Cherry-pick from master
pr: #45969
Related to #45960
Follow-up to #45961
After #45961 ensured that handleNodeUp is always called for nodes
discovered during rewatchNodes (including stopping nodes), this change
adds a safeguard in ResourceManager.handleNodeUp to skip adding stopping
nodes to resource groups.
1. **resource_manager.go**: Add check for IsStoppingState() in
handleNodeUp to prevent stopping nodes from being added to incomingNode
set and assigned to resource groups.
2. **server.go**:
- Delete processed nodes from sessionMap to avoid duplicate processing
in the subsequent loop
- Add warning logs for stopping state transitions during rewatch
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Cherry-pick from master
pr: #45147
Related to #44509
Fix a bug where QueryNodeNumEntities metrics were not updated for
collections with zero segments, causing stale metrics when all segments
are flushed or compacted.
The previous implementation used separate loops: one to update size
metrics for all collections, and another to update num entities metrics
only for collections present in the grouped segments map. Collections
with no segments were skipped in the second loop, leaving their
NumEntities metrics stale.
Changes:
- Consolidate size and num entities metric updates into single loop
- Iterate over all collections instead of grouped segments
- Get collection metadata from manager instead of segment instances
- Correctly set NumEntities to 0 for collections with no segments
- Apply the same fix to both growing and sealed segment processing
- Add nil check for collection metadata before processing
This ensures all collection metrics are updated consistently, even when
segment count drops to zero.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Cherry-pick from master
pr: #45961
Related to #45960
When QueryCoord restarts or reconnects to etcd, the rewatchNodes
function previously skipped handleNodeUp for QueryNodes in stopping
state. This caused stopping balance to fail because necessary components
were not initialized:
- Task scheduler executor was not added
- Dist handler was not started
- Node was not registered in resource manager
This fix ensures handleNodeUp is always called for new nodes regardless
of their stopping state, followed by handleNodeStopping if the node is
stopping. This allows the graceful shutdown process to correctly migrate
segments and channels away from stopping nodes.
Signed-off-by: Congqi Xia <congqi.xia@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
pr: https://github.com/milvus-io/milvus/pull/45743
Signed-off-by: Buqian Zheng <zhengbuqian@gmail.com>
issue: #45623
When etcd reconnects, the DataCoord rewatches DataNodes and calls
ChannelManager.Startup again without closing the previous instance. This
causes multiple contexts and goroutines to accumulate, leading to Close
hanging indefinitely waiting for untracked goroutines.
Root cause:
- Etcd reconnection triggers rewatch flow and calls Startup again
- Startup was not idempotent, allowing repeated calls
- Multiple context cancellations and goroutines accumulated
- Close would wait indefinitely for untracked goroutines
Changes:
- Add started field to ChannelManagerImpl
- Refactor Startup to check and handle restart scenario
- Add state check in Close to prevent hanging
---------
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
When there're no growing segments in the collection, L0 Compaction will
try to choose all L0 segments that hits all L1/L2 segments.
However, if there's Sealed Segment still under flushing in DataNode at
the same time L0 Compaction selects satisfied L1/L2 segments, L0
Compaction will ignore this Segment because it's not in "FlushState",
which is wrong, causing missing deletes on the Sealed Segment.
This quick solution here is to fail this L0 compaction task once
selected a Sealed segment.
See also: #45339
pr: #45340
pr: #45341
Signed-off-by: yangxuan <xuan.yang@zilliz.com>
Cherry-pick from master
pr: #45037Fixes#45035
This commit addresses a data race issue where refreshCollection was
updating the collection notifier without proper lock protection.
Changes:
- Add UpdateCollection method to CollectionManager with proper locking
- Introduce CollectionOperator pattern for thread-safe collection
updates
- Make setRefreshNotifier private and use it through the operator
pattern
- Update refreshCollection to use the new UpdateCollection method
- Handle collection not found error gracefully in refreshCollection
The CollectionOperator pattern ensures all collection modifications go
through the CollectionManager's lock, preventing concurrent access
issues.
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #44909
pr: #44917
When requery optimization is enabled, search results contain IDs but
empty FieldsData. During reduce/rerank operations, if the first shard
has empty FieldsData while others have data, PrepareResultFieldData
initializes an empty array, causing AppendFieldData to panic when
accessing array indices.
Changes:
- Find first non-empty FieldsData as template in 5 functions:
reduceAdvanceGroupBY, reduceSearchResultDataWithGroupBy,
reduceSearchResultDataNoGroupBy, rankSearchResultDataByGroup,
rankSearchResultDataByPk
- Add length check before 4 AppendFieldData calls to prevent panic
- Add unit tests for empty and partial empty FieldsData scenarios
This fix handles both pure requery (all empty) and mixed scenarios
(some empty, some with data) without breaking normal search flow.
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
issue: #44961
master pr: #44966
This PR fixes 3 geometry related bugs:
1. Implement ToString interface for GisFunctionFilter.
2. Ignore GisFunctionFilter MoveCursor for growing segment.
3. Don't skip null geometry for building R-Tree index, should be record
in null_offsets.
---------
Signed-off-by: Cai Zhang <cai.zhang@zilliz.com>
issue: #44802
master pr: #44889
After a Geometry object is serialized into WKB, the resulting binary may
contain '\0' bytes.
When growing mmap is enabled, the append data logic uses strcpy, which
stops copying at the first '\0' bytes.
This causes only part of the WKB---typically the portion up to the
geometry type field to be copied, leading to corrupted data.
As a result, during parsing, all POINT geometries are incorrectly
interperted as POINT(0 0).
To fix this issue, memcpy will be used instead of strcpy.
Signed-off-by: Cai Zhang <cai.zhang@zilliz.com>
Cherry-pick from master
pr: #44870
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: #43858
pr: #44834
Fix the issue introduced in PR #43992 where deactivating the balance
checker incorrectly stops stopping balance operations.
Changes:
- Move IsActive() check after stopping balance logic
- Only skip normal balance when checker is inactive
- Allow stopping balance to proceed regardless of checker state
This ensures stopping balance can execute even when the balance checker
is deactivated.
---------
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
issue: #43828
pr: #43829#43909
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>
Co-authored-by: Zhen Ye <chyezh@outlook.com>
Cherry-pick from master
pr: #44789
This commit addresses issue #44788 where the
`datacoord_stored_binlog_size` metric could become inaccurate when
multiple concurrent `GetMetrics` calls arrived at DataCoord.
### Problem
The original implementation called `Reset()` followed by `Add()`
operations on Prometheus metrics within the `GetQuotaInfo()` method.
When multiple goroutines invoked this method concurrently, race
conditions occurred:
- Thread 1: Reset() → Add(value1)
- Thread 2: Reset() → Add(value2)
- Result: Metrics could be reset multiple times and values added in an
interleaved fashion, leading to inaccurate and inflated metric values
### Solution
Changed the approach from `Reset() + Add()` to aggregating metric values
in local maps first, then using `Set()` to update metrics atomically:
1. Collect segment size data into local maps:
- `storedBinlogSize`: tracks size per collection per segment state
- `binlogFileSize`: tracks total file count per collection
- `coll2DbName`: maps collection IDs to database names
2. After aggregation is complete, use `Set()` (instead of `Add()`) to
update metrics in a single operation per label combination
This ensures that concurrent `GetMetrics` calls don't interfere with
each other, as each invocation works with its own local state and only
updates the final metric value atomically.
---------
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
Cherry-pick from master
pr: #44723
Related to #36672
Add accesslog field displaying value length for search/query request may
help developers debug related issues
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #44699
Fix the issue where L0 segment checking logic incorrectly identifies L0
segments as missing when they exist on multiple delegators during
channel balance process, which blocks sealed segment loading and target
progression.
Changes include:
- Replace GetLatestShardLeaderByFilter with GetByFilter to check all
delegators instead of only the latest leader
- Iterate through all delegator views to identify which ones lack the L0
segment
The original logic only checked the latest shard leader, causing false
positive detection of missing L0 segments when they actually exist on
other delegators in the same channel during balance operations. This led
to continuous generation of duplicate L0 segment load tasks, preventing
normal sealed segment loading flow.
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
issue: #43427
master pr: #44606
The GISFunction asserts that the segment_offsets cannot be nullptr. When
size is 0, the segment_offsets is nullptr, so the loop is skiped.
Signed-off-by: Cai Zhang <cai.zhang@zilliz.com>
Cherry-pick from master
pr: #44706
Related to #44703
This PR:
- Add `SetActualConsistencyLevel` to `info.AccessInfo` interface and
related util method processing it
- Make `$consistency_level` returning actual value if set
---------
Signed-off-by: Congqi Xia <congqi.xia@zilliz.com>
issue: #43858
pr: #43992
Refactor the balance checker implementation to use priority queues for
managing collection balance operations, improving processing efficiency
and order control.
Changes include:
- Export priority queue interfaces (Item, BaseItem, PriorityQueue)
- Replace collection round-robin with priority-based queue system
- Add BalanceCheckCollectionMaxCount configuration parameter
- Optimize balance task generation with batch processing limits
- Refactor processBalanceQueue method for different strategies
- Enhance test coverage with comprehensive unit tests
The new priority queue system processes collections based on row count
or collection ID order, providing better control over balance operation
priorities and resource utilization.
---------
Signed-off-by: Wei Liu <wei.liu@zilliz.com>
