See also #33561
This PR:
- Use zero copy when buffering insert messages
- Make `storage.InsertCodec` support serialize multiple insert data
chunk into same batch binlog files
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #32995
To speed up the construction and querying of Bloom filters, we chose a
blocked Bloom filter instead of a basic Bloom filter implementation.
WARN: This PR is compatible with old version bf impl, but if fall back
to old milvus version, it may causes bloom filter deserialize failed.
In single Bloom filter test cases with a capacity of 1,000,000 and a
false positive rate (FPR) of 0.001, the blocked Bloom filter is 5 times
faster than the basic Bloom filter in both querying and construction, at
the cost of a 30% increase in memory usage.
- Block BF construct time {"time": "54.128131ms"}
- Block BF size {"size": 3021578}
- Block BF Test cost {"time": "55.407352ms"}
- Basic BF construct time {"time": "210.262183ms"}
- Basic BF size {"size": 2396308}
- Basic BF Test cost {"time": "192.596229ms"}
In multi Bloom filter test cases with a capacity of 100,000, an FPR of
0.001, and 100 Bloom filters, we reuse the primary key locations for all
Bloom filters to avoid repeated hash computations. As a result, the
blocked Bloom filter is also 5 times faster than the basic Bloom filter
in querying.
- Block BF TestLocation cost {"time": "529.97183ms"}
- Basic BF TestLocation cost {"time": "3.197430181s"}
---------
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
See also #32642
`LocationCache` used map to store different locations for different K
which may cause lots of CPU time when get locations many times.
This PR change the implementation of LocationCache to store only the
location for the largest K used to totally remove the map access
operation.
See pprof from test of @XuanYang-cn
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #33005
1. add `MemorySize` field for insert binlog.
2. `LogSize` means the file size in the storage object.
3. `MemorySize` means the size of the data in the memory.
---------
Signed-off-by: Cai Zhang <cai.zhang@zilliz.com>
Signed-off-by: cai.zhang <cai.zhang@zilliz.com>
See also #32642
This PR reuses hash locations for bloom filter prediction utilizing
`storage.Location`, like enhancement #32642.
Also adds a utility struct in storage: `LocationCache` to storage
locations for variable K (numbers of hash functions)
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #32530
when try to match segment bloom filter with pk, we can reuse the hash
locations. This PR maintain the max hash Func, and compute hash location
once for all segment, reuse hash location can speed up bf access
---------
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
issue: #19095,#29655,#31718
- Change `ListWithPrefix` to `WalkWithPrefix` of OOS into a pipeline
mode.
- File garbage collection is performed in other goroutine.
- Segment Index Recycle clean index file too.
---------
Signed-off-by: chyezh <chyezh@outlook.com>
issue: #29419
added helper functions to parse JSON representation of sparse float
vectors, will be used by both the restful server and the import utils.
Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
add sparse float vector support to different milvus components,
including proxy, data node to receive and write sparse float vectors to
binlog, query node to handle search requests, index node to build index
for sparse float column, etc.
https://github.com/milvus-io/milvus/issues/29419
---------
Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
Define FieldValue, FieldStats, PartitionStats
FieldValue is largely copied from PrimaryKey
FieldStats is largely copied from PrimaryKeyStats
PartitionStats is map[segmentid][]FieldStats
Each partition can have a PartitionStats file
/kind feature
related: #30287
related: #30633
---------
Signed-off-by: wayblink <anyang.wang@zilliz.com>
This PR introduces novel managerial roles for importv2:
1. ImportMeta: To manage all the import tasks;
2. ImportScheduler: To process tasks and modify their states;
3. ImportChecker: To ascertain the completion of all tasks and instigate
relevant operations.
issue: https://github.com/milvus-io/milvus/issues/28521
---------
Signed-off-by: bigsheeper <yihao.dai@zilliz.com>
See also #30404
`PrimaryKey` is used to hold pk values for both int64 & varchar data
type. Since it is an interface it may occupies more memory than pure
slices when holding a group of pks.
This PR add `PrimaryKeys` interface when some other module need to hold
lots of PrimaryKeys.
By using this interface, it could reduce the memory of pk slice to half
when using Int64 Pk data type and reduce interface cost for each row of
varchar as well.
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
This PR introduces novel importv2 roles for datanode:
1. Executor: To execute tasks, a import task will be divided into the
following steps: read data -> hash data -> sync data;
2. Manager: To manage all the tasks;
issue: https://github.com/milvus-io/milvus/issues/28521
---------
Signed-off-by: bigsheeper <yihao.dai@zilliz.com>
fix: #29757
In previous code, `ColumnBasedInsertMsgToInsertData` adds empty field if
the insertMsg parameter does not have the column schema defined. This
may lead to unexpected behavior of caller functions.
This PR:
- Add column missing check
- Add column length check
- Generate BlobInfo for ColumnBasedInsertMsgToInsertData result
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
This PR defines the new import reader interfaces and implement a binlog
reader for import.
issue: https://github.com/milvus-io/milvus/issues/28521
---------
Signed-off-by: bigsheeper <yihao.dai@zilliz.com>
issue: https://github.com/milvus-io/milvus/issues/27704
Add inverted index for some data types in Milvus. This index type can
save a lot of memory compared to loading all data into RAM and speed up
the term query and range query.
Supported: `INT8`, `INT16`, `INT32`, `INT64`, `FLOAT`, `DOUBLE`, `BOOL`
and `VARCHAR`.
Not supported: `ARRAY` and `JSON`.
Note:
- The inverted index for `VARCHAR` is not designed to serve full-text
search now. We will treat every row as a whole keyword instead of
tokenizing it into multiple terms.
- The inverted index don't support retrieval well, so if you create
inverted index for field, those operations which depend on the raw data
will fallback to use chunk storage, which will bring some performance
loss. For example, comparisons between two columns and retrieval of
output fields.
The inverted index is very easy to be used.
Taking below collection as an example:
```python
fields = [
FieldSchema(name="pk", dtype=DataType.VARCHAR, is_primary=True, auto_id=False, max_length=100),
FieldSchema(name="int8", dtype=DataType.INT8),
FieldSchema(name="int16", dtype=DataType.INT16),
FieldSchema(name="int32", dtype=DataType.INT32),
FieldSchema(name="int64", dtype=DataType.INT64),
FieldSchema(name="float", dtype=DataType.FLOAT),
FieldSchema(name="double", dtype=DataType.DOUBLE),
FieldSchema(name="bool", dtype=DataType.BOOL),
FieldSchema(name="varchar", dtype=DataType.VARCHAR, max_length=1000),
FieldSchema(name="random", dtype=DataType.DOUBLE),
FieldSchema(name="embeddings", dtype=DataType.FLOAT_VECTOR, dim=dim),
]
schema = CollectionSchema(fields)
collection = Collection("demo", schema)
```
Then we can simply create inverted index for field via:
```python
index_type = "INVERTED"
collection.create_index("int8", {"index_type": index_type})
collection.create_index("int16", {"index_type": index_type})
collection.create_index("int32", {"index_type": index_type})
collection.create_index("int64", {"index_type": index_type})
collection.create_index("float", {"index_type": index_type})
collection.create_index("double", {"index_type": index_type})
collection.create_index("bool", {"index_type": index_type})
collection.create_index("varchar", {"index_type": index_type})
```
Then, term query and range query on the field can be speed up
automatically by the inverted index:
```python
result = collection.query(expr='int64 in [1, 2, 3]', output_fields=["pk"])
result = collection.query(expr='int64 < 5', output_fields=["pk"])
result = collection.query(expr='int64 > 2997', output_fields=["pk"])
result = collection.query(expr='1 < int64 < 5', output_fields=["pk"])
```
---------
Signed-off-by: longjiquan <jiquan.long@zilliz.com>
We have implemented the chunkcache (in cpp) to retrieve vectors, hence
rendering the vectorchunkcache (in golang) obsolete.
issue: https://github.com/milvus-io/milvus/issues/28568
---------
Signed-off-by: bigsheeper <yihao.dai@zilliz.com>
see also: https://github.com/milvus-io/milvus/issues/28509
Currently Minio latency monitoring for get operation only collects the
duration of getting object (which just returns an io.Reader and does not
really read from minio), this pr will correct this behavior.
Signed-off-by: bigsheeper <yihao.dai@zilliz.com>
See also #27675
When L0 segment contains only delta data, merged statslog shall be
skiped when performing sync task
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
See also #28575
Add zero-length check for `storage.NewPrimaryKeyStats`. This function
shall return error when non-positive rowNum passed.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Benchmark Milvus with https://github.com/qdrant/vector-db-benchmark and
specify the datasets as 'deep-image-96-angular'. Meanwhile, do perf
profiling during 'upload + index' stage of vector-db-benchmark and see
the following hot spots.
39.59%--github.com/milvus-io/milvus/internal/storage.MergeInsertData
|
|--21.43%--github.com/milvus-io/milvus/internal/storage.MergeFieldData
| |
| |--17.22%--runtime.memmove
| |
| |--1.53%--asm_exc_page_fault
| ......
|
|--18.16%--runtime.memmove
|
|--1.66%--asm_exc_page_fault
......
The hot code path is in storage.MergeInsertData() which updates
buffer.buffer by creating a new 'InsertData' instance and merging both
the old buffer.buffer and addedBuffer into it. When it calls golang
runtime.memmove to move buffer.buffer which is with big size (>1M), the
hot spots appear.
To avoid the above overhead, update storage.MergeInsertData() by
appending addedBuffer to buffer.buffer, instead of moving buffer.buffer
and addedBuffer to a new 'InsertData'. This change removes the hot spots
'runtime.memmove' from perf profiling output. Additionally, the 'upload
+ index' time, which is one performance metric of vector-db-benchmark,
is reduced around 60% with this change.
Signed-off-by: Cathy Zhang <cathy.zhang@intel.com>
smellthemoon
Ted Xu
shaoting-huang
congqixia
XuanYang-cn
wei liu
wayblink
Cai Yudong
cai.zhang
aoiasd
Aldrin
chyezh
Buqian Zheng
cqy123456
SimFG
groot
yihao.dai
Chun Han
Xu Tong
yah01
Jiquan Long
Bingyi Sun
PowderLi
Enwei Jiao
zhenshan.cao
zhagnlu
bjzhjing
xige-16
Xiaofan
yah01