Release Notes
Find out what’s new in Milvus! This page summarizes information about new features, improvements, known issues, and bug fixes in each release. You can find the release notes for each released version after v2.2.0 in this section. We suggest that you regularly visit this page to learn about updates.
2.2.9
Release date: 2 June, 2023
| Milvus version | Python SDK version | Java SDK version | Go SDK version | Node.js SDK version |
|---|---|---|---|---|
| 2.2.9 | 2.2.9 | 2.2.5 | 2.2.3 | 2.2.11 |
Milvus 2.2.9 has added JSON support, allowing for more flexible schemas within collections through dynamic schemas. The search efficiency has been improved through partition keys, which enable data separation for different data categories, such as multiple users, in a single collection. Additionally, database support has been integrated into Role-Based Access Control (RBAC), further fortifying multi-tenancy management and security. Support has also been extended to Alibaba Cloud OSS, and connection management has been refined, resulting in an improved user experience.
As always, this release includes bug fixes, enhancements, and performance improvements. Notably, disk usage has been significantly reduced, and performance has been improved, particularly for filtered searches.
We hope you enjoy the latest release!
New Features
-
JSON support
-
Dynamic schema
-
Partition key
-
Database support in RBAC
-
Connection management
-
Alibaba Cloud OSS support
-
Additional features
Bug fixes
- Added temporary disk data cleaning upon the start of Milvus (#24400).
- Fixed crash issue of bulk insert caused by an invalid Numpy array file (#24480).
- Fixed an empty result set type for Int8~Int32 (#23851).
- Fixed the panic that occurs while balancing releasing a collection (#24003) (#24070).
- Fixed an error that occurs when a role removes a user that has already been deleted (#24049).
- Fixed an issue where session stop/goingStop becomes stuck after a lost connection (#23771).
- Fixed the panic caused by incorrect logic of getting unindexed segments (#24061).
- Fixed the panic that occurs when a collection does not exist in quota effect (#24321).
- Fixed an issue where refresh may be notified as finished early (#24438) (#24466).
Enhancement
-
Added an error response to return when an unimplemented request is received (#24546)
-
Reduced disk usage for Milvus Lite and Standalone:
-
Optimized quota to avoid OOM on search
-
Added consistency_level in search/query request (#24541)
-
(pr24562) Supported search with default parameters (#24516)
-
Put DataNode load statslog lazy if SkipBFStatsLog is true (#23779)
-
Put QueryNode lazy load statslog if SkipBFLoad is true (#23904)
-
Fixed concurrent map read/write in rate limiter (#23957)
-
Improved load/release performance:
-
Optimized PrivilegeAll permission check (#23972)
-
Fixed the "not shard leader" error when gracefully stopping (#24038)
-
Lowered the task merge cap to mitigate an insufficient memory error (#24233)
-
Removed constraint that prevents creating an index after load (#24415)
-
Optimized the search performance with a high filtering ratio (#23948)
Performance improvements
- Added SIMD support for several filtering expressions (#23715, #23781).
- Reduced data copying during insertion into growing segments (#24492).
v2.2.8
Release date: 3 May, 2023
| Milvus version | Python SDK version | Java SDK version | Go SDK version | Node.js SDK version |
|---|---|---|---|---|
| 2.2.8 | 2.2.8 | 2.2.5 | 2.2.2 | 2.2.8 |
In this update, we fixed 1 critical bug.
Bugfix
- Fixed RootCoord panic caused by the upgrades from v2.2.x to v2.2.7 (#23828).
v2.2.7
Release date: 28 April, 2023
| Milvus version | Python SDK version | Java SDK version | Go SDK version | Node.js SDK version |
|---|---|---|---|---|
| 2.2.7 | 2.2.8 | 2.2.5 | 2.2.2 | 2.2.7 |
In this update, we have focused on resolving various issues reported by our users, enhancing the software's overall stability and functionality. Additionally, we have implemented several optimizations, such as load balancing, search grouping, and memory usage improvements.
Bugfix
- Fixed a panic caused by not removing metadata of a dropped segment from the DataNode. (#23492)
- Fixed a bug that caused forever blocking due to the release of a non-loaded partition. (#23612)
- To prevent the query service from becoming unavailable, automatic balancing at the channel level has been disabled as a workaround. (#23632) (#23724)
- Cancel failed tasks in the scheduling queue promptly to prevent an increase in QueryCoord scheduling latency. (#23649)
- Fixed compatibility bug and recalculate segment rows to prevent service queries from being unavailable. (#23696)
- Fixed a bug in the superuser password validation logic. (#23729)
- Fixed the issue of shard detector rewatch failure, which was caused by returning a closed channel. (#23734)
- Fixed a loading failure caused by unhandled interrupts in the AWS SDK. (#23736)
- Fixed the "HasCollection" check in DataCoord. (#23709)
- Fixed the bug that assigned all available nodes to a single replica incorrectly. (#23626)
Enhancement
- Optimized the display of RootCoord histogram metrics. (#23567)
- Reduced peak memory consumption during collection loading. (#23138)
- Removed unnecessary handoff event-related metadata. (#23565)
- Added a plugin logic to QueryNode to support the dynamic loading of shared library files. (#23599)
- Supports load balancing with replica granularity. (#23629)
- Released a load-balancing strategy based on scores. (#23805)
- Added a coroutine pool to limit the concurrency of cgo calls triggered by "delete". (#23680)
- Improved the compaction algorithm to make the distribution of segment sizes tend towards the ideal value. (#23692)
- Changed the default shard number to 1. (#23593)
- Improved search grouping algorithm to enhance throughput. (#23721)
- Code refactoring: Separated the read, build, and load DiskANN parameters. (#23722)
- Updated etcd and Minio versions. (#23765)
v2.2.6
Release date: 18 April, 2023
| Milvus version | Python SDK version | Java SDK version | Go SDK version | Node.js SDK version |
|---|---|---|---|---|
| 2.2.6 | 2.2.7 | 2.2.5 | 2.2.1 | 2.2.4 |
You are advised to refrain from using version 2.2.5 due to several critical issues that require immediate attention. Version 2.2.6 addresses these issues. One of the critical issues is the inability to recycle dirty binlog data. We highly recommend using version 2.2.6 version instead of version 2.2.5 to avoid any potential complications.
If you hit the issue where data on object storage cannot be recycled, upgrade your Milvus to v2.2.6 to fix these issues.
Bugfix
- Fixed the problem of DataCoord GC failure (#23298)
- Fixed the problem that index parameters passed when creating a collection will override those passed in subsequent create_index operations (#23242)
- Fix the problem that the message backlog occurs in RootCoord, which causes the delay of the whole system to increase (#23267)
- Fixed the accuracy of metric RootCoordInsertChannelTimeTick (#23284)
- Fixed the issue that the timestamp reported by the proxy may stop in some cases (#23291)
- Fixed the problem that the coordinator role may self-destruct by mistake during the restart process (#23344)
- Fixed the problem that the checkpoint is left behind due to the abnormal exit of the garbage collection goroutine caused by the etcd restart (#23401)
Enhancement
- Added slow logging performance for query/search when the latency is not less than 5 seconds (#23274)
v2.2.5
Release date: 29 March, 2023
| Milvus version | Python SDK version | Java SDK version | Go SDK version | Node.js SDK version |
|---|---|---|---|---|
| 2.2.5 | 2.2.5 | 2.2.4 | 2.2.1 | 2.2.4 |
Security
Fixed MinIO CVE-2023-28432 by upgrading MinIO to RELEASE.2023-03-20T20-16-18Z.
New Features
-
First/Random replica selection policy
This policy allows for a random replica selection if the first replica chosen under the round-robin selection policy fails. This improves the throughput of database operations.
Bug fixes
-
Fixed index data loss during the upgrade from Milvus 2.2.0 to 2.2.3 or higher.
-
Fixed DataCoord Out-of-Memory (OOM) with large fresh pressure.
-
Fixed a concurrency issue in the LRU cache that was caused by concurrent queries with specified output fields.
-
Fixed shard leader cache
-
Other fixes
- Fixed query performance issue with a large number of segments (#23028)
- Fixed an issue to enable batch delete files on GCP of MinIO (#23052) (#23083)
- Fixed flush delta buffer if SegmentID equals 0 (#23064)
- fixed unassigned from resource group (#22800)
- Fixed load partition timeout logic still using createdAt (#23022)
- Fixed unsub channel always removes QueryShard (#22961)
Enhancements
- Added memory Protection by using the buffer size in memory synchronization policy (#22797)
- Added dimension checks upon inserted records (#22819) (#22826)
- Added configuration item to disable BF load (#22998)
- Aligned the maximum dimensions of the DisANN index and that of a collection (#23027)
- Added checks whether all columns aligned with same num_rows (#22968) (#22981)
- Upgraded Knowhere to 1.3.11 (#22975)
- Added the user RPC counter (#22870)
v2.2.4
Release date: 17 March, 2023
| Milvus version | Python SDK version | Java SDK version | Go SDK version | Node.js SDK version |
|---|---|---|---|---|
| 2.2.4 | 2.2.3 | 2.2.3 | 2.2.1 | 2.2.4 |
Milvus 2.2.4 is a minor update to Milvus 2.2.0. It introduces new features, such as namespace-based resource grouping, collection-level physical isolation, and collection renaming.
In addition to these features, Milvus 2.2.4 also addresses several issues related to rolling upgrades, failure recovery, and load balancing. These bug fixes contribute to a more stable and reliable system.
We have also made several enhancements to make your Milvus cluster faster and consume less memory with reduced convergence time for failure recovery.
New Features
-
Resource grouping
Milvus has implemented resource grouping for QueryNodes. A resource group is a collection of QueryNodes. Milvus supports grouping QueryNodes in the cluster into different resource groups, where access to physical resources in different resource groups is completely isolated. See Manage Resource Group for more information.
-
Collection renaming
The Collection-renaming API provides a way for users to change the name of a collection. Currently, PyMilvus supports this API, and SDKs for other programming languages are on the way. See Rename a Collection for details.
-
Google Cloud Storage support
Milvus now supports Google Cloud Storage as the object storage.
-
New option to the search and query APIs
If you are more concerned with performance rather than data freshness, enabling this option will skip search on all growing segments and offer better search performance under the scenario search with insertion. See search() and query() for details.
Bug fixes
- Fixed segment not found when forwarding delete to empty segment (#22528)(#22551)
- Fixed possible broken channel checkpoint in v2.2.2 (#22205) (#22227)
- Fixed entity number mismatch with some entities inserted (#22306)
- Fixed DiskANN recovery failure after QueryNode reboots (#22488) (#22514)
- Fixed search/release on same segment (#22414)
- Fixed file system crash during bulk-loading files prefixed with a '.' (#22215)
- Added tickle for DataCoord watch event (#21193) (#22209)
- Fixed deadlock when releasing segments and removing nodes concurrently (#22584)
- Added channel balancer on DataCoord (#22324) (#22377)
- Fixed balance generated reduce task (#22236) (#22326)
- Fixed QueryCoord panic caused by balancing (#22486)
- Added scripts for rolling update Milvus's component installed with helm (#22124)
- Added
NotFoundTSaferandNoReplicaAvailableto retriable error code (#22505) - Fixed no retires upon gRPC error (#22529)
- Fixed an issue for automatic component state update to healthy after start (#22084)
- Added graceful-stop for sessions (#22386)
- Added retry op for all servers (#22274)
- Fixed metrics info panic when network error happens (#22802)
- Fixed disordered minimum timestamp in proxy's pchan statistics (#22756)
- Fixed an issue to ensure segment ID recovery upon failures to send time-tick (#22771)
- Added segment info retrieval without the binlog path (#22741)
- Added distribution.Peek for GetDataDistribution in case of blocked by release (#22752)
- Fixed the segment not found error (#22739)
- Reset delta position to vchannel in packSegmentLoadReq (#22721)
- Added vector float data verification for bulkinsert and insert (#22729)
- Upgraded Knowhere to 1.3.10 to fix bugs (#22746)
- Fixed RootCoord double updates TSO (#22715) (#22723)
- Fixed confused time-tick logs (#22733) (#22734)
- Fixed session nil point (#22696)
- Upgraded Knowhere to 1.3.10 (#22614)
- Fixed incorrect sequence of timetick statistics on proxy(#21855) (#22560)
- Enabled DataCoord to handle GetIndexedSegment error from IndexCoord (#22673)
- Fixed an issue for Milvus writes flushed segment key only after the segment is flushed (#22667)
- Marked cache deprecated instead of removing it (#22675)
- Updated shard leader cache (#22632)
- Fixed an issue for the replica observer to assign node (#22635)
- Fixed the not found issue when retrieving collection creation timestamp (#22629) (#22634)
- Fixed time-tick running backwards during DDLs (#22617) (#22618)
- Fixed max collection name case (#22601)
- Fixed DataNode tickle not run default (#22622)-
- Fixed DataCoord panic while reading timestamp of an empty segment (#22598)
- Added scripts to get etcd info (#22589)
- Fixed concurrent loading timeout during DiskANN indexing (#22548)
- Fixed an issue to ensure index file not finish early because of compaction (#22509)
- Added
MultiQueryNodestag for resource group (#22527) (#22544)
Enhancement
-
Performance
- Improved query performance by avoiding counting all bits (#21909) (#22285)
- Fixed dual copy of varchar fields while loading (#22114) (#22291)
- Updated DataCoord compaction panic after DataNode update plan to ensure consistency (#22143) (#22329)
- Improved search performance by avoiding allocating a zero-byte vector during searches (#22219) (#22357)
- Upgraded Knowhere to 1.3.9 to accelerate IVF/BF (#22368)
- Improved search task merge policy (#22006) (#22287)
- Refined Read method of MinioChunkManager to reduce IO(#22257)
-
Memory Usage
-
Others
v2.2.3
Release date: 10 February, 2023
| Milvus version | Python SDK version | Java SDK version | Go SDK version | Node.js SDK version |
|---|---|---|---|---|
| 2.2.3 | 2.2.2 | 2.2.3 | 2.2.0 | 2.2.3 |
Milvus 2.2.3 introduces the rolling upgrade capability to Milvus clusters and brings high availability settings to RootCoords. The former greatly reduces the impacts brought by the upgrade and restart of the Milvus cluster in production to the minimum, while the latter enables coordinators to work in active-standby mode and ensures a short failure recovery time of no more than 30 seconds.
In this release, Milvus also ships with a lot of improvements and enhancements in performance, including a fast bulk-insert experience with reduced memory usage and less loading time.
Breaking changes
In 2.2.3, the maximum number of fields in a collection is reduced from 256 to 64. (#22030)
Features
-
Rolling upgrade
The rolling upgrade feature allows Milvus to respond to incoming requests during the upgrade, which is not possible in previous releases. In such releases, upgrading a Milvus instance requires it to be stopped first and then restarted after the upgrade is complete, leaving all incoming requests unanswered.
Related issues:
-
Coordinator HA
Coordinator HA allows Milvus coordinators to work in active-standby mode to avoid single-point of failures.
Related issues:
- HA-related issues identified and fixed in QueryCoordV2 (#21501)
- Auto-registration on the startup of Milvus was implemented to prevent both coordinators from working as the active coordinators. (#21641)
- HA-related issues identified and fixed in RootCoords (#21700)
- Issues identified and fixed in active-standby switchover (#21747)
Enhancements
-
Bulk-insert performance enhanced
- Bulk-insert enhancement implemented (#20986 #21532)
- JSON parser optimized for data import (#21332)
- Stream-reading NumPy data implemented (#21540)
- Bulk-insert progress report implemented (#21612)
- Issues identified and fixed so that Milvus checks indexes before flushes segments before bulk-insert is complete (#21604)
- Issues related to bulk-insert progress identified and fixed (#21668)
- Issues related to bulk-insert report identified and fixed (#21758)
- Issues identified and fixed so that Milvus does not seal failed segments while performing bulk-insert operations. (#21779)
- Issues identified and fixed so that bulk-insert operations do not cause a slow flush (#21918)
- Issues identified and fixed so that bulk-insert operations do not crash the DataNodes (#22040)
- Refresh option added to LoadCollection and LoadPartition APIs (#21811)
- Segment ID update on data import implemented (#21583)
-
Memory usage reduced
-
Monitoring metrics optimized
-
Meta storage performance improved
- Improved ListSegments performance for Datacoord catalog. (#21600)
- Improved LoadWithPrefix performance for SuffixSnapshot. (#21601)
- Removed redundant LoadPrefix requests for Catalog ListCollections. (#21551) (#21594)
- Added A WalkWithPrefix API for MetaKv interface. (#21585)
- Added GC for snapshot KV based on time-travel. (#21417) (#21763)
-
Performance improved
- Upgraded Knowhere to 1.3.7. (#21735)
- Upgraded Knowhere to 1.3.8. (#22024)
- Skipped search GRPC call for standalone. (#21630)
- Optimized some low-efficient code. (#20529) (#21683)
- Fixed fill the string field twice when string index exists. (#21852) (#21865)
- Used all() API for bitset check. (#20462) (#21682)
-
Others
- Implemented the GetLoadState API. (#21533)
- Added a task to unsubscribe dmchannel. (#21513) (#21794)
- Explicitly list the triggering reasons when Milvus denies reading/writing. (#21553)
- Verified and adjusted the number of rows in a segment before saving and passing SegmentInfo. (#21200)
- Added a segment seal policy by the number of binlog files. (#21941)
- Upgraded etcd to 3.5.5. (#22007）
Bug Fixes
-
QueryCoord segment replacement fixed
- Fixed the mismatch of sealed segments IDs after enabling load-balancing in 2.2. (#21322)
- Fixed the sync logic of the leader observer. (#20478) (#21315)
- Fixed the issues that observers may update the current target to an unfinished next target. (#21107) (#21280)
- Fixed the load timeout after the next target updates. (#21759) (#21770)
- Fixed the issue that the current target may be updated to an invalid target. (#21742) (#21762)
- Fixed the issue that a failed node may update the current target to an unavailable target. (#21743)
-
Improperly invalidated proxy cache fixed
-
CheckPoint and GC Related issues fixed
-
Issues related to not being able to use embedded etcd with Milvus fixed
- Added setup config files for embedded etcd (#22076)
-
Others
- Fixed the offset panic in queries (#21292) (#21751)
- Fixed the issue that small candidate compaction should only happen with more than one segment (#21250)
- Fixed the issue of memory usage calculation (#21798)
- Fixed the issue that a timestamp allocation failure blocks compaction queue forever (#22039) (#22046)
- Fixed the issue that QueryNode may panic when stopped (#21406) (#21419)
- Modified lastSyncTime in advance to prevent multiple flush binlogs (#22048)
- Fixed the issue that a collection does not exist when users try to recover it (#21471) (#21628)
- Use tt msg stream for consume delete msg (#21478)
- Prevent users from deleting entities by any non-primary-key field (#21459) (#21472)
- Fixed potential nil access on segments (#22104)
v2.2.2
Release date: 22 December, 2022
| Milvus version | Python SDK version | Java SDK version | Go SDK version | Node.js SDK version |
|---|---|---|---|---|
| 2.2.2 | 2.2.1 | 2.2.1 | 2.2.0 | 2.2.1 |
Milvus 2.2.2 is a minor fix of Milvus 2.2.1. It fixed a few loading failure issues as of the upgrade to 2.2.1 and the issue that the proxy cache is not cleaned upon some types of errors.
Bug Fixes
- Fixed the issue that the proxy doesn't update the cache of shard leaders due to some types of errors. (#21320)
- Fixed the issue that the loaded info is not cleaned for released collections/partitions. (#21321)
- Fixed the issue that the load count is not cleared on time. (#21314)
v2.2.1
Release date: 15 December, 2022
| Milvus version | Python SDK version | Java SDK version | Go SDK version | Node.js SDK version |
|---|---|---|---|---|
| 2.2.1 | 2.2.1 | 2.2.1 | 2.2.0 | 2.2.1 |
Milvus 2.2.1 is a minor fix of Milvus 2.2.0. It supports authentication and TLS on all dependencies, optimizes the performance ludicrously on searches and fixes some critical issues. With tremendous contribution from the community, this release managed to resolve over 280 issues, so please try the new release and give us feedback on stability, performance and ease of use.
New Features
Performance
After upgrading the Knowhere vector engine and changing the parallelism strategy, Milvus 2.2.1 improves search performance by over 30%.
Optimizes the scheduler, and increases merge tasks probability. (#20931)
Bug Fixes
- Fixed term filtering failures on indexed scalar fields. (#20840)
- Fixed the issue that only partial data returned upon QueryNode restarts. (#21139)(#20976)
- Fixed IndexNode panic upon failures to create an index. (#20826)
- Fixed endless BinaryVector compaction and generation of data on Minio. (#21119) (#20971)
- Fixed the issue that
meta_cacheof proxy partially updates. (#21232) - Fixed slow segment loading due to staled checkpoints. (#21150)
- Fixed concurrently loaded Casbin model causing concurrent write operations. (#21132)(#21145)(#21073)
- Forbade garbage-collecting index meta when creating an index. (#21024)
- Fixed a bug that the index data can not be garbage-collected because
ListWithPrefixfrom Minio with recursive is false. (#21040) - Fixed an issue that an error code is returned when a query expression does not match any results. (#21066)
- Fixed search failures on disk index when
search_listequals tolimit. (#21114) - Filled collection schema after DataCoord restarts. (#21164)
- Fixed an issue that the compaction handler may double release and hang. (#21019)
- [restapi] Fixed precision loss for Int64 fields upon insert requests. (#20827)
- Increased
MaxWatchDurationand make it configurable to prevent shards with large data loads from timing out. (#21010) - Fixed the issue that the compaction target segment
rowNumis always 0. (#20941) - Fixed the issue that IndexCoord deletes segment index by mistake because IndexMeta is not stored in time. (#21058)
- Fixed the issue that DataCoord crushes if auto-compaction is disabled. (#21079)
- Fixed the issue that searches on growing segments even though the segments are indexed. (#21215)
Improvements
- Refined logs and the default log level is set to INFO.
- Fixed incorrect metrics and refined the metric dashboard.
- Made TopK limit configurable (#21155)
Breaking changes
Milvus now limits each RPC to 64 MB to avoid OOM and generating large message packs.
v2.2.0
Release date: 18 November, 2022
| Milvus version | Python SDK version | Java SDK version | Go SDK version | Node.js SDK version |
|---|---|---|---|---|
| 2.2.0 | 2.2.0 | 2.2.1 | 2.2.0 | 2.2.0 |
Milvus 2.2.0 introduces many new features including support for Disk-based approximate nearest neighbor (ANN) algorithm, bulk insertion of entities from files, and role-based access control (RBAC) for an improved security. In addition, this major release also ushers in a new era for vector search with enhanced stability, faster search speed, and more flexible scalability.
Breaking changes
Since metadata storage is refined and API usage is normalized, Milvus 2.2 is not fully compatible with earlier releases. Read this guide to learn how to safely upgrade from Milvus 2.1.x to 2.2.0.
Features
- Support for bulk insertion of entities from files
Milvus now offers a new set of bulk insertion APIs to make data insertion more efficient. You can now upload entities in a Json file directly to Milvus. See Insert Entities from Files for details.
- Query result pagination
To avoid massive search and query results returned in a single RPC, Milvus now supports configuring offset and filtering results with keywords in searches and queries. See Search and Query for details.
- Role-based access control (RBAC)
Like other traditional databases, Milvus now supports RBAC so that you can manages users, roles and privileges. See Enable RBAC for details.
- Quotas and limits
Quota is a new mechanism that protects the system from OOM and crash under a burst of traffic. By imposing quota limitations, you can limit ingestion rate, search rate, etc. See Quota and Limitation Configurations for details.
- Time to live (TTL) at a collection level
In prior releases, we only support configuring TTL at a cluster level. Milvus 2.2.0 now supports configuring collection TTL when you create or modify a collection. After setting TTL for a collection, the entities in this collection automatically expires after the specified period of time. See Create a collection or Modify a collection for details.
- Support for disk-based approximate nearest neighbor search (ANNS) indexes (Beta)
Traditionally, you need to load the entire index into memory before search. Now with DiskANN, an SSD-resident and Vamana graph-based ANNS algorithm, you can directly search on large-scale datasets and save up to 10 times the memory.
- Data backup (Beta)
Thanks to the contribution from Zilliz, Milvus 2.2.0 now provides a tool to back up and restore data. The tool can be used either in a command line or an API server for data security.
Bug fixes and stability
- Implements query coord V2, which handles all channel/segment allocation in a fully event-driven and asynchronous mode. Query coord V2 address all issues of stuck searches and accelerates failure recovery.
- Root coord and index coord are refactored for more elegant handling of errors and better task scheduling.
- Fixes the issue of invalid RocksMQ retention mechanism when Milvus Standalone restarts.
- Meta storage format in etcd is refactored. With the new compression mechanism, etcd kv size is reduced by 10 times and the issues of etcd memory and space are solved.
- Fixes a couple of memory issues when entities are continuously inserted or deleted.
Improvements
-
Performance
- Fixes performance bottleneck to that Milvus can fully utilize all cores when CPU is more than 8 cores.
- Dramatically improves the search throughput and reduce the latency.
- Decreases load speed by processing load in parallel.
-
Observability
- Changes all log levels to
infoby default. - Added collection-level latency metrics for search, query, insertion, and deletion.
- Changes all log levels to
-
Debug tool
- BirdWatcher, the debug tool for Milvus, is further optimized as it can now connect to Milvus meta storage and inspect the part of the internal status of the Milvus system.
Others
- Index and load
- A collection can only be loaded with an index created on it.
- Indexes cannot be created after a collection is loaded.
- A loaded collection must be released before dropping the index created on this collection.
- Flush
- Flush API, which forces a seal on a growing segment and syncs the segment to object storage, is now exposed to users. Calling
flush()frequently may affect search performance as too many small segments are created. - No auto-flush is triggered by any SDK APIs such as
num_entities(),create_index(), etc.
- Flush API, which forces a seal on a growing segment and syncs the segment to object storage, is now exposed to users. Calling
- Time Travel
- In Milvus 2.2, Time Travel is disabled by default to save disk usage. To enable Time Travel, configure the parameter
common.retentionDurationmanually.
- In Milvus 2.2, Time Travel is disabled by default to save disk usage. To enable Time Travel, configure the parameter