storage v2 chunked seal segment loading is based on caching layer. A
cell unit in storage v2 is a parquet row group in remote object storage,
containing all fields. Therefore, each field needs a proxy to do related
one field operations.
<img width="965" alt="Screenshot 2025-04-28 at 10 59 30"
src="https://github.com/user-attachments/assets/83e93a10-3b1d-4066-ac17-b996d5650416"
/>
related: #39173
---------
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
Related to #39718Fixesmilvus-io/pymilvus#2771
This PR:
- Make AsyncRetrieve task triggers "schema check" logic as well
- Rename `AddField` related methods to align with code standard
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Related to #39718
This PR:
- Add reopen logic for growing & sealed segments
- Lazy reopen when schema version increases
- Add FinishLoad api for loading progress
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
json stats add map null check before insert into tantivity. Json stats
index may fail if there is no data
issue:https://github.com/milvus-io/milvus/issues/41494
---------
Signed-off-by: Xianhui.Lin <xianhui.lin@zilliz.com>
Related to #39718
This PR:
- Use WAL broadcast timestamp as Collection update timestamp
- Remove request_fields size assertion
- Remove proxy schema cache loaded field check & skip related cases
- other minor issues
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Optimized JSON filter execution by introducing
ProcessJsonStatsChunkPos() for unified position calculation and
GetNextBatchSize() for better batch processing.
Improved JSON key generation by replacing manual path joining with
milvus::Json::pointer() and adjusted slot size calculation for JSON key
index jobs.
Updated the task slot calculation logic in calculateStatsTaskSlot() to
handle the increased resource needs of JSON key index jobs.
issue: https://github.com/milvus-io/milvus/issues/41378https://github.com/milvus-io/milvus/issues/41218
---------
Signed-off-by: Xianhui.Lin <xianhui.lin@zilliz.com>
issue: #40942
Add simde package, which can make porting SIMD code to other
architectures much easier.
Signed-off-by: Shawn Wang <shawn.wang@zilliz.com>
fix: #39755
The following shows a simple benchmark where insert 1M docs where all
rows are "hello", the latency is segcore level, CPU is 9900K:
master: 2.62ms
this PR: 2.11ms
bench mark code:
```
TEST(TextMatch, TestPerf) {
auto schema = GenTestSchema({}, true);
auto seg = CreateSealedSegment(schema, empty_index_meta);
int64_t N = 1000000;
uint64_t seed = 19190504;
auto raw_data = DataGen(schema, N, seed);
auto str_col = raw_data.raw_->mutable_fields_data()
->at(1)
.mutable_scalars()
->mutable_string_data()
->mutable_data();
for (int64_t i = 0; i < N - 1; i++) {
str_col->at(i) = "hello";
}
SealedLoadFieldData(raw_data, *seg);
seg->CreateTextIndex(FieldId(101));
auto now = std::chrono::high_resolution_clock::now();
auto expr = GetMatchExpr(schema, "hello", OpType::TextMatch);
auto final = ExecuteQueryExpr(expr, seg.get(), N, MAX_TIMESTAMP);
auto end = std::chrono::high_resolution_clock::now();
auto duration =
std::chrono::duration_cast<std::chrono::microseconds>(end - now);
std::cout << "TextMatch query time: " << duration.count() << "ms"
<< std::endl;
}
```
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
support parallel loading sealed and growing segments with storage v2
format by async reading row groups.
related: #39173
---------
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
Issue: #41210
After https://github.com/zilliztech/tantivy/pull/5, we can provide
milvus row id directly to tantivy rather than record it in the fast
field "doc_id".
So rather than search tantivy doc id and then get milvus row id from
"doc_id", now, the searched tantivy doc id is the milvus row id,
eliminating the expensive acquiring row id phase.
The following shows a simple benchmark where insert **1M** docs where
all rows are "hello", the latency is **segcore** level, CPU is 9900K:
**The latency is 2.02 and 2.1 times respectively.**
bench mark code:
```
TEST(TextMatch, TestPerf) {
auto schema = GenTestSchema({}, true);
auto seg = CreateSealedSegment(schema, empty_index_meta);
int64_t N = 1000000;
uint64_t seed = 19190504;
auto raw_data = DataGen(schema, N, seed);
auto str_col = raw_data.raw_->mutable_fields_data()
->at(1)
.mutable_scalars()
->mutable_string_data()
->mutable_data();
for (int64_t i = 0; i < N - 1; i++) {
str_col->at(i) = "hello";
}
SealedLoadFieldData(raw_data, *seg);
seg->CreateTextIndex(FieldId(101));
auto now = std::chrono::high_resolution_clock::now();
auto expr = GetMatchExpr(schema, "hello", OpType::TextMatch);
auto final = ExecuteQueryExpr(expr, seg.get(), N, MAX_TIMESTAMP);
auto end = std::chrono::high_resolution_clock::now();
auto duration =
std::chrono::duration_cast<std::chrono::microseconds>(end - now);
std::cout << "TextMatch query time: " << duration.count() << "ms"
<< std::endl;
}
```
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
issue: https://github.com/milvus-io/milvus/issues/40897
After this, the document add operations scheduling duration is decreased
roughly from 6s to 0.9s for the case in the issue.
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
issue: #41172
Elements with type int8 or int16 in Array is encoded using int32, so we
should parse it as int32 when creating index.
Signed-off-by: sunby <sunbingyi1992@gmail.com>
Ref: https://github.com/milvus-io/milvus/issues/40823
It does not make any sense to create single segment tantivy index for
old version such as 2.4 by using tantivy V7.
So, clean the relevant code.
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
issue: https://github.com/milvus-io/milvus/issues/40006
This PR make tantivy document add by batch. Add document by batch can
greately reduce the latency of scheduling the document add operation
(call tantivy `add_document` only schdules the add operation and it
returns immediately after scheduled) , because each call involes a tokio
block_on which is relatively heavy.
Reduce scheduling part not necessarily reduces the overall latency if
the index writer threads does not process indexing quickly enough.
But if scheduling itself is pretty slow, even the index writer threads
process indexing very fast (by increasing thread number), the overall
performance can still be limited.
The following codes bench the PR (Note, the duration only counts for
scheduling without commit)
```
fn test_performance() {
let field_name = "text";
let dir = TempDir::new().unwrap();
let mut index_wrapper = IndexWriterWrapper::create_text_writer(
field_name,
dir.path().to_str().unwrap(),
"default",
"",
1,
50_000_000,
false,
TantivyIndexVersion::V7,
)
.unwrap();
let mut batch = vec![];
for i in 0..1_000_000 {
batch.push(format!("hello{:04}", i));
}
let batch_ref = batch.iter().map(|s| s.as_str()).collect::<Vec<_>>();
let now = std::time::Instant::now();
index_wrapper
.add_data_by_batch(&batch_ref, Some(0))
.unwrap();
let elapsed = now.elapsed();
println!("add_data_by_batch elapsed: {:?}", elapsed);
}
```
Latency roughly reduces from 1.4s to 558ms.
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
fix: https://github.com/milvus-io/milvus/issues/40823
To solve the problem in the issue, we have to support building tantivy
index with low version
for those query nodes with low tantivy version.
This PR does two things:
1. refactor codes for IndexWriterWrapper to make it concise
2. enable IndexWriterWrapper to build tantivy index by different tantivy
crate
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
after the pr merged, we can support to insert, upsert, build index,
query, search in the added field.
can only do the above operates in added field after add field request
complete, which is a sync operate.
compact will be supported in the next pr.
#39718
---------
Signed-off-by: lixinguo <xinguo.li@zilliz.com>
Co-authored-by: lixinguo <xinguo.li@zilliz.com>
- Feat: Support Mix compaction. Covering tests include compatibility and
rollback ability.
- Read v1 segments and compact with v2 format.
- Read both v1 and v2 segments and compact with v2 format.
- Read v2 segments and compact with v2 format.
- Compact with duplicate primary key test.
- Compact with bm25 segments.
- Compact with merge sort segments.
- Compact with no expiration segments.
- Compact with lack binlog segments.
- Compact with nullable field segments.
- Feat: Support Clustering compaction. Covering tests include
compatibility and rollback ability.
- Read v1 segments and compact with v2 format.
- Read both v1 and v2 segments and compact with v2 format.
- Read v2 segments and compact with v2 format.
- Compact bm25 segments with v2 format.
- Compact with memory limit.
- Enhance: Use serdeMap serialize in BuildRecord function to support all
Milvus data types.
related: #39173
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
two point:
(1) reoder conjucts expr's subexpr, postpone heavy operations
sequence: int(column) -> index(column) -> string(column) -> light
conjuct
...... -> json(column) -> heavy conjuct -> two_column_compare
(2) support pre filter for expr execute, skip scan raw data that had
been skipped
because of preceding expr result.
#39869
Signed-off-by: luzhang <luzhang@zilliz.com>
Co-authored-by: luzhang <luzhang@zilliz.com>
issue: #40730
also see: https://github.com/milvus-io/cgosymbolizer/pull/2
After these PR, at linux:
- the milvus will always enable jemalloc by default.
- jemalloc will always compiled with --enable-prof options.
- all image will always enable the jemalloc prof by default.
- a pprof http service for jemalloc at `/debug/jemalloc/` will be
registered into restful.
- `jeprof` can remote profile the memory of milvus.
Signed-off-by: chyezh <chyezh@outlook.com>
issue: #40607
tantivy change: https://github.com/zilliztech/tantivy/pull/3
Benchmarks:
Test Envrioment: CPU 9900K
The data is insert by:
```
for i in 0..N {
for j in 0..UNIQUE {
let key = format!("hello{}", j);
index_writer.add_string(&key, i * UNIQUE + j).unwrap();
}
}
```
So the unique influences the locality of the matched docs.
The latency is the avg latency over 1000 repeate quries.
The result shows 22.5%-34.8% latency reduction.
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
issue: #40308
This issue fixes these two concurrent issues:
1. element in null_offset is used to set bitset where the size of bitset
is initialized by tantivy document count. However, there may still be
some documents that are not committed in tantivy but are null in
null_offset. So array out of range occurs.
2. null_offset can be read and write concurrently but there's no
synchronization protection.
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
issue: #40198
Fix random sample does not consider empty input, that is no data is hit
by filter expression.
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
issue: https://github.com/milvus-io/milvus/issues/35528
If the query data type does not match the index type, fall back to a
brute-force search
---------
Signed-off-by: sunby <sunbingyi1992@gmail.com>
* use the new packed reader and writer api to be compatible with current
etcd meta
* For the new packed writer API: column groups and paths are explicitly
defined by users and won't split column groups by memory in storage v2.
Packed writer follows the user-defined column groups to split arrow
record and write into the corresponding file path.
* For the new packed reader API: read paths are explicitly defined by
users.
related: #39173
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
issue: https://github.com/milvus-io/milvus/issues/39818
This PR mimics Varchar data type, allows insert, search, query, delete,
full-text search and others.
Functionalities related to filter expressions are disabled temporarily.
Storage changes for Text data type will be in the following PRs.
Signed-off-by: Patrick Weizhi Xu <weizhi.xu@zilliz.com>
fix: #39711
Unlike English sentence where each words are parsed exactly once and one
after one with position length 1, one Chinese word may be parsed to
multiple words with position length larger than 1.
For example, "badminton and skiing" will be parsed to Token{ start: 0,
length: 1, text: "badminton" }, Token{ start: 1, length: 1, text: "and"
}, and Token{ start: 2, length: 1, text: "tennis" }.
While for exmaple for Chinsese: "羽毛球和滑雪" may be parsed to Token{ start:
0, length: 2, text: "羽毛" }, Token{ start: 0, length: 3, text: "羽毛球" },
Token{ start: 3, length: 1, text: "和" }, and Token{ start: 4, length: 2,
text: "滑雪" }.
This PR fix that the code not recognizes this situation.
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
issue: #39541
This PR implements random sample, the syntax is:
```
filter="random_sample(factor)"
or
filter="boolean_expression && random_sample(factor)"
where
factor is a float between (0, 1) and
boolean_expression is like
"1 <= number < 10", "color in ["read, "blue"]" or others
```
---------
Signed-off-by: SpadeA-Tang <tangchenjie1210@gmail.com>
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
1. skip get expr arguments which deserialize proto for every batch
execute.
2. replace unordered_set with sort array that has better performace for
small set.
#39688
Co-authored-by: luzhang <luzhang@zilliz.com>
https://github.com/milvus-io/milvus/issues/35528
This PR adds json index support for json and dynamic fields. Now you can
only do unary query like 'a["b"] > 1' using this index. We will support
more filter type later.
basic usage:
```
collection.create_index("json_field", {"index_type": "INVERTED",
"params": {"json_cast_type": DataType.STRING, "json_path":
'json_field["a"]["b"]'}})
```
There are some limits to use this index:
1. If a record does not have the json path you specify, it will be
ignored and there will not be an error.
2. If a value of the json path fails to be cast to the type you specify,
it will be ignored and there will not be an error.
3. A specific json path can have only one json index.
4. If you try to create more than one json indexes for one json field,
sdk(pymilvus<=2.4.7) may return immediately because of internal
implementation. This will be fixed in a later version.
---------
Signed-off-by: sunby <sunbingyi1992@gmail.com>
Previously the grep with regex does not work and failed to match lots of
.cpp files
This PR:
- use "-E" flag to use regex match
- commit the fixed result of current cpp code
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Related to #39339
Extra indexes can be ignored for most cases since sorted pk column
already provided indexing features
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Related to #39003
Previous PR #39004 has to clone & flip bitset due to bitset does not
support find0 operator. #39176 added this feature so clone & flip could
be removed now.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #38715
- Current milvus use a serialized index size(compressed) for estimate
resource for loading.
- Add a new field `MemSize` (before compressing) for index to estimate
resource.
---------
Signed-off-by: chyezh <chyezh@outlook.com>
issue: #39124
`bitset::find_first()` and `bitset::find_next()` now accept one more
parameter, which allows to search for `0` bit instead of `1` bit
Signed-off-by: Alexandr Guzhva <alexanderguzhva@gmail.com>
instead of marking as not supported,
`ChunkedSparseFloatColumn::DataByteSize` can simply use the impl of
super class.
issue: https://github.com/milvus-io/milvus/issues/39158
Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
Related to #39003
Copying bitset value bit by bit is slow and CPU heavy, this PR utilizes
bitset operator "|=" to accelerate this procedure
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Related to previous PR #38157
If mmapped row is too small, frequent fwrite call still cost too much
cpu time for context switching. This PR add buffered write to avoid this
bad case with extra buffer per variable field.
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
`File.Write` and `File.WriteInt` use `write`, which may be just direct
syscall in some systems. When mappding field data and write line by
line, this could cost lost of CPU time when the row number is large.
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #33285
- move most cgo opeartions related to search/query into segcore package
for reusing for streamingnode.
- add go unittest for segcore operations.
Signed-off-by: chyezh <chyezh@outlook.com>
Related to #37999
This PR add `SetThreadName` API for marking cgo thread and utilize it
when initializing cgo worker.
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #37851
- make a global thread pool at tantivy temporally.
- set 1 but not 4 threads for inverted text index.
Signed-off-by: chyezh <chyezh@outlook.com>
issue: #37665#37631#37620#37587#36906
knowhere has add default nlist value, so some invalid param test ut with
no nlist param will be valid.
Signed-off-by: xianliang.li <xianliang.li@zilliz.com>
Bingyi Sun
Buqian Zheng
sthuang
zhagnlu
foxspy
congqixia
Spade A
cai.zhang
Chun Han
Xianhui Lin
aoiasd
cqy123456
sparknack
Ted Xu
Zhen Ye
smellthemoon
Xiaofan
Cai Yudong
sre-ci-robot
Micka
yihao.dai
Patrick Weizhi Xu
sparknack
Gao
Alexander Guzhva
tinswzy