Optimizations copy-on-write move points out of their source segments in
memory; WAL replay can only re-derive those moves from the sources'
on-disk pre-images. Destroying the replaced segments' files right at the
swap breaks that: the moved copies may still sit unflushed in appendable
segments, so a restart before they are persisted loses the points.
Instead of dropping source data in finish_optimization, retire the
segments. Their files survive until a flush proves the durable waterline
covers the optimization, at which point flush_all destroys them. While
the files remain on disk the WAL acknowledge is capped at the wrapped
segment's persisted version, so every operation the files contradict
(deletions in particular) is replayed and re-applied on restart. A crash
in the meantime loads the retirees next to their replacement and
load-time deduplication resolves the overlap.
Optimizer tests that assert source files are gone now flush first to
mature the retirees before counting/asserting.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* override `for_values_map` for `OnDiskMapIndex`
* refactor and apply to filter as id iterator
* add comment
---------
Co-authored-by: generall <andrey@vasnetsov.com>
* feat: feature-flagged segment manifest for LocalShard
Add an on-disk segment manifest (`segments/manifest.json`) that lists a
shard's segments and their state, so out-of-process readers (e.g. a
read-only follower, possibly over object storage) can discover segments
without scanning the filesystem. Gated by the new `write_segment_manifest`
feature flag (off by default).
shard: define the structure + helpers (`SegmentsManifest`,
`SegmentManifestState`, `from_segment_holder`) in a new `segment_manifest`
module, plus the `SEGMENT_MANIFEST_FILE` constant and path helper. The
manifest is a flat `{ "<uuid>": "<state>" }` map; only `active` is written
today, with `under_construction`/`retiring` defined so the format can be
extended without breaking compatibility.
collection: LocalShard owns the writing logic. The manifest is persisted
via `SaveOnDisk<SegmentsManifest>`, initialized from the live segment set
on load/build and refreshed by the optimization worker whenever the
segment set changes (the helper re-derives from the holder and no-ops when
unchanged). No changes to `lib/shard/src/optimize.rs` internals.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Emj8TFxdtrf3K32eWhGgor
* feat: make append_only_mutations a proper feature flag
Replace the debug-only `QDRANT_APPEND_ONLY_MUTATIONS=1` env-var escape
hatch in segment construction with a `FeatureFlags::append_only_mutations`
flag, so it works in release builds and is configurable like the other
flags (config / `QDRANT__FEATURE_FLAGS__APPEND_ONLY_MUTATIONS`).
Deliberately left out of `FeatureFlags::all()`: it changes mutation
semantics and `all` is enabled in dev and e2e configs, so it stays
explicit opt-in.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Emj8TFxdtrf3K32eWhGgor
* upd openapi schema
* feat: register segments in manifest via must-use token, holder builder
Wire segment-manifest maintenance through the segment lifecycle so a
newly created segment is registered as soon as it exists on disk, and
construction can't silently skip it:
- build_segment now returns a #[must_use] NewSegmentToken carrying the
new segment UUID; the lint forces callers to register or drop it.
- SegmentHolder owns the manifest and reconciles it on sync; new
segments are registered ASAP (even before being added to the holder)
via the token, before they can receive writes.
- SegmentHolderBuilder is the only way to obtain a shard's holder; its
build() wires up the manifest, so it can't be forgotten. init/set
manifest helpers are now private / test-only.
- Optimization registers the optimized segment before dropping the
superseded segments' data; deletion is intentionally lenient.
- Document the consistency assumptions on SegmentsManifest: it is a
superset-biased view that may list not-yet-finalized or already-deleted
segments, which readers must tolerate.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add two-phase config reload to `ReadOnlySegment`, alongside the existing
live-reload:
- `config_reload_diff(&self, fs)` re-reads the on-disk config, diffs it
against the in-memory config, and eagerly loads every new or changed
component (dense/sparse vector storages, indexes, quantized vectors and
payload field indexes) under a shared `&self` borrow, so the segment
keeps serving reads while loading.
- `apply_config_reload(&mut self, diff)` installs the pre-loaded
components and drops removed ones under `&mut self` — a cheap swap with
no I/O, so the exclusive borrow is held only briefly.
A vector or field whose config changed is reloaded (drop + load). The
payload-index half lives on `ReadOnlyStructPayloadIndex` with the same
diff/apply split, plus register/unregister helpers for its `has_vector`
storage map.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
`version()` is the segment's update version — only meaningful for the
mutable/storage path. Every real caller already reaches it through
`StorageSegmentEntry` (flush), `SegmentEntry` (update), or a concrete
`Segment`/`ProxySegment`; none use it through the read-only base trait.
Move the declaration down to `StorageSegmentEntry` and relocate the
`Segment`/`ProxySegment` impls accordingly. `ReadOnlySegment` no longer
needs it, so drop the method and the write-only `version`/`initial_version`
fields it carried (`live_reload` never refreshed `version`, and neither
field was ever read).
This leaves `ReadSegmentEntry` a clean read-only surface and stops a
read-only segment from exposing a meaningless (stale) version.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat: immutable field index live_reload
* feat: read-only segment live_reload
* fix: linter
* fix: linter
* fix: review comments
* fix: linter
* feat: ReadOnlyVectorData::live_reload covering all components
Extract the per-vector reload into a dedicated method that destructures
`ReadOnlyVectorData` so storage, index and quantized vectors are all
covered. Adding a field without reloading it won't compile, which guards
against silently skipping a component. The segment orchestrator now just
delegates per named vector.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix: keep id-tracker delta pending until all reloads succeed
ReadOnlySegment::live_reload drained the id-tracker delta (which advances
tracker state and cannot be replayed) and only then ran the fallible
payload/vector reloads. On error the delta was lost, so un-updated
components drifted out of sync permanently.
Accumulate the delta into a new `pending_reload` field and clear it only
once every component has reloaded successfully. On a later reload the
tracker's fresh delta is folded in via `LiveReloadResult::merge` and the
union is replayed, so a partial failure self-heals. Offsets are monotonic,
so the only merge conflict is an inserted-but-unapplied offset later
deleted: it is dropped from `inserted` and kept in `deleted` so a
partially-applied component drops it on replay.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: generall <andrey@vasnetsov.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat(facet): add sampling strategy for high-cardinality fields
For approximate facet queries the current per-segment implementation
walks every distinct value in the field index — O(unique_values_count)
even when the user asks for a tiny top-K. On UUID-style fields with
millions of unique values this dominates the request latency even
after #9208 capped the cross-shard payload.
This commit adds a parallel sampling strategy that runs in O(limit)
instead of O(unique_values_count):
1. Phase 1 — iterative novelty sampling. Stream point IDs in random
order (filtered if requested), look up each point's value via the
facet index, and collect distinct values into a candidate set until
`limit * 10` (min 1000) candidates have been gathered. Uses a
batch size of 32 to amortise the inner `for_points_values` call,
and bails out early after 128 consecutive empty batches when the
long tail is too thin to keep finding novel values.
2. Phase 2 — exact-count post-pass. For each candidate value, compose
`field == value` with the user filter and count via the payload
index. This guarantees the returned counts are exact (matching the
semantics of the full-scan path); only the *set* of returned values
is approximate.
The two strategies live side by side; `SegmentReadView::approximate_facet`
picks between them per-request based on
`unique_values_count > limit * FACET_FULL_SCAN_FACTOR` (FACTOR = 4).
Below that, the existing scan path runs unchanged — it'd visit most of
the index either way, and the post-pass adds no value.
The Monte-Carlo simulation behind this design (see thread context for
Zipf-distributed fields with cardinality up to 10^5 in ~1000 samples,
and trivially-correct results on UUID-style fields where every value
has count 1.
Adds a new `unique_values_count` method on the `FacetIndex` trait
(implemented for `MapIndex`, `ReadOnlyMapIndex`, `BoolIndex`,
`ReadOnlyBoolIndex`, and the `FacetIndexEnum` dispatcher) so the
strategy switch can run without touching the index.
Co-authored-by: Cursor <cursoragent@cursor.com>
* [AI] simplify, use single file
[AI] better selection of filtering approach
fmt
[AI] simplify, use single file
* manual simplification
* [AI] implement candidate-based lookups
[AI] 🧹
* precollect filter into bitmap
* avoid sampling with restrictive filter
* fix rebase + clippy
* refactor tests
* no duplicate values in map index
* polish comments
---------
Co-authored-by: root <111755117+qdrant-cloud-bot@users.noreply.github.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
`GridstoreReader`)
- Make the entrypoint of payload storage choose the `Populate` variant
- Mutable payload indexes now populate gridstore blockingly before using
it to load
- Propagate `Populate` into `flags` module
- Add `UniversalRead::populate_auto` to know whether a backend chooses
to populate or not when `Populate::Auto`
* feat(bm25): add explicit Disabled stemmer; deprecate language hack
Adds a `Disabled` variant to `StemmingAlgorithm` (`stemmer: {"type": "none"}`)
so stemming can be turned off explicitly in both the main engine and Edge,
instead of relying on the undocumented `language: "none"` footgun that
silently disabled both stemming and stopwords.
For language-neutral text processing the supported setup is now:
1. set the stemmer to disabled, and
2. configure an empty stopword set.
The main engine still tolerates unsupported languages (so existing
`language: "none"` configs keep working on upgrade) but now logs a
deprecation warning pointing users to the explicit setup. Edge continues
to reject unsupported languages, and now has a real way to disable stemming.
Refs: https://github.com/qdrant/qdrant/issues/9289
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(edge-py): handle Disabled stemmer in python bindings; fix openapi schema
- Handle the new StemmingAlgorithm::Disabled variant in the qdrant-edge-py
bindings (FromPyObject/IntoPyObject/Repr) and add a DisabledStemmer pyclass
plus its .pyi stub entry.
- Match generator output for the StemmingAlgorithm OpenAPI schema (plain $ref
in anyOf) so docs/redoc/master/openapi.json stays consistent.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(openapi): regenerate StemmingAlgorithm schema with generator output
Ran tools/generate_openapi_models.sh so docs/redoc/master/openapi.json
exactly matches generator output: DisabledStemmerParams/NoStemmer are placed
after SnowballLanguage, and the StemmingAlgorithm anyOf entry is a plain $ref
(the schema2openapi step flattens the allOf+description wrapper).
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(test): avoid wildcard enum match arm in bm25 sparse_len helper
clippy --all-targets flags `other => panic!()` as wildcard_enum_match_arm;
match the Dense/MultiDense variants explicitly instead.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix: issues
* fix: log::warn as call once
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Daniel Boros <dancixx@gmail.com>
* test(model_testing): add basic Query API coverage
Add an Op::Query variant that exercises the unified Query API with a
plain Nearest scoring query (ScoringQuery::Vector(QueryEnum::Nearest)),
routed through collection.query. Verification mirrors the existing
Search op: exact dense / multi-dense scans must return exactly the
top-k, sparse and approximate paths are an upper bound, and every
returned id must exist in the model with the queried vector populated
and matching the optional num filter.
Co-authored-by: Cursor <cursoragent@cursor.com>
* test(model_testing): unify Search/Query invariant check
Extract the shared candidate set, strict-vs-upper-bound decision,
per-result membership checks, and the rich diagnostic probes
(retrieve/count/scroll/retry) into helpers reused by both apply_search
and apply_query. apply_query now gets the same strict top-k check plus
failure diagnostics it previously lacked.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat(model_testing): add --duration-sec to run for a fixed wall-clock time
Bound the soak by wall-clock time instead of op count: when --duration-sec is set, the loop runs until the deadline (or Ctrl-C) and --op-num is ignored.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Fix clippy
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Extend ScrollFilter with a HasVector variant so paginated scroll
exercises Condition::HasVector. The matcher targets an active vector
name, and the model verifier checks it against each point's populated
vector set (which varies via DeleteVectors / partial UpdateVectors),
restricting results to a known, model-checkable subset.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Buffer multi-dense offsets store to prevent reload corruption
The appendable multi-dense storage flushes its `vectors` and `offsets`
chunked stores independently. Each flusher snapshots its `status.len` at
creation time but msyncs chunk bytes at execution time. A re-upsert that
grows a point past its reserved capacity rewrites the point's `offsets`
entry in place to a freshly-appended row region; if that lands in a
flush's creation-to-execution window, the relocated entry becomes durable
while the `vectors` store's recorded length still predates the rows it
references. On reload that point is unreadable and a WAL append reuses the
rows, clobbering another point.
Wrap the offsets store in a write-back buffer (`BufferedOffsets`) so the
durable offsets can never reference rows beyond the durable `vectors`
length. Offset writes stage in a pending overlay and only land in the
durable store while a flush executes; the flusher snapshots the pending
set at creation time, so any write after that stays buffered for the next
flush. Both flushers snapshot at the same instant, yielding a consistent
durable cut. Rows written after the cut are unreferenced garbage the next
append overwrites. This prevents the skew at the source instead of
patching it on reload, and also closes the residual offsets-length smear.
The buffer follows the Gridstore flusher convention: pending writes live
inside the single lock-guarded store, reads consult the overlay then the
durable bytes under one lock, and the durable msync runs after releasing
the write lock so it never stalls concurrent reads.
Enable the "m" multivector in the collection model test now that its
reload divergence is resolved.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Narrow offsets flush write-lock scope
Build and sort the pending snapshot before taking the write lock; only the
apply + reconcile need it. Shortens the lock hold so concurrent reads block
less. Addresses review feedback.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* perf: use Entry API to avoid redundant map double-lookups
Replace get_mut/contains_key followed by insert with the entry API
across several maps, collapsing two hash lookups into one. Limited to
sites where the key is Copy or already owned and moved, so no extra
key clone is added to any hot path.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* More Entry API usage in mutable_geo_index
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: xzfc <xzfcpw@gmail.com>
* test(model_testing): add has_id matcher to generated scroll filter
Extend the ScrollPaged filter selector with a HasId variant so the
generated filter exercises a has_id matcher (restrict to an explicit
point-id set), in addition to the existing none / num / tag filters.
The id set mixes ids present in the model with ids drawn from the id pool
that may not be, so the matcher meaningfully restricts. The model mirrors
it with a set-membership predicate, and the existing paged-scroll
id-set assertion validates the engine result.
Co-authored-by: Cursor <cursoragent@cursor.com>
* chore: fix clippy lints (wildcard_enum_match_arm, from_iter_instead_of_collect)
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(model_testing): clamp has_id sample count to id_pool
Addresses CodeRabbit review: random_distinct_ids could spin forever when
the requested distinct count (up to 15) exceeds id_pool. Clamp to
id_pool.min(15). Only relevant for tiny --id-pool values; the default
pool (500) is far above 15.
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Add a ScrollPaged verification op that scrolls the (optionally filtered)
collection in pages of a small limit, following next_page_offset until
exhausted. The other scroll ops always read everything in one page
(offset: None, limit: usize::MAX), so the offset cursor and a real
(non-MAX) limit had no coverage.
Asserts each page holds at most `limit` points, no id repeats across
pages, and the union of all pages equals the model's expected id set for
the chosen filter (none / num == X / tag == X). Includes a page-count
guard against a stuck cursor.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add a RetrieveSelective verification op that exercises the
with_payload/with_vector selector forms, which the soak test previously
never used (every read passed Bool only).
- with_payload covers Bool, Fields, Selector::Include, Selector::Exclude.
- with_vector covers both Bool forms and Selector(names) over a subset of
active vector names.
The verifier asserts the engine's returned payload equals the model entry
filtered by the engine's own PayloadSelector::process, and the returned
vectors equal the requested name subset that the point actually has
populated.
Co-authored-by: Cursor <cursoragent@cursor.com>
Add a SetPayloadByKey op to the model_testing soak test so the keyed
set-payload path (SetPayloadOp.key) is exercised. Previously every
set-payload op passed key: None, leaving the key-scoped assignment path
(merge_by_key / JsonPath::value_set) uncovered.
The op samples existing point ids, a random payload, and a single
top-level schema field as the key. The model mirrors the engine via
Payload::merge_by_key (the same JsonPath::value_set the engine's
set_by_key uses), keeping model and engine in lockstep.
Co-authored-by: Cursor <cursoragent@cursor.com>
* Add routing token structure
* Implement routing token in read operation executor as per design doc
* Add TODO to glue routing token to user requests
* Implement routing header for REST API
* Source routing token from request, not from JWT token
* Implement routing token in gRPC API
* Add test
* Review remarks
Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>
* Use lower case header name to prevent panic
* Rename header to X-Qdrant-Route-Affinity
* Assert routing consistency in test on all peers
---------
Co-authored-by: coderabbitai[bot] <136622811+coderabbitai[bot]@users.noreply.github.com>
* Add DynConditionChecker type alias
* Add DynConditionChecker type alias (point_scorer.rs)
* Move trait ConditionChecker to `common`
Reason: will be used both in `segment` and `sparse` crates.
Add a new named dense vector "i" to the model_testing fixture configured
with HNSW inline_storage backed by scalar quantization, so the soak harness
exercises the inline-storage index layout. It behaves like a normal dense
vector to the model; only the on-disk HNSW layout differs.
Co-authored-by: Cursor <cursoragent@cursor.com>
The 6 turbo vector storage model tests are flagged SLOW by nextest on
Windows CI (>60s, some >240s):
- vector_storage::turbo::tests::turbo_model_test_random_ops_{dot,cosine}
- vector_storage::turbo::multi::tests::turbo_multi_model_test_random_ops_{dot,cosine}
- vector_storage::turbo::test::congruent_random_ops_{dot,cosine}
These are dim x seed x ops sweeps; Windows runners are several times slower.
Lower SEEDS_PER_CELL on Windows only via #[cfg(windows)] so the runs stay
within the slow-timeout, while keeping full coverage on Linux/macOS.
Co-authored-by: Cursor <cursoragent@cursor.com>
* feat: add open for read-only sparse vector index + enum sparse dispatcher
* fix: universal-IO loads for read-only sparse index open
* refactor: rename load_via/open_via to load_universal/open_universal
* refactor: drop StorageVersion::load in favor of load_universal
The regular-IO `load` duplicated `load_universal` over plain `File` IO.
Remove it and route all callers through `load_universal(&MmapFs, ..)`.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* refactor(sparse): decouple inverted index from concrete storage; split segment_constructor_base (#9461)
Make the read-only index enum generic over storage `S` (no concrete MmapFile),
and remove construction callbacks from the sparse index open paths.
- InvertedIndex is now a pure read/search trait: `open`/`from_ram_index`/
`type Fs` moved off the trait to inherent methods on each concrete index
type, so construction no longer requires `S::Fs: Default`.
- SparseVectorIndex open split into a generic `plan` (load-vs-build decision +
RAM-index build) and generic `finish` (assembly); callers do the concrete
per-type construction, so no construction callbacks are needed.
- ReadOnlySparseVectorIndex::open takes the already-constructed inverted index
and caller-loaded config instead of a `load_inverted_index` callback.
- VectorIndexReadEnum is generic over `S: UniversalRead`; sparse mmap variants
hold `InvertedIndexCompressedMmap<_, S>` rather than a concrete `MmapFile`.
- Split the 1182-line segment_constructor_base.rs into a module (paths,
vector_storage, payload_storage, id_tracker, vector_index,
sparse_vector_index, create_segment, segment, legacy_state); the sparse
dispatcher's match arms collapse into three per-family helpers.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
* feat: LiveReload (no-op) dispatch for read-only vector index enum (#9436)
---------
Co-authored-by: generall <andrey@vasnetsov.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* feat: live_reload for read-only immutable id tracker + enum dispatch
* deleted should be already sorted
---------
Co-authored-by: Luis Cossío <luis.cossio@outlook.com>
* feat: add open to read-only HNSW index
* fix: dedicated universal-IO load for read-only graph
* fix: report actual graph residency in read-only is_on_disk
* refactor: rename load_via to load_universal
* refactor: unify graph links loading on universal IO
Replace the dual GraphLinks load paths (regular mmap + universal IO) with a
single universal-IO loader, and the `Mmap` GraphLinksEnum variant with a
`Universal` variant that keeps a type-erased `UniversalRead` handle alive
behind `Box<dyn GraphLinksStorage>` (UniversalRead is not object-safe).
- GraphLinksEnum::from_storage picks the variant from UniversalKind:
borrowable (mmap-backed) kinds stay `Universal`, others are materialized
into `Ram`, so the borrowability invariant in GraphLinksStorage::bytes
holds by construction.
- Loading takes a `Populate` parameter, derived from `hnsw_config.on_disk`:
on_disk -> Populate::No (lazy), otherwise Populate::Blocking.
- is_on_disk is reported from config again instead of the enum variant.
- Split graph_links.rs into a graph_links/ module (format, vectors, storage,
links, tests) with a relationship diagram in mod.rs.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* refactor: drop LoadOption in favor of a Populate parameter
After unifying the link load paths on universal IO, LoadOption only encoded a
populate choice with a fs that was always MmapFs. Replace it with a `Populate`
argument to GraphLayers::load, removing the enum, its constructors, the
load_links helper, and the unused generic backend.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
---------
Co-authored-by: generall <andrey@vasnetsov.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Reduce the default in-RAM update worker queue size to limit memory
held by pending operations and provide faster backpressure.
Co-authored-by: Cursor <cursoragent@cursor.com>
