run test:mocha:esm -- --grep '^Snapshots writes snapshot in correct format$' --reporter dot — 1 passed, with zero mismatches across all current fixture datasets and no special normalization for the new flag.
• Targeted ESLint and Biome checks on the changed source files — passed.
• From the repository root, pnpm policy-check --path '^([.]changeset.sticky-shared-string-snapshot-format[.]md|packages.dds.(sequence|merge-tree)|packages.test.test-end-to-end-tests)' — passed.

On the preceding fuzz-harness revision, the existing SharedString fuzz workloads also passed 497 cases with 3 pre-existing skips using npm run test:mocha:esm -- --reporter dot with MOCHA_SPEC=lib\test\fuzz\sharedString.fuzz.spec.js and FUZZ_TEST_COUNT=100. The complete DDS fuzz harness passed 41 cases using npm run test:mocha:esm -- --grep '^DDS Fuzz Harness' --reporter dot from packages\dds\test-dds-utils.

Broader CI-readiness and API-report generation checks have not been run locally.