0dce7c8d18
* feat(engine): unify constraint validation across all write surfaces
Constraint enforcement (value/range/check, enum, uniqueness, edge
referential integrity, cardinality) was implemented three times — once
each in the bulk loader, the mutation executor, and the branch-merge
path — and had drifted: merge validated @range/@check but not enum, and
neither the mutation nor the load path enforced cross-version uniqueness
against already-committed rows.
Introduce one catalog-derived evaluator (`crate::validate`) that all
three surfaces route through. It is delta-scoped (checks only the change
set, not the whole graph) and index-backed (probes committed state
through the @key/@unique/src/dst BTREEs instead of full-scanning every
catalog table), reusing the existing leaf checks
(validate_value_constraints, validate_enum_constraints,
composite_unique_key) so the surfaces cannot drift again. A one-row-delta
merge now opens ~3 data tables instead of ~6+, and validation cost is
flat in graph size rather than O(V+E).
Behavior changes (all stricter, none relaxed):
- Enum constraints are now enforced on the merge path (was a gap).
- A write or load whose @unique value collides with an already-committed
different row is now rejected (cross-version uniqueness); re-upserting
an existing @key still upserts.
- Uniqueness distinguishes a duplicate key WITHIN one input batch (two
distinct records -> rejected, e.g. a bulk load listing a @key twice)
from the SAME id reappearing ACROSS batches (ordered supersession of
one logical row -> coalesced, e.g. a mutation insert-then-update).
- Overwrite loads validate per-table: a touched table's committed view is
its replacement image (empty), but a table absent from the batch keeps
its committed rows, so an edges-only overwrite still resolves
referential integrity against retained nodes.
Remove the per-surface validation orchestration the evaluator supersedes,
and the now-orphaned version-pinned dataset opener from the sealed
storage trait (reads route through the snapshot path). Docs (invariants,
testing) updated; full engine suite green.
* test(engine): pin orphan-edge validation on adopt-by-pointer merge
Regression for a gap in the unified merge validation: when a table is
adopted by pointer switch (AdoptSourceState) — source on main, target on a
branch — build_merge_changeset skips it, so referential integrity is never
checked for it. Merging main into a branch that deleted a node while main
added an edge to that node silently publishes the orphan edge.
This test merges main -> feature where feature deleted Bob and main added
Knows Alice->Bob, and asserts an OrphanEdge conflict. Red against HEAD
(merge returns Merged); turns green with the AdoptSourceState validation fix.
* fix(engine): validate adopt-by-pointer merge tables (AdoptSourceState)
The unified merge validator skipped any table classified AdoptSourceState
(a pointer switch / fork), so referential integrity, uniqueness, and
cardinality were never checked for it. Merging main into a branch that
deleted a node while main added an edge to that node silently published the
orphan edge — the prior full-scan validation caught it.
Root cause: classify_adopt keyed AdoptSourceState on the publish mechanism
("does it advance Lance HEAD") and returned before computing any delta, and
build_merge_changeset then skipped the table. Fix decouples the validation
input from the publish mechanism: classify_adopt now always computes the
source-vs-target delta (base == target on this path, so it is the right
validation delta) and carries it as AdoptSourceState { validation_delta };
build_merge_changeset validates it exactly like AdoptWithDelta. The publish
stays a pointer/fork (delta ignored) and remains excluded from recovery
pins, so publish/recovery semantics are unchanged — only validation is
restored. Closes the class: no publish optimization can bypass validation.
Turns the orphan-edge regression test green.
* test(engine): pin typed committed-uniqueness probe on non-String columns
The cross-version @unique check pushes a committed-state filter built from
the stringified key. On a non-String @unique column (e.g. Date) this compares
a Date32 column to a Utf8 literal — and the stringified key is the raw day
count, so the probe raises "Cannot cast string '20633' to Date32" for ANY
second write to the table (colliding or not).
Two regressions: a colliding Date value must surface a proper "@unique
violation" (not a coercion error), and a non-colliding write must succeed.
Both red against HEAD; green with the typed-literal probe fix.
* fix(engine): build committed uniqueness probe from typed column values
The cross-version @unique check pushed a Lance filter built with a
stringified key (lit(String)) against the real, typed column. On a
non-String @unique column this compared a Date32/numeric/bool column to a
Utf8 literal: a coercion error on Date/Bool (failing every write to the
table) or a silent miss on Float. For Date the stringified key was even the
raw day count, so the literal could never parse.
unique_holders now takes typed ScalarValues, built at the call site via
ScalarValue::try_from_array(group_column, row), so the pushed-down predicate
compares like-typed for any scalar @unique. The in-memory intra-delta dedup
keeps the stringified key (a type-agnostic equality grouping, unaffected).
Turns the Date @unique cross-version regression tests green.
* test(engine): pin id-keyed cardinality on merge-load edge moves/dups
Two cardinality drifts between validation and what commit persists:
- Move (B): a Merge-load that moves an edge to a new src only recounts the
new src, so vacating a src and dropping it below @card min is missed —
moving Alice's only WorksAt to Bob silently succeeds under @card(1..).
- Dup (A): a Merge-load batch listing one edge id under two srcs counts it
under both, but commit dedupes by id (last-wins). Alice gets a phantom
second edge and a spurious "has 2 edges (max 1)" violation under @card(0..1).
Both red against HEAD; green with the id-keyed last-wins cardinality model.
* fix(engine): key merge/load cardinality by edge id, last-wins
@card validation diverged from what commit persists in two ways: (1) it only
recounted the new src of a delta edge, so a Merge-load that moves an edge to a
new src never rechecked the vacated src and missed a drop below @card min; (2)
it counted raw delta rows, so the same edge id under two srcs in one batch was
counted under both, while commit dedupes by id (last-wins) — a phantom edge
and a spurious max violation.
evaluate_cardinality now coalesces the delta by edge id (last-wins, matching
dedupe_merge_batches_by_id) and builds the affected-src set from both the new
src of each delta edge AND the old committed src of each changed/deleted edge
id; a committed edge is dropped from its src when the delta deletes or
re-places it. The validated edge set per src now equals the committed image.
Turns the edge-move and duplicate-id cardinality regression tests green.
* docs(rfcs): add RFC 0001 — branch merge by fragment adoption
Proposed design for the by-design fix to merge cost/OOM: adopt the source
branch's Lance fragments by reference (base_paths) instead of re-materializing
rows, with a re-home reconciler + branch-delete reference guard closing the
dangling-reference lifecycle, and a reachability-complete cleanup sweep. Grounded
in the public Lance 7.0.0 multi-base APIs and the prior art (Delta shallow/deep
clone, Iceberg/lakeFS reachability GC). Status: Proposed.
* test(engine): pin @card validation on direct edge delete
Deletes stage as predicates, not constructive batches, so a delete-only
mutation produces an empty change-set and validate_changeset no-ops — a
`delete WorksAt where from = X` that removes a source's only edge commits
below @card(1..), while the merge path (which carries deleted_ids) rejects it.
Red against HEAD (the delete commits); green once the delete path resolves
its predicates into the validation change-set.
* fix(engine): validate edge cardinality on delete via resolved predicates
A delete-only mutation produced an empty change-set (deletes stage as
predicates, not constructive batches), so validate_changeset no-op'd and a
`delete Edge` that dropped a source below @card min committed silently — while
the merge path, which carries deleted_ids, rejects it.
validate_staged_mutation now resolves each staged delete predicate against the
live committed table (CommittedState::deleted_ids_matching, a SQL-filter scan
projecting id) and folds the matched ids into the change-set's deleted_ids for
that table. The existing evaluator then recounts the srcs a delete empties
(@card min) and sees removed rows for RI/node-delete — the same faithful
change-set the merge path already builds, so validation matches what commits.
Covers direct edge deletes, node deletes, and node-delete edge cascades
uniformly (all are staged predicates).
Turns the direct-edge-delete @card regression test green.
* refactor(engine): capture deleted ids at delete time, drop validation re-scan
The delete-cardinality fix resolved staged delete predicates a second time at
validation. Instead, capture the removed ids during the delete op's own scan:
execute_delete_edge and the node-delete edge cascade now scan id (not
count_rows), record the ids via MutationStaging::record_deleted_ids, and
to_changeset() folds them into the change-set's deleted_ids. validate_staged_
mutation reverts to plain to_changeset(); CommittedState::deleted_ids_matching
and scan_filtered_sql are removed.
Behavior-preserving (the @card-on-delete test stays green) and strictly fewer
scans — one scan at delete time replaces count-here + resolve-at-validation.
Node deletes already scanned their ids; this reuses that via a shared
ids_from_batches helper. Full engine suite green; workspace builds clean.
* test(engine): pin overwrite-removal RI + coalesced-unique final image
Two reviewer findings, both red against HEAD:
- F1 (High): overwriting a node table removes nodes without expressing them as
deleted_ids, so a retained edge in a non-overwritten table that references a
removed node is published as an orphan (edge-RI path-b never runs).
overwrite_node_removal_rejects_retained_orphan_edge.
- F2 (Medium): evaluate_unique accumulates superseded keys across batches, so a
mutation that frees a @unique value (Alice.email temp -> final) and reuses it
(insert Carol.email = temp) false-rejects a valid final image.
chained_unique_update_then_reuse_freed_value_is_not_a_violation.
* fix(engine): validate overwrite removals (orphan edges, emptied srcs)
An Overwrite load replaces each touched table, but to_changeset() only recorded
the new batch, never the committed rows the overwrite removes. So overwriting
node:Person to drop Bob while a retained edge:Knows(Alice->Bob) referenced him
published an orphan edge unchecked — edge-RI path-b is gated on the node's
deleted_ids, which were empty.
The loader now computes per overwritten table the removed ids (committed ids in
the pinned base minus the replacement batch's ids, via validate::
overwrite_removed_ids) and folds them into the change-set's deleted_ids. The
evaluator then runs RI path-b and cardinality against them — the same faithful
change-set the merge path builds. Overwrite is per-table, so a table absent from
the batch is untouched; a removed node referenced by a retained edge is now a
loud OrphanEdge.
Updates two tests that asserted the old silent-orphan behavior to
self-consistent overwrites (per-table Overwrite can't drop edge endpoints
without also overwriting the edge tables): end_to_end::overwrite_replaces_data
and writes::load_overwrite_with_bad_edge_reference_unblocks_next_load. The
orphan-rejection case itself is pinned by the new validators test.
* fix(engine): evaluate @unique against the coalesced final delta image
evaluate_unique iterated the raw delta batches and accumulated every key it saw
into one cross-batch map, so a coalesced write that frees then reuses a @unique
value within a query — update a row's email to 'temp', update the same row to
'final', insert a new row with 'temp' — false-rejected: 'temp' lingered in the
seen-set from the superseded first write though it no longer holds in the final
image that commits.
Restructure to validate the final coalesced image — the bytes that actually
publish:
- Pass 1 coalesces the delta by id (last-wins) into each id's final key, and
flags genuine within-ONE-batch duplicates (two distinct input records — the
bulk-load contract) before coalescing, so an unordered load batch with a real
dup still rejects.
- Pass 2 checks two distinct final ids holding the same key.
- Pass 3 does the committed cross-version lookup, excluding the delta's own ids.
Entries are sorted by id before the cross-row/committed passes so violation
order never depends on HashMap iteration. Coalescing first also drops the
redundant committed probes a superseded key used to issue.
Pinned by the chained-update red test; preserves intra-batch dup rejection
(consistency::loader_rejects_intra_batch_duplicate_keys) and cross-version
uniqueness (validators).
* style(engine): drop trailing blank line at staging.rs EOF
Left by a block-delete in an earlier refactor; flagged by git diff --check.
* docs(engine): refresh validate.rs module doc to current consumers
The module doc still said the merge path was the only consumer and the write
path a later, mechanical migration, and listed cardinality as a later
increment. Mutation and bulk load have since migrated onto the evaluator and
cardinality ships — correct both so the doc reflects that all three write
surfaces route through one evaluator.
26 KiB
RFC 0001: Branch merge by fragment adoption
| Status | Proposed |
| Author(s) | Ragnor Comerford |
| Discussion | — |
| Implementation | — |
Status is maintained by maintainers:
Proposedwhile the PR is open,Acceptedon merge,Declinedon close,Superseded by NNNNlater.
Summary
Make branch merge adopt the source branch's Lance fragments by reference
instead of re-materializing rows. Today, merging a branch whose table diverged
from the merge base copies every new/changed row through a temp Lance dataset
(stage_append + stage_merge_insert) and then rebuilds the table's indexes.
On embedding-heavy graphs this is the dominant merge cost and an out-of-memory
risk at scale. Because OmniGraph branches are Lance-native branches whose data
already lives durably under {table}/tree/{branch}/data/, the merged table can
instead be committed as a new table version whose data files point at the
source's fragments via Lance's base_paths mechanism — an O(metadata) commit,
no row copy, no hash-join, no synchronous reindex. A background re-home
reconciler then copies those referenced files into the target's own storage so
the source branch can be safely deleted, closing the only correctness hazard the
reference introduces.
Motivation
Branch merge has three per-table shapes (exec/merge.rs,
CandidateTableState):
AdoptSourceState— the target table is unchanged since the fork and the source is wholly adoptable. This is already metadata-only: a manifest pointer switch or a Lance-nativecreate_branchfork, zero row copy. No change is proposed here.AdoptWithDelta— the target is unchanged since the fork but the source added / changed / deleted rows. Today this copies the delta row-by-row into the target's storage. This is the target of this RFC.RewriteMerged— both sides changed; a real three-way row merge. Phase 2 (see Unresolved questions).
For AdoptWithDelta, the merged result is the source's table version (the
target equals the base, so there is nothing of the target's to preserve). Yet we
rebuild it row by row and then reindex. For a connector that incrementally
upserts embeddings, the changed set is large and each row carries a 3072-dim
vector; the copy + merge-insert hash-join + vector reindex is what hangs and
OOMs production merges.
The long-run liability argument: the row-copy machinery
(OrderedTableCursor, row_signature, compute_adopt_delta,
publish_adopted_delta) is already documented as transitional — to be removed
wholesale once the substrate offers a fragment-level branch merge. Investing in
patching its cost (e.g. a cheaper change-signature) spends effort on
deletion-bound code without removing the copy. Fragment adoption removes the
copy now, on public Lance 7.0.0 APIs, and deletes the transitional machinery
for the AdoptWithDelta path — it converges the design toward the same end
state the code already anticipates, rather than forking it.
Guide-level explanation
Nothing changes in the user-facing merge contract. branch merge produces the
same committed graph state, the same conflict kinds, and the same atomic
visibility. What changes is cost and latency:
- A merge whose source diverged by appends/upserts/deletes commits in time proportional to the fragment count, not the row count — typically a handful of metadata operations per table instead of a full table copy.
- The vector/FTS index is not rebuilt synchronously on the merge path.
- A new background maintenance step, re-home, runs as part of
optimize/cleanup(and is forced before a branch that the target still references can be reclaimed). Operators see it the way they see compaction: a convergence step that reclaims the "borrowed" layout into target-local files.
The one new operational rule: a branch cannot be reclaimed while the target
still references its fragments. Branch deletion (the logical operation) is
unchanged and immediate — the manifest flips authority as it does today — but the
physical reclamation of tree/{branch}/data is deferred until re-home has run,
exactly as branch-fork reclamation is already deferred to the cleanup
reconciler.
Reference-level design
Background: how Lance multi-base works (verified, Lance 7.0.0)
All required APIs are pub and constructible (no #[non_exhaustive]):
- A manifest carries
base_paths: HashMap<u32, BasePath>(lance-table/format/manifest.rs:103); aBasePath{id, name, is_dataset_root, path: String}holds a full URI to another dataset/branch root (manifest.rs:556). - A
DataFilecarriesbase_id: Option<u32>(lance-table/format/fragment.rs:56) indexing into that map; at read timeDataset::data_file_dir_for_base(dataset.rs:1954) resolves the file from the referenced base's store. Dataset::create_data_file(path, base_id)(dataset.rs:1861) reads an already-existing file's metadata and returns aDataFilestamped withbase_id— no copy.Operation::Overwrite{fragments, schema, initial_bases},Operation::Append{fragments}, andOperation::UpdateBases{new_bases}are public (transaction.rs:308-456). A registration step is required and it is a separate commit:initial_basesonOverwriteis honored only in create mode and is explicitly rejected on an existing dataset ("OVERWRITE mode cannot register new bases",transaction.rs:1802-1808), and a Lance transaction carries exactly oneOperation. So registering the base (UpdateBases) and committing the base-referencing fragments are two commits, not one. They are CAS-safe —UpdateBases/Append/Overwrite/CreateIndexeach conflict only with their own variant — and ride the per-table write queue. (Overwritealso clears the index section, so an adopt that uses it rebuilds indexes — see Component 6.)- Lance's own
shallow_clone(dataset.rs:2505,Manifest::shallow_clone:237) is exactly this recipe applied whole-dataset: stampbase_idon everybase_id == Nonefragment, register oneBasePath, clear the index section to rebuild. We apply the same recipe per table version, in place (a new version of the existing table dataset), whichshallow_clonedoes not do (it targets a fresh URI) — so we build it from the underlying primitives.
LanceDB was surveyed for a reusable higher-level pattern; it has none (it
delegates branches/clone/maintenance straight to lance and never touches
base_paths or branch merge). This is the right layer to build at.
Component 1 — TableStorage::stage_adopt_fragments (new sealed primitive)
A new staged primitive on the sealed TableStorage trait (storage_layer.rs),
shaped like the existing stage_* methods and consumed by the unchanged
commit_staged:
stage_adopt_fragments(
target: &SnapshotHandle, // target table dataset (== base)
source_branch_uri: &str, // {table}/tree/{source_branch}
source_fragments: &[Fragment], // source table version's fragments
) -> Result<StagedWrite>
It builds the referencing DataFiles via create_data_file(path, Some(base_id))
for each source fragment whose data is not already target-local, and produces a
two-commit staged sequence (per the registration constraint above):
Operation::UpdateBases{new_bases:[BasePath→source_branch_uri]}— register the source branch's data dir as a base on the target table dataset.Operation::Overwrite{fragments, schema}— replace the target table with the source's fragments (Overwrite becausetarget == base: the table becomes the source's image wholesale). The fragments carrybase_idfor the source-tree ones andNonefor those already target-local.
Both commits ride the existing commit_staged + per-(table, branch) write
queue + publisher CAS and are recovery-pinned together (the merge path already
does multi-commit-per-table under one sidecar — Component 3). No inline
HEAD/manifest coupling, so Invariant 1 holds by construction. (A
fragment-id-preserving variant — Update with removed_fragment_ids instead of
Overwrite — is required only if Component 6 Phase 2 index adoption is pursued,
since Overwrite clears indexes; Phase 1 uses Overwrite + rebuild.)
Note on lazy fork: a source that lazy-forked from the target already references
the target's root fragments (base_id → target root) for unchanged rows; only
the source's own tree/{branch} fragments (base_id == None, physically in the
branch tree) become target-referencing after adoption. So the cross-tree
reference surface — and therefore the re-home work — is bounded to the genuinely
diverged fragments, not the whole table.
Component 2 — merge integration
In exec/merge.rs, the AdoptWithDelta classification and its
compute_adopt_delta row walk + publish_adopted_delta copy are replaced: when
target == base, classify the table as adopt-by-reference and publish via
stage_adopt_fragments(target, source_branch_uri, source_version_fragments).
The merge already holds everything required — source_entry.table_branch,
table_version, and table_path (SubTableEntry) — to locate the source's
fragments. AdoptSourceState is untouched (already optimal). RewriteMerged
stays on the existing row path until Phase 2.
The per-table adopt commit still publishes through the unified merge manifest CAS
(one graph_commit), so cross-table atomic visibility (Invariant 2) is
unchanged.
Component 3 — recovery coverage
stage_adopt_fragments's commit advances Lance HEAD before the manifest publish,
so it is covered by the existing SidecarKind::BranchMerge recovery sidecar
(recovery.rs), which the merge path already writes with per-table version pins
and Phase-B confirmation. No new sidecar kind is required for the adopt step
itself; re-home (Component 4) runs inside optimize, so it rides the existing
SidecarKind::Optimize pin when in the same pass (a distinct SidecarKind::Rehome
is needed only if re-home runs independently of compaction).
Component 4 — the re-home reconciler (the load-bearing correctness piece)
This is where the design earns its RFC. Referencing a branch's fragments from the target creates a dangling-reference / data-loss trap that Lance does not protect against (verified):
- Lance's cleanup base-path protection is descendant-directional
(
cleanup.rs:655-868): it keeps files a descendant branch references, but a source branch is the target's ancestor (forked from it), so cleaning the source never consults the target's manifest — the target's referenced files look unreferenced and become GC-eligible. - Branch delete (
refs.rs:552) checks only descendant lineage, has zerobase_pathsawareness, thenremove_dir_all(tree/{branch})(refs.rs:611) — silently deleting files the target still needs. - There is no native re-home API:
UpdateBasesis add-only;compact_filesre-homes only fragments it happens to rewrite;deep_cloneis wholesale. We build re-home ourselves.
Re-home copies each target fragment whose DataFile.base_id resolves into a
branch tree into the target's own data/, then commits
Operation::DataReplacement{replacements} swapping the DataFile to
base_id = None. DataReplacement is chosen over Rewrite because it keeps
fragment ids and row ids stable (so scalar/vector indexes stay valid and no
bitmap remap is needed); the object-store copy goes through the storage adapter
(no raw FS I/O, per the deny-list). Re-home runs:
- as a background pass folded into
optimize(the natural maintenance home), and - forced on demand before a referencing branch is reclaimed.
Concrete integration (verified): re-home slots into optimize_one_table
(db/omnigraph/optimize.rs) as a Phase-B step — it needs the same &mut Dataset
handle (open_dataset_head_for_write(...).into_dataset()), runs under the same
per-(table, main) write queue, and advances HEAD before publish, so it can ride
the existing SidecarKind::Optimize pin when it runs in the same pass (a distinct
SidecarKind::Rehome is only needed if re-home runs independently of
compaction — decide by whether the two can be separated). It inherits
optimize's bounded maint_concurrency() budget and its refuse-on-unrecovered /
skip-uncovered-drift guards. Progress is reported by extending the returned,
#[non_exhaustive] TableOptimizeStats with bytes_rehomed / files_rehomed /
files_still_referenced (the last modeled on the existing pending_indexes
"work this pass could not finish, reported not fatal" field) — additive, no
breakage.
Component 5 — branch-delete / reclaim guard
The branch-delete reclaim path (force_delete_branch,
reconcile_orphaned_branches, cleanup_deleted_branch_tables) gains a
reference check: a branch tree is reclaimable only when no other branch's
manifest (the target's, in particular) holds a base_id reference into it.
While references remain, reclaim re-homes them first (Component 4) and only
then drops the tree. This is authority-derived and degrades to a no-op if Lance
ships an atomic multi-dataset branch merge — the same shape as the existing
branch-delete reconciler. The logical branch delete (manifest authority flip)
stays immediate; only physical reclamation waits, exactly as fork reclamation
already does.
Posture: prefer reachability-complete cleanup over a per-merge guard. The
prior art (Iceberg, lakeFS, Delta + Unity Catalog) makes "don't delete what any
ref still reaches" a property of GC computed over all references, not a guard
bolted onto each merge — and Delta's own evolution (classic shallow clone left
the FileNotFoundException exposed; the fix moved reference tracking into the
catalog) shows the per-merge guard is the early-stage posture. OmniGraph is
well-placed to do the reachability version because the manifest is its single
source of truth (Invariant 15): cleanup's existing per-table reconciler can
compute the live set as "every fragment any branch's manifest references,
following base_id" and refuse to delete anything in it — the same derive-from-
the-manifest shape as the index and orphaned-fork reconcilers. Build the reachability sweep from the start (verified recommendation). The
per-merge check and the reachability sweep are the same code at a different call
site, so deferring the sweep buys little and leaves a real cliff: the per-merge
guard protects only the reclaim path, not cleanup_old_versions itself, which
on an unrelated run would still GC a base-referenced file in a source tree (the
Delta "VACUUM-on-source" hazard the RFC cites). The integration point already
exists: reconcile_orphaned_branches (optimize.rs) — which cleanup_all_tables
already calls first — already enumerates all branches and caches each branch's
snapshot. Extend it to open each (branch, table) dataset at its pinned version,
union its manifest base_paths into a live set, then (a) thread that set into the
per-table cleanup_old_versions closure as a "refuse to delete in-set files"
filter and (b) into the reclaim decision so a tree with inbound references is
re-homed before force_delete_branch. Cost (document it): this adds
N_branches × N_tables per-table Lance manifest reads to cleanup — the same
kind of read the reconciler already does per branch, one level deeper — bounded
by the existing maint_concurrency() budget and the per-branch snapshot cache.
Concurrency (Q3, verified). The guard inherits the existing in-process-only
serialization (the per-(table, branch) write queue + manifest CAS) and the
documented "fork reclaim is in-process-safe only" gap: it reads manifest
authority, acquires the queue, re-validates under it, then re-homes + drops —
structurally identical to today's reconcile_orphaned_branches. Cross-process,
the worst case stays retry/contention, not data loss, provided the
reachability check + re-home always run before any remove_dir_all (the
manifest CAS still mediates the publish winner). No new primitive is needed for
single-process or one-winner-CAS topologies; a cross-process serialization
primitive (a lease on the schema-apply lock branch, or an engine-level
write_text_if_absent lease) must be designed with that existing gap before
multi-process write topologies rely on the guard — not separately.
Note the substrate gap this all works around: Lance cleanup is reachability-aware
only for within-dataset descendant branches (retain_branch_lineage_files,
base check scoped to base_path.path == self.dataset.uri), so the
ancestor-references-descendant case this merge creates (the target
referencing the source's data) and cross-dataset bases are not protected by Lance
today — the gap Issue 1 (below) raises upstream and Lance #7185 partly covers.
Component 6 — indexes (scoped)
Lance's shallow_clone clears the index section and rebuilds on access, and
Operation::Overwrite (Phase-1's data-adopt op) clears indexes too.
Phase 1 mirrors that: after an adopt, the index reconciler (ensure_indices
/ optimize, build_indices_on_dataset_for_catalog) rebuilds the table's indexes
as it does today (unchanged). This keeps Phase 1 correct and simple and adds no
index re-home surface, but does not remove the reindex cost — it defers it from
the synchronous merge path to the reconciler.
Phase 2 adopts the source's built indexes by reference, gated on evidence that
the reconciler reindex is the dominant cost for embedding-heavy tables. It is
feasible on Lance 7.0.0 (verified): IndexMetadata carries base_id +
fragment_bitmap, Operation::CreateIndex{new_indices} accepts a source index
verbatim, index files resolve from the base (Dataset::indice_files_dir), and the
bitmap stays valid because adopt preserves fragment ids. The staged-commit
mechanism already exists (our scalar-index staging builds CreateIndex via a
StagedWrite), so only a stage_adopt_index construction is new. But Phase 2
has real extra cost the evidence bar must clear: (a) index files become
base-referenced, doubling the re-home surface (and there is no IndexReplacement
op — re-home must copy the index files and re-commit CreateIndex with
base_id: None); and (b) it forces the data-adopt op off Overwrite (which
clears indexes) onto a fragment-id-preserving sequence (UpdateBases →
Update{removed_fragment_ids, new_fragments(base_id)} → CreateIndex), coupling
the two. Hence Phase 2 is deferred, not folded into Phase 1.
Invariants & deny-list check
- Invariant 2 (manifest-atomic visibility): preserved — adopt commits
publish through the unified merge CAS; one
graph_commitper merge. - Invariant 5 (recovery is part of the commit protocol): the adopt commits
reuse
SidecarKind::BranchMerge; re-home ridesSidecarKind::Optimize(it runs insideoptimize). No new HEAD-before-publish gap ships without sidecar coverage. - Invariant 7 / 15 (derived state, one source of truth): the base reference is a view into the source's fragments; re-home converges it target-local. No maintained shadow copy; the manifest stays the source of truth.
- Deny-list — "new write paths that advance Lance HEAD before manifest publish without a recovery sidecar": addressed (Component 3/4).
- Deny-list — "mutating immutable substrate state in place": not done — adopt and re-home both commit new manifest versions and write new local files.
- Deny-list — "raw filesystem I/O for cluster-stored state": the re-home
copy goes through the storage adapter /
TableStore.
New hazard explicitly closed: the cross-branch dangling-reference trap. Lance will not protect us (verified); Components 4 + 5 close it by construction. This is a new Known Gap only if the guard is not shipped with the adopt path — the RFC's position is that they land together.
Drawbacks & alternatives
- Do nothing. Incremental and three-way merges keep OOMing at scale (the FF case is already fine). Rejected — this is an active production failure mode.
- Symptomatic change-detection fix (cheaper
row_signaturevia content hash or row-version key). Investigated and rejected: it lands on deletion-bound code, and it only trims change-detection memory — it does not remove the row copy or the reindex, which are the actual OOM drivers. - Wait for native Lance branch merge. This is the cleanest end state and the
substrate is actively building it: Lance #7263 ("Branch merge and rebase",
open) specs this design — "graft the source branch's base into the target
branch's
base_paths… generalizesshallow_clone's base grafting from 'the whole manifest' to 'a single transaction's fragments,' replayed onto an existing target branch" — and depends on #7185 ("Can not delete branches referenced by other branches", open), which names the reclaim hazard and proposes the guard. These are likely Lance 8.x/9.x (LanceDB already pinslance 9.0.0-beta.8). The adopt design is the bridge: built so that when native branch merge lands,stage_adopt_fragments+ the re-home reconciler are removed, not reworked.exec/merge.rsalready anticipates this. - Re-home timing. Eager re-home at merge would defeat the speedup (it's the
copy again). Re-home-at-delete-only would make the first delete after a big
merge slow. The design does lazy-background re-home (in
optimize) with a forced fallback at reclaim — amortized, with a hard correctness backstop.
Prior art
This is a well-precedented pattern, not a novel one — which raises confidence in the shape and tells us where it converges:
- Delta Lake
SHALLOW CLONE→DEEP CLONEis the closest named twin. Shallow clone references the source's data files by path without copying (= our adopt); the docs carry the identical hazard ("runVACUUMon the source table → clients can no longer read those data files →FileNotFoundException") and the identical mitigations: deep clone (= re-home to independence) and Unity Catalog cross-clone reference tracking (= our reclaim guard, lifted into the catalog/GC). - Reflink / copy-on-write filesystems (
cp --reflink, btrfs/XFS/APFS): shared extents with refcounting; deleting one referer doesn't free shared blocks; writing/breaking the share copies. The same shape at the block layer. - Iceberg, lakeFS, Nessie, Dolt: merge by referencing shared immutable objects (metadata-only, zero-copy) and GC by reachability across all refs. Iceberg/lakeFS make "don't delete what any ref still reaches" a property of GC, not a per-merge guard — see the GC-posture note below.
Reversibility
Substrate-adjacent and partly irreversible — the reason this is an RFC. The
adopt commit puts base_id references from the target's manifests into branch
trees: an on-disk layout the readers must understand and the lifecycle must
protect. A single bad merge is recoverable (re-home or rewrite the references
away), but once branches are reclaimed in reliance on re-home, the layout
decision is committed. The merge contract (result, conflicts, atomicity) is
unchanged and the change is gated behind the existing staged-write + recovery
machinery, but the format/lifecycle dimension earns the up-front design review.
Unresolved questions
The lifecycle questions (cross-process safety, re-home throughput, GC posture)
were investigated against the code and are now resolved into the design above:
re-home rides optimize (Component 4), the reachability sweep is built from the
start in reconcile_orphaned_branches (Component 5), and the guard inherits the
documented in-process-only gap (Component 5, Q3). What remains genuinely open:
- Three-way (
RewriteMerged) — deferred, confirmed not feasible now. Investigated: our three-way path is row-at-a-time with no fragment identity, and Lance fragments are whole-or-nothing adopt units (any fragment containing a both-changed row must be rewritten). Fragment-granular partial adopt would need a different algorithm (per-fragment changed-row maps) and is a substantial redesign the substrate's native branch-merge (#7263) would supersede. Phase 1 leaves three-way on the existing row merge. Open: is it worth a Phase-3 attempt, or wait for #7263? - Index adoption — Phase 2, evidence-gated. Feasible (verified) but it doubles
the re-home surface to index files (no
IndexReplacementop) and forces the data-adopt offOverwriteonto a fragment-id-preserving sequence. Open: does the reconciler reindex actually dominate for embedding-heavy tables enough to clear that bar, or is Phase-1 rebuild sufficient indefinitely? - Cross-process write topologies. The guard is safe in-process and under
one-winner-CAS; before multi-process writers on one graph rely on it, a
cross-process lease (on the schema-apply lock branch, or an engine
write_text_if_absentlease) must be designed — together with the existing "fork reclaim is in-process-safe only" gap, not separately. - Upstream filing (filed). Two validated gaps, now raised upstream:
Lance #7514 (
cleanup_old_versionsdoes not protectbase_pathsreferences in the ancestor / cross-dataset direction — the live Delta "VACUUM-on-source" hazard; Lance protection is within-dataset-descendant only) and Lance #7515 (no in-place "materialize/promote shallow clone to independent" op —deep_clonemakes a new dataset,UpdateBasesis add-only). Closing these upstream lets this RFC's reachability sweep + re-home reconciler degrade to thin wrappers.