Related to #43936
This PR:
- Use `folly::SharedMutex` instead of `std::shared_mutex` preventing
starvation
- Use `folly::SharedMutex::WriteHolder/ReadHolder` instead of
std::shared_lock and std::unique_lock to get better performance
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #29735
Implement partial field update functionality for upsert operations,
supporting scalar, vector, and dynamic JSON fields without requiring all
collection fields.
Changes:
- Add queryPreExecute to retrieve existing records before upsert
- Implement UpdateFieldData function for merging data
- Add IDsChecker utility for efficient primary key lookups
- Fix JSON data creation in tests using proper map marshaling
- Add test cases for partial updates of different field types
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
issue: #42416
- Rename the InsertMetric into ModifiedMetric.
- Add L0 control configuration.
- Add some L0 current state collect.
Signed-off-by: chyezh <chyezh@outlook.com>
Enables the compilation of SVE code for the bitset library if a C++
compiler supports it.
There are two conditions for enabling the SVE code
* a C++ compiler needs to have a `-march=armv8-a+sve`
* `arm_sve.h` header must be available
AFAIK, `gcc 7 does not support SVE`, `gcc 8` and `gcc 9` support SVE,
but have no `arm_sve.h` file, and only `gcc 10` has both.
Signed-off-by: Alexandr Guzhva <alexanderguzhva@gmail.com>
The Out of Memory (OOM) error occurs because a handler retains the
entire ImportRecordBatch in memory. Consequently, even when child arrays
within the batch are flushed, the memory for the complete batch is not
released. We temporarily fixed by deep copying record batch in #43724.
The proposed fix is to split the RecordBatch into smaller sub-batches by
column group. These sub-batches will be transferred via CGO, then
reassembled before being written to storage using the Storage V2 API.
Thus we can achieve zero-copy and only transferring references in CGO.
related: #43310
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
issue: #43828
Implement robust rewatch mechanism to handle etcd connection failures
and node reconnection scenarios in DataCoord and QueryCoord, along with
heartbeat lag monitoring capabilities.
Changes include:
- Implement rewatchDataNodes/rewatchQueryNodes callbacks for etcd
reconnection scenarios
- Add idempotent rewatchNodes method to handle etcd session recovery
gracefully
- Add QueryCoordLastHeartbeatTimeStamp metric for monitoring node
heartbeat lag
- Clean up heartbeat metrics when nodes go down to prevent metric leaks
---------
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
Related to #43230
This PR
- Move segcore setup function to `initcore` package to remove cgo
dependency from pkg
- Register core callback only for components depends on segcore
- Rectify `UpdateLogLevel` implementation
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Each ColumnReader consumes ReaderProperties.BufferSize memory
independently. Therefore, the bufferSize should be divided by the number
of columns to ensure total memory usage stays within the intended limit.
issue: https://github.com/milvus-io/milvus/issues/43755
Signed-off-by: bigsheeper <yihao.dai@zilliz.com>
issue: #43767
- Enhance serviceable check logic to properly handle full vs partial
result requirements
- For full result (requiredLoadRatio >= 1.0): check
queryView.Serviceable()
- For partial result (requiredLoadRatio < 1.0): check load ratio
satisfaction
- Add comprehensive unit tests covering all serviceable check scenarios
This enhancement ensures delegator correctly validates serviceability
based on the requested result completeness, improving reliability of
query operations.
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
issue: #43745
Add timestamp filtering capability to L0Reader to match the
functionality available in the regular Reader. This enhancement allows
filtering delete records based on timestamp range during L0 import
operations.
Changes include:
- Add tsStart and tsEnd fields to l0Reader struct for timestamp
filtering
- Modify NewL0Reader function signature to accept tsStart and tsEnd
parameters
- Implement timestamp filtering logic in Read method to skip records
outside the specified range
- Update L0ImportTask and L0PreImportTask to parse timestamp parameters
from request options and pass them to NewL0Reader
- Add comprehensive test case TestL0Reader_ReadWithTsFilter to verify ts
filtering functionality using mockey framework
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
Related to #43725
This patch add assertion preventing segment reloading same field column.
Also improve the message info when pk already exists.
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #41435
After introducing the caching layer's lazy loading and eviction
mechanisms, most parts of a segment won't be loaded into memory or disk
immediately, even if the segment is marked as LOADED. This means
physical resource usage may be very low. However, we still need to
reserve enough resources for the segments marked as LOADED. Thus, the
logic of resource usage estimation during segment loading, which based
on physcial resource usage only for now, should be changed.
To address this issue, we introduced the concept of logical resource
usage in this patch. This can be thought of as the base reserved
resource for each LOADED segment.
A segment’s logical resource usage is derived from its final evictable
and inevictable resource usage and calculated as follows:
```
SLR = SFPIER + evitable_cache_ratio * SFPER
```
it also equals to
```
SLR = (SFPIER + SFPER) - (1.0 - evitable_cache_ratio) * SFPER
```
`SLR`: The logical resource usage of a segment.
`SFPIER`: The final physical inevictable resource usage of a segment.
`SFPER`: The final physical evictable resource usage of a segment.
`evitable_cache_ratio`: The ratio of a segment's evictable resources
that can be cached locally. The higher the ratio, the more physical
memory is reserved for evictable memory.
When loading a segment, two types of resource usage are taken into
account.
First is the estimated maximum physical resource usage:
```
PPR = HPR + CPR + SMPR - SFPER
```
`PPR`: The predicted physical resource usage after the current segment
is allowed to load.
`HPR`: The physical resource usage obtained from hardware information.
`CPR`: The total physical resource usage of segments that have been
committed but not yet loaded. When one new segment is allow to load,
`CPR' = CPR + (SMR - SER)`. When one of the committed segments is
loaded, `CPR' = CPR - (SMR - SER)`.
`SMPR`: The maximum physical resource usage of the current segment.
`SFPER`: The final physical evictable resource usage of the current
segment.
Second is the estimated logical resource usage, this check is only valid
when eviction is enabled:
```
PLR = LLR + CLR + SLR
```
`PLR`: The predicted logical resource usage after the current segment is
allowed to load.
`LLR`: The total logical resource usage of all loaded segments. When a
new segment is loaded, `LLR` should be updated to `LLR' = LLR + SLR`.
`CLR`: The total logical resource usage of segments that have been
committed but not yet loaded. When one new segment is allow to load,
`CLR' = CLR + SLR`. When one of the committed segments is loaded, `CLR'
= CLR - SLR`.
`SLR`: The logical resource usage of the current segment.
Only when `PPR < PRL && PLR < PRL` (`PRL`: Physical resource limit of
the querynode), the segment is allowed to be loaded.
---------
Signed-off-by: Shawn Wang <shawn.wang@zilliz.com>
`localDataRootPath` is used to init local chunk manager and has
`querynode` appended to it, thus is incorrect
#41435
Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
Related to #43660
This patch reduces the unwanted offset&ts entries having same timestamp
of delete record. Under large amount of upsert, this false hit could
increase large amount of memory usage while applying delete.
The next step could be passing a callback to `search_pk_func_` to handle
hit entry streamingly.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Related to #43655
This patch add a padding when writing mmap file for ScalarSortedIndex in
case of mmap falure due to 0 mmap length.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Related to #43626
Similar to previous pr #43321, null bitmap could be dislocated if the
bitset ptr does not count the offset of array
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Related to #43592
When delete records are large, search pk one by one will result into
many `Pincells` call which creates lots of futures.
This patch make search pk execute in batch to reduce this cost.
Also add `GetAllChunks` API to utilize `PinAllCells` to reduce pins.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Replace multiple per-table flush RPC calls with single FlushAll RPC to
improve performance in multi-table scenarios.
issue: #43338
- Implement server-side FlushAll request processing in
DataCoord/MixCoord
- Add flushAllTask to handle unified flush operations across all tables
- Replace proxy-side per-table flush iteration with single RPC call
- Support both streaming and non-streaming service execution paths
- Add comprehensive unit tests for new FlushAll implementation
---------
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
Related to #43584
There might be concurrent calls on `translator.get_cells`. The channel
cannot be shared among these calls, otherwise the logic will break.
This patch create new channel for each `get_cells` invocation in case of
data race.
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Related to #43584
Directly accessing `fields_` in `get_raw_data` may have race if load vec
index happens concurrently during getting raw data.
This PR make `bulk_subscript` hold shared_ptr of field column prevent
field column being release during reading it.
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
relate: https://github.com/milvus-io/milvus/issues/42792
Add char group tokenizer which support use costum char group or use some
build-in char group as delimiters.
Signed-off-by: aoiasd <zhicheng.yue@zilliz.com>
Related to #43584
When `LoadWithStrategy` throw exception, the ex was wrapped in the
returned future. If the future is not handled, this exception would be
ignored.
This patch add `future.get()` to get exception if any.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
1. Use blocking memory allocation to wait until memory becomes available
2. Perform memory allocation at the file level instead of per task
3. Limit Parquet file reader batch size to prevent excessive memory
consumption
4. Limit import buffer size from 20% to 10% of total memory
issue: https://github.com/milvus-io/milvus/issues/43387,
https://github.com/milvus-io/milvus/issues/43131
---------
Signed-off-by: bigsheeper <yihao.dai@zilliz.com>
Add `arrowBuild.Reserve` call for `ValueSerializer` to reduce repeated
resizing buffer when write size is large
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #43360
Enhance the partial result evaluation mechanism in delegator to use row
count based data ratio instead of simple segment count ratio for better
accuracy.
Key improvements:
- Introduce PartialResultEvaluator interface for flexible evaluation
strategy
- Implement NewRowCountBasedEvaluator using sealed segment row count
data
- Replace segment count based ratio with row count based data ratio
calculation
- Update PinReadableSegments to return sealedRowCount information
- Modify executeSubTasks to use configurable evaluator for partial
result decisions
- Add comprehensive unit tests for the new row count based evaluation
logic
This change provides more accurate partial result evaluation by
considering the actual data volume rather than just segment quantity,
leading to better query performance and consistency when some segments
are unavailable.
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
issue: #42688
- The channel cp is dropped by garbage collector
- The channel is dropped and the cp is marked as math.Uint64
- If we drop it here, the update channel checkpoints will write the
dirty cp back.
Signed-off-by: chyezh <chyezh@outlook.com>
Ref https://github.com/milvus-io/milvus/issues/42148https://github.com/milvus-io/milvus/pull/42406 impls the segcore part of
storage for handling with VectorArray.
This PR:
1. impls the go part of storage for VectorArray
2. impls the collection creation with StructArrayField and VectorArray
3. insert and retrieve data from the collection.
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
Signed-off-by: SpadeA-Tang <tangchenjie1210@gmail.com>
Signed-off-by: SpadeA-Tang <u6748471@anu.edu.au>
issue: #43261
`promise->setValue(folly::Unit());` may run callbacks inline and some of
them may attempt to grab `mtx_`. So we should not call
`promise->setValue(folly::Unit());` while holding the lock.
---------
Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
This parameter determines whether the returned value should be a copy or
a reference from the arrow array. The updates enhance memory management
and provide more control over data handling during deserialization.
See #43186
---------
Signed-off-by: Ted Xu <ted.xu@zilliz.com>
Related to #43261
Read error with catched in `LoadWithStrategy`. Caller could not detect
read failure when some error occurred. This patch make
`LoadWithStrategy` throw ex instead of swallowing it.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
We’ve frequently observed data loss caused by broken mutual exclusion in
compaction tasks. This PR introduces a post-check: before modifying
metadata upon compaction task completion, it verifies the state of the
input segments. If any input segment has been dropped, the compaction
task will be marked as failed.
issue: https://github.com/milvus-io/milvus/issues/43513
---------
Signed-off-by: bigsheeper <yihao.dai@zilliz.com>
See: #43186
In this PR:
1. Flush renamed to FlushChunk, while a new Flush primitive is
introduced to serialize values to records.
2. Segment mapping in clustering compaction now process data by records
instead of values, it calls flush to all buffers after each record is
processed.
Signed-off-by: Ted Xu <ted.xu@zilliz.com>
issue: #43088
issue: #43038
The current loading process:
* When loading an index, we first download the index files into a list
of buffers, say A
* then constructing(copying) them into a vector of FieldDatas(each file
is a FieldData), say B
* assembles them together as a huge BinarySet, say C
* lastly, copy into the actual index data structure, say D
The problem:
* We can see that, after each step, we don't need the data in previous
step.
* But currently, we release the memory of A, B, C only after we have
finished constructing D
* This leads to a up to 4x peak memory usage comparing with the raw
index size, during the loading process
* This PR allows timely releasing of B after we assembled C. So after
this PR, the peak memory usage during loading will be up to 3x of the
raw index size.
I will create another PR to release A after we created B, that seems
more complicated and need more work.
Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
Realted #43520
Datanode may have memory leakage when reader returns error. In
previously mention issue, datanodes got OOM killed due to continueous
error in read path.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
related: #42417
- update the isValidUsername function to accept dots and hyphens in
addition to letters, digits, and underscores
- this change improves compatibility with common username formats and
addresses feedback in issue #42417
Signed-off-by: Jean-Francois Weber-Marx <jfwm@hotmail.com>
Signed-off-by: Jean-Francois Weber-Marx <jf.webermarx@criteo.com>
issue: #42884
Fixes an issue where delete records for a segment are lost from the
delete buffer if `load segment` execution on the delegator is too slow,
causing `syncTargetVersion` or other cleanup operations to clear them
prematurely.
Changes include:
- Introduced `Pin` and `Unpin` methods in `DeleteBuffer` interface and
its implementations (`doubleCacheBuffer`, `listDeleteBuffer`).
- Added a `pinnedTimestamps` map to track timestamps protected from
cleanup by specific segments.
- Modified `LoadSegments` in `shardDelegator` to `Pin` relevant segment
delete records before loading and `Unpin` them afterwards.
- Added `isPinned` check in `UnRegister` and `TryDiscard` methods of
`listDeleteBuffer` to skip cleanup if corresponding timestamps are
pinned.
- Added comprehensive unit tests for `Pin`, `Unpin`, and `isPinned`
functionality, covering basic, multiple pins, concurrent, and edge
cases.
This ensures the integrity of delete records by preventing their
premature removal from the delete buffer during segment loading.
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
Related to #43522
Currently, passing partial schema to storage v2 packed reader may
trigger SEGV during clustering compaction unit test.
This patch implement `NeededFields` differently in each `RecordReader`
imlementation. For now, v2 will implemented as no-op. This will be
supported after packed reader support this API.
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #43072, #43289
- manage the schema version at recovery storage.
- update the schema when creating collection or alter schema.
- get schema at write buffer based on version.
- recover the schema when upgrading from 2.5.
---------
Signed-off-by: chyezh <chyezh@outlook.com>
1. Modify the binlog reader to stop reading a fixed 4096 rows and
instead use the calculated bufferSize to avoid generating small binlogs.
2. Use a fixed bufferSize (32MB) for the Parquet reader to prevent OOM.
issue: https://github.com/milvus-io/milvus/issues/43387
---------
Signed-off-by: bigsheeper <yihao.dai@zilliz.com>
currently we multiplied the requesting size when adding to loading, but
did not do so when estimating projected usage.
issue: #43088
Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
Related to #43113
This PR:
- Change member of FieldIndex from `FieldMeta &` to needed `DataType`
and dim member resolving dangling reference after schema change
- Add double check after acquiring lock to reduce multiple assignment
- Change `auto schema` to `auto& schema` to reduce schema copy
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
The root cause of the issue lies in the fact that when a sealed segment
contains multiple row groups, the get_cells function may receive
unordered cids. This can result in row groups being written into
incorrect cells during data retrieval.
Previously, this issue was hard to reproduce because the old Storage V2
writer had a bug that caused it to write row groups larger than 1MB.
These large row groups could lead to uncontrolled memory usage and
eventually an OOM (Out of Memory) error. Additionally, compaction
typically produced a single large row group, which avoided the incorrect
cell-filling issue during query execution.
related: https://github.com/milvus-io/milvus/issues/43388,
https://github.com/milvus-io/milvus/issues/43372,
https://github.com/milvus-io/milvus/issues/43464, #43446, #43453
---------
Signed-off-by: shaoting-huang <shaoting.huang@zilliz.com>
This patch makes `PackedReader` return EOF when try to calling
`ReadNext` after closing it.
This behavior make importv2.binlog reader could retry after EOF reached
and act normally.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
This PR fill default value for `PackedBinlogRecordWriter` timestamp
range so target segment meta will contains correct timestamp range
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #43434
- the segment start position can be carried by other segment sync
operation. so the sync start position operation can happens before
insert.
- TODO: It's a wired design should be removed.
Signed-off-by: chyezh <chyezh@outlook.com>
issue: #41435
this is to prevent AI from thinking of our exception throwing as a
dangerous PANIC operation that terminates the program.
Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
issue: https://github.com/milvus-io/milvus/issues/41435
turns out we have per file binlog size in golang code, by passing it
into segcore we can support eviction in storage v1
Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
Realted to #43407
When `MultiSaveAndRemove` like ops contains same key in saves and
removal keys it may cause data lost if the execution order is save first
than removal.
This PR make all the kv execute removal first then save the new values.
Even when same key appeared in both saves and removals, the new value
shall stay.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #41435
issue: https://github.com/milvus-io/milvus/issues/43038
This PR also:
1. removed ERROR state from ListNode
2. CacheSlot will do reserveMemory once for all requested cells after
updating the state to LOADING, so now we transit a cell to LOADING
before its resource reservation
3. reject resource reservation directly if size >= max_size
---------
Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
issue: #42995
- don't balance the wal if the producing-consuming lag is too long.
- don't balance if the rebalance is set as false.
- don't balance if the wal is balanced recently.
Signed-off-by: chyezh <chyezh@outlook.com>
issue: #43117, #42966, #43373
- also fix channel balance may not work at 2.6.
- fix error lost at delete path
- add mvcc into s/q log
- change the log level for TestCoordDownSearch
Signed-off-by: chyezh <chyezh@outlook.com>
Previous code uses diskSegmentMaxSize if and only if all of the
collection's vector fields are indexed with DiskANN index.
When introducing sparse vectors, since sparse vector cannot be indexed
with DiskANN index, collections with both dense and sparse vectors will
use maxSize instead.
This PR changes the requirments of using diskSegmentMaxSize to all dense
vectors are indexed with DiskANN indexs, ignoring sparse vector fields.
See also: #43193
Signed-off-by: yangxuan <xuan.yang@zilliz.com>
fix: https://github.com/milvus-io/milvus/issues/43354
The current implementation of stdsort index is not supported for
std::string. Remove the code.
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
Correct read and buffer size to 64MB to prevent OOM during clustering
compaction.
issue: https://github.com/milvus-io/milvus/issues/43310
---------
Signed-off-by: bigsheeper <yihao.dai@zilliz.com>
Related to #43262
This patch fixes following logic bug:
- When multiple chunks are loaded and size cannot be divided by 8, just
appending uint8_t as bitmap will cause null bitmap dislocation
- `null_bitmap_data()` points to start of whole row group, which may not
stand for current `arrow::Array`
The current solutions is:
- Reorganize the null_bitmap with currect size & offset
- Pass `array->offset()` in tuple to info the current offset
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Ref https://github.com/milvus-io/milvus/issues/42053
This PR enable ngram to support more kinds of matches such as prefix and
postfix match.
---------
Signed-off-by: SpadeA <tangchenjie1210@gmail.com>
Related to #43250
Use FieldIDList to check missing field. If column is missing, return
empty resultset
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #41570
Fix issue where growing and sealed segments could be searched
simultaneously, causing inflated count(*) results. This was caused by
logic introduced in PR #42009 that made sealed segments readable before
target version advancement.
Changes include:
- Fix conditional filtering logic in PinReadableSegments to prevent
sealed segments from becoming readable prematurely
- Use target version filter for full results (ratio=1.0) to ensure
sealed segments only become readable after target advancement
- Use query view segment list filter for partial results (ratio<1.0) to
maintain backward compatibility
- Simplify target version setting in AddDistributions to prevent
premature segment readability
- Add logging for redundant growing segments during sync
- Add comprehensive unit tests covering the duplicate segment scenario
This fix ensures count(*) queries return accurate results by preventing
the same segment from being counted in both growing and sealed states.
---------
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
issue: https://github.com/milvus-io/milvus/issues/42900
@sunby Unfortunately, it is not that easy to fix as it was thought in
#43177
Upd: also handles `Inf` and `NaN` values, and the division by zero case
for `fp32` and `fp64`
Signed-off-by: Alexandr Guzhva <alexanderguzhva@gmail.com>
issue: #42995
- The consuming lag at streaming node will be reported to coordinator.
- The consuming lag will trigger the write limit and deny by quota
center.
- Set the ttProtection by default.
---------
Signed-off-by: chyezh <chyezh@outlook.com>
Related to #43113
When schema change happens, insert shall not happen, otherwise:
- Data race may happen causing insertion failure
- Inconsistent data schema
This PR add shared_lock prevent this data race.
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #43107
- Add checkLoadConfigChanges() to apply load config during startup
- Call config check in startQueryCoord() after restart
- Skip auto-updates for collections with user-specified replica numbers
- Add is_user_specified_replica_mode field to preserve user settings
- Add comprehensive unit tests with mockey
Ensures existing collections use latest cluster-level config after
restart.
---------
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
Related to #39178
This PR add logs for segment schema change operations.
Also fixes the nit comments from PR #42490
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Alexander Guzhva
liliu-z
congqixia
aoiasd
wei liu
Gao
Zhen Ye
sthuang
Xianhui Lin
yihao.dai
foxspy
PjJinchen
Chun Han
zhagnlu
cai.zhang
sparknack
Buqian Zheng
Bingyi Sun
Ted Xu
SimFG
XuanYang-cn
Xiaofan
Spade A
Shuyoou
Jean-Francois Weber-Marx
junjiejiangjjj
tinswzy