milvus-2.0.0-rc8
Pre-releasev2.0.0-RC8
Release date: 2021-11-5
Compatibility
Milvus version | Python SDK version | Java SDK version | Go SDK version | Node SDK version |
---|---|---|---|---|
2.0.0-RC8 | 2.0.0rc8 | Coming soon | Coming soon | 1.0.18 |
Milvus 2.0.0-RC8 is the last release candidate of Milvus 2.0.0-GA. It supports handoff task, primary key deduplication and search by Time Travel functionalities. The mean time to recovery (MTTR) has also been greatly reduced with the enhancement of timetick mechanism. We had run stress test on 2.0.0-RC8 with 10M datasets, and both standalone and distributed cluster survived for 84 hours.
Improvements
-
Failure Recovery speed:
- #10737 Fixes Session checker for proxy.
- #10723 Fixes seek query channel error.
- #10907 Fixes
LatestPosition
option conflict with earliest patch. - #10616 Removes Common YAML.
- #10771 Changes
SeekPosition
to the earliest of all segments. - #10651 Fixes query coord set seek position error.
- #9543 Initializes global sealed segments and seek query channel when
AddQueryChannel
. - #9684 Skips re-consuming timetick MsgStream when data coord restarts.
-
Refactor meta snapshot:
-
#10563 Changes default balance policy.
-
#10730 Returns segment state when getting query segment information.
-
#10534 Supports reading MinIO configuration from environment variables.
-
#10114 Sets default
gracefulTime
to0
. -
#9860 Hides
liveChn
intosessionutil
and fix liveness initialization order. -
#7115 Uses etcd to watch channel on data node.
-
#7606 Makes
knowhere
compile independently.
Features
-
Handoff:
-
#10330 Adds
handoffTask
. -
#10084 Broadcasts
sealedSegmentChangeInfo
toqueryChannel
. -
#10619 Fixes removing segment when query node receives
segmentChangeInfo
. -
#10045 Watches
changeInfo
in query node. -
#10011 Updates excluded segments info when receiving
changeInfo
. -
#9606 Adds initialization information for
AddQueryChannelRequest
. -
#10619 Fixes removing segment when query node receives
segmentChangeInfo
.
-
-
Primary Deduplication:
- #10834 Removes primary key duplicated query result in query node.
- #10355 Removes duplicated search results in proxy.
- #10117 Removes duplicated search results in segcore reduce.
- #10949 Uses primary key only to check search result duplication.
- #10967 Removes primary key duplicated query result in proxy.
-
Auto-flush:
- #10659 Adds
injectFlush
method forflushManager
interface. - #10580 Adds injection logic for
FlushManager
. - #10550 Merges automatic and manual flush with same segment ID.
- #10539 Allows flushed segments to trigger flush process.
- #10197 Adds a timed flush trigger mechanism.
- #10142 Applies flush manager logic in data node.
- #10075 Uses single signal channel to notify flush.
- #9986 Adds flush manager structure.
- #10659 Adds
-
#10173 Adds binlog iterators.
-
#10193 Changes bloom filter use primary key.
-
#9782 Adds
allocIDBatch
for data node allocator.
Bug Fixes
-
Incorrect collection loading behavior if there is not enough memory:
- #10796 Fixes get container mem usage.
- #10800 Uses
TotalInactiveFile
inGetContainerMemUsed
. - #10603 Increases compatibility for
EstimateMemorySize
interface. - #10363 Adds
cgroups
to get container memory and check index memory in segment loader. - #10294 Uses proto size to calculate request size.
- #9688 Estimates memory size with descriptor event.
- #9681 Fixes the way that binlog stores the original memory size.
- #9628 Stores original memory size of binlog file to extra information.
-
Size of etcd-related request is too large:
- #10909 Fixes too many operations in
txn
request when savingsegmentInfo
. - #10812 Fixes too large request when loading segment.
- #10768 Fixes too large request when loading collection.
- #10655 Splits watch operations into many transactions.
- #10587 Compacts
multiSegmentChangeInfo
to a single info. - #10425 Trims
segmentinfo
binlog forVChaninfo
usage. - #10340 Fixes
multiSave
childTask
failed to etcd. - #10310 Fixes error when assigning load segment request.
- #10125 Splits large
loadSegmentReq
to multiple small requests.
- #10909 Fixes too many operations in
-
System panics:
-
RocksMQ-related issues:
- #10367 Stops retention gracefully.
- #9828 Fixes retention data race.
- #9933 Changes retention ticker time to 10 minutes.
- #9694 Deletes messages before deleting metadata in rocksmq retention.
- #11029 Fixes rocksmq
SeekToLatest
. - #11057 Fixes
SeekToLatest
memory leakage and remove redundant logic. - #11081 Fixes rocksdb retention ts not set.
- #11083 Adds topic lock for rocksmq
Seek
. - #11076 Moves topic lock to the front of final delete in retention expired cleanup.
-
#10751
loadIndex
keep retrying whenindexFilePathInfo
gets empty list. -
#10583
ParseHybridTs
returns type to INT64. -
#10599 Delete message hash error.
-
#10314 Index building task mistakenly canceled by index coord by mistake.
-
#9701 Incorrect
CreateAlias/DropAlias/AlterAlias
implementation. -
#9573 Timeout when data coord saves binlog.
-
#9788 Watch Channel canceled due to revision compacted.
-
#10994 Index node does not balances load.
-
#11152 Search is wrong when using Time Travel without filtering condition and call
num_entities
. -
#11222 Incorrect empty retrieve result handling.