related: #45993
This commit extends nullable vector support to the proxy layer,
querynode,
and adds comprehensive validation, search reduce, and field data
handling
for nullable vectors with sparse storage.
Proxy layer changes:
- Update validate_util.go checkAligned() with getExpectedVectorRows()
helper
to validate nullable vector field alignment using valid data count
- Update checkFloatVectorFieldData/checkSparseFloatVectorFieldData for
nullable vector validation with proper row count expectations
- Add FieldDataIdxComputer in typeutil/schema.go for logical-to-physical
index translation during search reduce operations
- Update search_reduce_util.go reduceSearchResultData to use
idxComputers
for correct field data indexing with nullable vectors
- Update task.go, task_query.go, task_upsert.go for nullable vector
handling
- Update msg_pack.go with nullable vector field data processing
QueryNode layer changes:
- Update segments/result.go for nullable vector result handling
- Update segments/search_reduce.go with nullable vector offset
translation
Storage and index changes:
- Update data_codec.go and utils.go for nullable vector serialization
- Update indexcgowrapper/dataset.go and index.go for nullable vector
indexing
Utility changes:
- Add FieldDataIdxComputer struct with Compute() method for efficient
logical-to-physical index mapping across multiple field data
- Update EstimateEntitySize() and AppendFieldData() with fieldIdxs
parameter
- Update funcutil.go with nullable vector support functions
<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit
* **New Features**
* Full support for nullable vector fields (float, binary, float16,
bfloat16, int8, sparse) across ingest, storage, indexing, search and
retrieval; logical↔physical offset mapping preserves row semantics.
* Client: compaction control and compaction-state APIs.
* **Bug Fixes**
* Improved validation for adding vector fields (nullable + dimension
checks) and corrected search/query behavior for nullable vectors.
* **Chores**
* Persisted validity maps with indexes and on-disk formats.
* **Tests**
* Extensive new and updated end-to-end nullable-vector tests.
<sub>✏️ Tip: You can customize this high-level summary in your review
settings.</sub>
<!-- end of auto-generated comment: release notes by coderabbit.ai -->
---------
Signed-off-by: marcelo-cjl <marcelo.chen@zilliz.com>
issue: #46349
When using brute-force search, the iterator results from multiple chunks
are merged; at that point, we need to pay attention to how the metric
affects result ranking.
Signed-off-by: xianliang.li <xianliang.li@zilliz.com>
issue: https://github.com/milvus-io/milvus/issues/42148
For a vector field inside a STRUCT, since a STRUCT can only appear as
the element type of an ARRAY field, the vector field in STRUCT is
effectively an array of vectors, i.e. an embedding list.
Milvus already supports searching embedding lists with metrics whose
names start with the prefix MAX_SIM_.
This PR allows Milvus to search embeddings inside an embedding list
using the same metrics as normal embedding fields. Each embedding in the
list is treated as an independent vector and participates in ANN search.
Further, since STRUCT may contain scalar fields that are highly related
to the embedding field, this PR introduces an element-level filter
expression to refine search results.
The grammar of the element-level filter is:
element_filter(structFieldName, $[subFieldName] == 3)
where $[subFieldName] refers to the value of subFieldName in each
element of the STRUCT array structFieldName.
It can be combined with existing filter expressions, for example:
"varcharField == 'aaa' && element_filter(struct_field, $[struct_int] ==
3)"
A full example:
```
struct_schema = milvus_client.create_struct_field_schema()
struct_schema.add_field("struct_str", DataType.VARCHAR, max_length=65535)
struct_schema.add_field("struct_int", DataType.INT32)
struct_schema.add_field("struct_float_vec", DataType.FLOAT_VECTOR, dim=EMBEDDING_DIM)
schema.add_field(
"struct_field",
datatype=DataType.ARRAY,
element_type=DataType.STRUCT,
struct_schema=struct_schema,
max_capacity=1000,
)
...
filter = "varcharField == 'aaa' && element_filter(struct_field, $[struct_int] == 3 && $[struct_str] == 'abc')"
res = milvus_client.search(
COLLECTION_NAME,
data=query_embeddings,
limit=10,
anns_field="struct_field[struct_float_vec]",
filter=filter,
output_fields=["struct_field[struct_int]", "varcharField"],
)
```
TODO:
1. When an `element_filter` expression is used, a regular filter
expression must also be present. Remove this restriction.
2. Implement `element_filter` expressions in the `query`.
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
See: #44956
This PR upgrades loon to the latest version and resolves building
conflicts.
---------
Signed-off-by: Ted Xu <ted.xu@zilliz.com>
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Co-authored-by: Congqi Xia <congqi.xia@zilliz.com>
Previously, mmap settings configured at the collection or field level
were not being applied during segment loading in segcore. This was
caused by:
1. A typo in the key name: "mmap.enable" instead of "mmap.enabled"
2. Missing logic to parse and apply mmap settings from schema
This commit fixes the issue by:
- Correcting the key name to "mmap.enabled" to match the standard
- Adding Schema::MmapEnabled() method to retrieve field/collection level
mmap settings with proper fallback logic
- Parsing mmap settings from field type_params and collection properties
during schema parsing
- Applying computed mmap settings in LoadColumnGroup() and Load()
methods instead of hardcoded false values
- Using global MmapConfig as fallback when no explicit setting exists
The mmap setting priority is now:
1. Field-level mmap setting (from type_params)
2. Collection-level mmap setting (from properties)
3. Global mmap config (from MmapManager)
For column groups, if any field has mmap enabled, the entire group uses
mmap (since they are loaded together).
Related issue: #45060
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
This PR adds support for reading data from StorageV2 using manifest
files and the Loon FFI interface during index building, providing an
alternative to the traditional segment insert files approach.
Key changes:
Core C++ changes:
- Add SEGMENT_MANIFEST_KEY and LOON_FFI_PROPERTIES_KEY constants for
manifest handling
- Extend FileManagerContext to carry loon_ffi_properties for FFI
operations
- Update index_c.cpp to pass manifest and loon properties to file
managers for all index types (vector, JSON key, text)
- Implement GetFieldDatasFromManifest() in Util.cpp using Arrow C Stream
interface:
* Create Arrow schema from field metadata
* Initialize FFI reader with manifest content and storage properties
* Import record batches from C data interface
* Convert to FieldData for index building
- Update DiskFileManagerImpl and MemFileManagerImpl to support
manifest-based data reading with fallback to traditional paths
Loon FFI utilities (internal/core/src/storage/loon_ffi/):
- Add ToCStorageConfig() to convert StorageConfig to C-compatible
structure
- Implement GetManifest() to parse manifest JSON and retrieve column
groups via FFI
- Enhance MakePropertiesFromStorageConfig() integration
Storage V2 integration:
- Update milvus-storage dependency from 0883026 to 302143c for latest
FFI support
Protobuf changes:
- Add manifest field to BuildIndexInfo for passing manifest path to C++
layer
Configuration:
- Add common.storageV2.useLoonFFI config option (default: false) for
feature toggle
This change is part of issue #44956 to integrate the StorageV2 FFI
interface as the unified storage layer. The implementation maintains
backward compatibility by checking for manifest presence and falling
back to existing segment insert files approach when manifest is not
provided.
Related issue: #44956
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #45486
This commit refactors the chunk writing system by introducing a
two-phase
approach: size calculation followed by writing to a target. This enables
efficient group chunk creation where multiple fields share a single mmap
region, significantly reducing the number of mmap system calls and VMAs.
- Optimize `mmap` usage: single `mmap` per group chunk instead of per
field
- Split ChunkWriter into two phases:
- `calculate_size()`: Pre-compute required memory without allocation
- `write_to_target()`: Write data to a provided ChunkTarget
- Implement `ChunkMmapGuard` for unified mmap region lifecycle
management
- Handles `munmap` and file cleanup via RAII
- Shared via `std::shared_ptr` across multiple chunks in a group
Signed-off-by: Shawn Wang <shawn.wang@zilliz.com>
---------
Signed-off-by: Shawn Wang <shawn.wang@zilliz.com>
1. Array.h: Add output_data(ScalarFieldProto&) overload for both Array
and ArrayView classes
2. Use std::string_view instead of std::string for VARCHAR and GEOMETRY
types to avoid extra string copies
3. Call Reserve(length_) before writing to proto objects to reduce
memory reallocations
a simple test shows those optimizations improve the Array of Varchar
bulk_subscript performance by 20%
issue: https://github.com/milvus-io/milvus/issues/45679
Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
Related to #45060
Refactor segment loading architecture to make segments autonomously
manage their own loading process, moving the orchestration logic from Go
(segment_loader.go) to C++ (segcore).
**C++ Layer (segcore):**
- Added `SetLoadInfo()` and `Load()` methods to `SegmentInterface` and
implementations
- Implemented `ChunkedSegmentSealedImpl::Load()` with parallel loading
strategy:
- Separates indexed fields from non-indexed fields
- Loads indexes concurrently using thread pools
- Loads field data for non-indexed fields in parallel
- Implemented `SegmentGrowingImpl::Load()` to convert and load field
data
- Extracted `LoadIndexData()` as a reusable utility function in
`Utils.cpp`
- Added `SegmentLoad()` C binding in `segment_c.cpp`
**Go Layer:**
- Added `Load()` method to segment interfaces
- Updated mock implementations and test interfaces
- Integrated new C++ `SegmentLoad()` binding in Go segment wrapper
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Related to #45445
Previously, FillFieldData for JSON fields would assert and fail when a
default_value was provided, blocking index creation for JSON fields with
default values (including dynamic fields like $meta).
This change enables JSON default value support by:
- Removing the assertion that blocked default values
- Parsing bytes_data into Json objects when default_value is present
- Properly filling data_ array and setting valid_data_ bitset to true
- Maintaining null behavior when no default_value is provided
Impact:
- Fixes index creation failure for JSON fields with default values
- Resolves upgrade issues from 2.5 to 2.6.5 where dynamic fields with
default values couldn't be indexed
- Index builds that were stuck in InProgress state can now complete
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Related to #44819
This fix addresses an issue(#44819) where the offset parameter did not
work correctly during searches when multiple results had identical
scores. The problem occurred because results with equal scores were not
consistently ordered, leading to unpredictable pagination behavior.
The solution adds a new sorting step (SortEqualScoresByPks) in the
reduce phase that sorts results with identical scores by their primary
keys in ascending order. This ensures deterministic ordering and enables
proper offset functionality.
Changes:
- Add SortEqualScoresByPks() to sort results with equal scores by PK
- Add SortEqualScoresOneNQ() to handle per-query sorting logic
- Invoke sorting step after FillPrimaryKey() in Reduce() workflow
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #43427
This pr's main goal is merge #37417 to milvus 2.5 without conflicts.
# Main Goals
1. Create and describe collections with geospatial type
2. Insert geospatial data into the insert binlog
3. Load segments containing geospatial data into memory
4. Enable query and search can display geospatial data
5. Support using GIS funtions like ST_EQUALS in query
6. Support R-Tree index for geometry type
# Solution
1. **Add Type**: Modify the Milvus core by adding a Geospatial type in
both the C++ and Go code layers, defining the Geospatial data structure
and the corresponding interfaces.
2. **Dependency Libraries**: Introduce necessary geospatial data
processing libraries. In the C++ source code, use Conan package
management to include the GDAL library. In the Go source code, add the
go-geom library to the go.mod file.
3. **Protocol Interface**: Revise the Milvus protocol to provide
mechanisms for Geospatial message serialization and deserialization.
4. **Data Pipeline**: Facilitate interaction between the client and
proxy using the WKT format for geospatial data. The proxy will convert
all data into WKB format for downstream processing, providing column
data interfaces, segment encapsulation, segment loading, payload
writing, and cache block management.
5. **Query Operators**: Implement simple display and support for filter
queries. Initially, focus on filtering based on spatial relationships
for a single column of geospatial literal values, providing parsing and
execution for query expressions.Now only support brutal search
7. **Client Modification**: Enable the client to handle user input for
geospatial data and facilitate end-to-end testing.Check the modification
in pymilvus.
---------
Signed-off-by: Yinwei Li <yinwei.li@zilliz.com>
Signed-off-by: Cai Zhang <cai.zhang@zilliz.com>
Co-authored-by: ZhuXi <150327960+Yinwei-Yu@users.noreply.github.com>
And support set function mode and boost mode when run search with boost.
RandomScore support get random function score between [0, weight).
FunctionMode decide how to calculate boost score for multiple boost
function scores.
BoostMode decide how to calculate final score for origin score and boost
score.
relate: https://github.com/milvus-io/milvus/issues/43867
---------
Signed-off-by: aoiasd <zhicheng.yue@zilliz.com>
issue: https://github.com/milvus-io/milvus/issues/27467
>My plan is as follows.
>- [x] M1 Create collection with timestamptz field
>- [x] M2 Insert timestamptz field data
>- [x] M3 Retrieve timestamptz field data
>- [x] M4 Implement handoff
>- [x] M5 Implement compare operator
>- [x] M6 Implement extract operator
>- [x] M8 Support database/collection level default timezone
>- [x] M7 Support STL-SORT index for datatype timestamptz
---
The third PR of issue: https://github.com/milvus-io/milvus/issues/27467,
which completes M5, M6, M7, M8 described above.
## M8 Default Timezone
We will be able to use alter_collection() and alter_database() in a
future Python SDK release to modify the default timezone at the
collection or database level.
For insert requests, the timezone will be resolved using the following
order of precedence: String Literal-> Collection Default -> Database
Default.
For retrieval requests, the timezone will be resolved in this order:
Query Parameters -> Collection Default -> Database Default.
In both cases, the final fallback timezone is UTC.
## M5: Comparison Operators
We can now use the following expression format to filter on the
timestamptz field:
- `timestamptz_field [+/- INTERVAL 'interval_string'] {comparison_op}
ISO 'iso_string' `
- The interval_string follows the ISO 8601 duration format, for example:
P1Y2M3DT1H2M3S.
- The iso_string follows the ISO 8601 timestamp format, for example:
2025-01-03T00:00:00+08:00.
- Example expressions: "tsz + INTERVAL 'P0D' != ISO
'2025-01-03T00:00:00+08:00'" or "tsz != ISO
'2025-01-03T00:00:00+08:00'".
## M6: Extract
We will be able to extract sepecific time filed by kwargs in a future
Python SDK release.
The key is `time_fields`, and value should be one or more of "year,
month, day, hour, minute, second, microsecond", seperated by comma or
space. Then the result of each record would be an array of int64.
## M7: Indexing Support
Expressions without interval arithmetic can be accelerated using an
STL-SORT index. However, expressions that include interval arithmetic
cannot be indexed. This is because the result of an interval calculation
depends on the specific timestamp value. For example, adding one month
to a date in February results in a different number of added days than
adding one month to a date in March.
---
After this PR, the input / output type of timestamptz would be iso
string. Timestampz would be stored as timestamptz data, which is int64_t
finally.
> for more information, see https://en.wikipedia.org/wiki/ISO_8601
---------
Signed-off-by: xtx <xtianx@smail.nju.edu.cn>
issue: #44212
Implement search/query storage usage statistics in go side(result
reduce), only record storage usage in vector search C++ path. Need to be
implemented in query c++ path in next prs.
---------
Signed-off-by: chasingegg <chao.gao@zilliz.com>
Signed-off-by: marcelo.chen <marcelo.chen@zilliz.com>
Co-authored-by: marcelo.chen <marcelo.chen@zilliz.com>
issue: https://github.com/milvus-io/milvus/issues/42032
- Use bytes to estimate load resource in the whole estimation procedure
- Add num_rows and dim info for vector index to better estimate
- Disable eviction for tiered index's meta
---------
Signed-off-by: chasingegg <chao.gao@zilliz.com>
issue: https://github.com/milvus-io/milvus/issues/42148
Optimized from
Go VectorArray → VectorArray Proto → Binary → C++ VectorArray Proto →
C++ VectorArray local impl → Memory
to
Go VectorArray → Arrow ListArray → Memory
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
issue: #42942
This pr includes the following changes:
1. Added checks for index checker in querycoord to generate drop index
tasks
2. Added drop index interface to querynode
3. To avoid search failure after dropping the index, the querynode
allows the use of lazy mode (warmup=disable) to load raw data even when
indexes contain raw data.
4. In segcore, loading the index no longer deletes raw data; instead, it
evicts it.
5. In expr, the index is pinned to prevent concurrent errors.
---------
Signed-off-by: sunby <sunbingyi1992@gmail.com>
