milvus/configs/milvus.yaml

# Licensed to the LF AI & Data foundation under one
# or more contributor license agreements. See the NOTICE file
# distributed with this work for additional information
# regarding copyright ownership. The ASF licenses this file
# to you under the Apache License, Version 2.0 (the
# "License"); you may not use this file except in compliance
# with the License. You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# Related configuration of etcd, used to store Milvus metadata & service discovery.
etcd:
  endpoints: localhost:2379
  rootPath: by-dev # The root path where data is stored in etcd
  metaSubPath: meta # metaRootPath = rootPath + '/' + metaSubPath
  kvSubPath: kv # kvRootPath = rootPath + '/' + kvSubPath
  log:
    level: info # Only supports debug, info, warn, error, panic, or fatal. Default 'info'.
    # path is one of:
    #  - "default" as os.Stderr,
    #  - "stderr" as os.Stderr,
    #  - "stdout" as os.Stdout,
    #  - file path to append server logs to.
    # please adjust in embedded Milvus: /tmp/milvus/logs/etcd.log
    path: stdout
  ssl:
    enabled: false # Whether to support ETCD secure connection mode
    tlsCert: /path/to/etcd-client.pem # path to your cert file
    tlsKey: /path/to/etcd-client-key.pem # path to your key file
    tlsCACert: /path/to/ca.pem # path to your CACert file
    # TLS min version
    # Optional values: 1.0, 1.1, 1.2, 1.3。
    # We recommend using version 1.2 and above.
    tlsMinVersion: 1.3
  use:
    embed: false # Whether to enable embedded Etcd (an in-process EtcdServer).
  data:
    dir: default.etcd # Embedded Etcd only. please adjust in embedded Milvus: /tmp/milvus/etcdData/

metastore:
  # Default value: etcd
  # Valid values: [etcd, mysql]
  type: etcd

# Related configuration of mysql, used to store Milvus metadata.
mysql:
  username: root
  password: 123456
  address: localhost
  port: 3306
  dbName: milvus_meta
  maxOpenConns: 20
  maxIdleConns: 5

localStorage:
  path: /var/lib/milvus/data/ # please adjust in embedded Milvus: /tmp/milvus/data/

# Related configuration of MinIO/S3/GCS or any other service supports S3 API, which is responsible for data persistence for Milvus.
# We refer to the storage service as MinIO/S3 in the following description for simplicity.
minio:
  address: localhost # Address of MinIO/S3
  port: 9000 # Port of MinIO/S3
  accessKeyID: minioadmin # accessKeyID of MinIO/S3
  secretAccessKey: minioadmin # MinIO/S3 encryption string
  useSSL: false # Access to MinIO/S3 with SSL
  bucketName: a-bucket # Bucket name in MinIO/S3
  rootPath: files # The root path where the message is stored in MinIO/S3
  # Whether to useIAM role to access S3/GCS instead of access/secret keys
  # For more information, refer to
  # aws: https://docs.aws.amazon.com/IAM/latest/UserGuide/id_roles_use.html
  # gcp: https://cloud.google.com/storage/docs/access-control/iam
  # aliyun (ack): https://www.alibabacloud.com/help/en/container-service-for-kubernetes/latest/use-rrsa-to-enforce-access-control 
  # aliyun (ecs): https://www.alibabacloud.com/help/en/elastic-compute-service/latest/attach-an-instance-ram-role
  useIAM: false
  # Cloud Provider of S3. Supports: "aws", "gcp", "aliyun". 
  # You can use "aws" for other cloud provider supports S3 API with signature v4, e.g.: minio
  # You can use "gcp" for other cloud provider supports S3 API with signature v2
  # You can use "aliyun" for other cloud provider uses virtual host style bucket 
  # When useIAM enabled, only "aws", "gcp", "aliyun" is supported for now
  cloudProvider: aws
  # Custom endpoint for fetch IAM role credentials. when useIAM is true & cloudProvider is "aws".
  # Leave it empty if you want to use AWS default endpoint
  iamEndpoint: 

# Milvus supports three MQ: rocksmq(based on RockDB), Pulsar and Kafka, which should be reserved in config what you use.
# There is a note about enabling priority if we config multiple mq in this file
# 1. standalone(local) mode: rocksmq(default) > Pulsar > Kafka
# 2. cluster mode:  Pulsar(default) > Kafka (rocksmq is unsupported)

# Related configuration of pulsar, used to manage Milvus logs of recent mutation operations, output streaming log, and provide log publish-subscribe services.
pulsar:
  address: localhost # Address of pulsar
  port: 6650 # Port of Pulsar
  webport: 80 # Web port of pulsar, if you connect direcly without proxy, should use 8080
  maxMessageSize: 5242880 # 5 * 1024 * 1024 Bytes, Maximum size of each message in pulsar.
  tenant: public
  namespace: default

# If you want to enable kafka, needs to comment the pulsar configs
# kafka:
#   brokerList: 
#   saslUsername: 
#   saslPassword: 
#   saslMechanisms: PLAIN
#   securityProtocol: SASL_SSL

rocksmq:
  # The path where the message is stored in rocksmq
  # please adjust in embedded Milvus: /tmp/milvus/rdb_data
  path: /var/lib/milvus/rdb_data
  lrucacheratio: 0.06 # rocksdb cache memory ratio
  rocksmqPageSize: 268435456 # 256 MB, 256 * 1024 * 1024 bytes, The size of each page of messages in rocksmq
  retentionTimeInMinutes: 7200 # 5 days, 5 * 24 * 60 minutes, The retention time of the message in rocksmq.
  retentionSizeInMB: 8192 # 8 GB, 8 * 1024 MB, The retention size of the message in rocksmq.
  compactionInterval: 86400 # 1 day, trigger rocksdb compaction every day to remove deleted data

# Related configuration of rootCoord, used to handle data definition language (DDL) and data control language (DCL) requests
rootCoord:
  dmlChannelNum: 256 # The number of dml channels created at system startup
  maxPartitionNum: 4096 # Maximum number of partitions in a collection
  minSegmentSizeToEnableIndex: 1024 # It's a threshold. When the segment size is less than this value, the segment will not be indexed
  importTaskExpiration: 900 # (in seconds) Duration after which an import task will expire (be killed). Default 900 seconds (15 minutes).
  importTaskRetention: 86400 # (in seconds) Milvus will keep the record of import tasks for at least `importTaskRetention` seconds. Default 86400, seconds (24 hours).
  enableActiveStandby: false
  port: 53100
  grpc:
    serverMaxSendSize: 536870912
    serverMaxRecvSize: 536870912
    clientMaxSendSize: 268435456
    clientMaxRecvSize: 268435456

# Related configuration of proxy, used to validate client requests and reduce the returned results.
proxy:
  timeTickInterval: 200 # ms, the interval that proxy synchronize the time tick
  msgStream:
    timeTick:
      bufSize: 512
  maxNameLength: 255 # Maximum length of name for a collection or alias
  # Maximum number of fields in a collection.
  # As of today (2.2.0 and after) it is strongly DISCOURAGED to set maxFieldNum >= 64.
  # So adjust at your risk!
  maxFieldNum: 64
  maxShardNum: 64 # Maximum number of shards in a collection
  maxDimension: 32768 # Maximum dimension of a vector
  # Whether to produce gin logs.\n
  # please adjust in embedded Milvus: false
  ginLogging: true
  maxTaskNum: 1024 # max task number of proxy task queue
  accessLog:
    localPath: /tmp/milvus_accesslog
    filename: milvus_access_log.log # Log filename, leave empty to disable file log.
  http:
    enabled: true # Whether to enable the http server
    debug_mode: false # Whether to enable http server debug mode
  port: 19530
  internalPort: 19529
  grpc:
    serverMaxSendSize: 67108864
    serverMaxRecvSize: 67108864
    clientMaxSendSize: 268435456
    clientMaxRecvSize: 268435456

# Related configuration of queryCoord, used to manage topology and load balancing for the query nodes, and handoff from growing segments to sealed segments.
queryCoord:
  taskMergeCap: 16
  taskExecutionCap: 256
  autoHandoff: true # Enable auto handoff
  autoBalance: true # Enable auto balance
  overloadedMemoryThresholdPercentage: 90 # The threshold percentage that memory overload
  balanceIntervalSeconds: 60
  memoryUsageMaxDifferencePercentage: 30
  checkInterval: 1000
  channelTaskTimeout: 60000 # 1 minute
  segmentTaskTimeout: 120000 # 2 minute
  distPullInterval: 500
  heartbeatAvailableInterval: 10000 # 10s, Only QueryNodes which fetched heartbeats within the duration are available
  loadTimeoutSeconds: 600
  checkHandoffInterval: 5000
  enableActiveStandby: false
  port: 19531
  grpc:
    serverMaxSendSize: 536870912
    serverMaxRecvSize: 536870912
    clientMaxSendSize: 268435456
    clientMaxRecvSize: 268435456

# Related configuration of queryNode, used to run hybrid search between vector and scalar data.
queryNode:
  dataSync:
    flowGraph:
      maxQueueLength: 1024 # Maximum length of task queue in flowgraph
      maxParallelism: 1024 # Maximum number of tasks executed in parallel in the flowgraph
  stats:
    publishInterval: 1000 # Interval for querynode to report node information (milliseconds)
  segcore:
    chunkRows: 1024 # The number of vectors in a chunk.
    smallIndex:
      nlist: 128 # small index nlist, recommend to set sqrt(chunkRows), must smaller than chunkRows/8
      nprobe: 16 # nprobe to search small index, based on your accuracy requirement, must smaller than nlist
  loadMemoryUsageFactor: 3 # The multiply factor of calculating the memory usage while loading segments
  enableDisk: true # enable querynode load disk index, and search on disk index
  maxDiskUsagePercentage: 95
  cache:
    enabled: true
    memoryLimit: 2147483648 # 2 GB, 2 * 1024 *1024 *1024
  grouping:
    enabled: true
    maxNQ: 1000
    topKMergeRatio: 10
  scheduler:
    receiveChanSize: 10240
    unsolvedQueueSize: 10240
    # maxReadConcurrentRatio is the concurrency ratio of read task (search task and query task).
    # Max read concurrency would be the value of runtime.NumCPU * maxReadConcurrentRatio.
    # It defaults to 2.0, which means max read concurrency would be the value of runtime.NumCPU * 2.
    # Max read concurrency must greater than or equal to 1, and less than or equal to runtime.NumCPU * 100.
    # (0, 100]
    maxReadConcurrentRatio: 2
    cpuRatio: 10 # ratio used to estimate read task cpu usage.
    maxTimestampLag: 86400
  gracefulStopTimeout: 30
  port: 21123
  grpc:
    serverMaxSendSize: 536870912
    serverMaxRecvSize: 536870912
    clientMaxSendSize: 268435456
    clientMaxRecvSize: 268435456

indexCoord:
  bindIndexNodeMode:
    enable: false
    address: localhost:22930
    withCred: false
    nodeID: 0
  segment:
    minSegmentNumRowsToEnableIndex: 1024 # It's a threshold. When the segment num rows is less than this value, the segment will not be indexed

indexNode:
  scheduler:
    buildParallel: 1
  enableDisk: true # enable index node build disk vector index
  maxDiskUsagePercentage: 95
  gracefulStopTimeout: 30
  port: 21121
  grpc:
    serverMaxSendSize: 536870912
    serverMaxRecvSize: 536870912
    clientMaxSendSize: 268435456
    clientMaxRecvSize: 268435456

dataCoord:
  channel:
    watchTimeoutInterval: 30 # Timeout on watching channels (in seconds). Datanode tickler update watch progress will reset timeout timer.
    balanceSilentDuration: 300 # The duration before the channelBalancer on datacoord to run
    balanceInterval: 360 #The interval for the channelBalancer on datacoord to check balance status
  segment:
    maxSize: 512 # Maximum size of a segment in MB
    diskSegmentMaxSize: 2048 # Maximun size of a segment in MB for collection which has Disk index
    sealProportion: 0.23
    assignmentExpiration: 2000 # The time of the assignment expiration in ms
    maxLife: 86400 # The max lifetime of segment in seconds, 24*60*60
    # If a segment didn't accept dml records in maxIdleTime and the size of segment is greater than
    # minSizeFromIdleToSealed, Milvus will automatically seal it.
    # The max idle time of segment in seconds, 10*60.
    maxIdleTime: 600
    minSizeFromIdleToSealed: 16 # The min size in MB of segment which can be idle from sealed.
    # The max number of binlog file for one segment, the segment will be sealed if
    # the number of binlog file reaches to max value.
    maxBinlogFileNumber: 32
    smallProportion: 0.5 # The segment is considered as "small segment" when its # of rows is smaller than
    # (smallProportion * segment max # of rows).
    # A compaction will happen on small segments if the segment after compaction will have
    compactableProportion: 0.5
    # over (compactableProportion * segment max # of rows) rows.
    # MUST BE GREATER THAN OR EQUAL TO <smallProportion>!!!
    # During compaction, the size of segment # of rows is able to exceed segment max # of rows by (expansionRate-1) * 100%. 
    expansionRate: 1.25
  enableCompaction: true # Enable data segment compaction
  compaction:
    enableAutoCompaction: true
  enableGarbageCollection: true
  gc:
    interval: 3600 # gc interval in seconds
    missingTolerance: 86400 # file meta missing tolerance duration in seconds, 60*24
    dropTolerance: 3600 # file belongs to dropped entity tolerance duration in seconds. 3600
  enableActiveStandby: false
  port: 13333
  grpc:
    serverMaxSendSize: 536870912
    serverMaxRecvSize: 536870912
    clientMaxSendSize: 268435456
    clientMaxRecvSize: 268435456

dataNode:
  dataSync:
    flowGraph:
      maxQueueLength: 1024 # Maximum length of task queue in flowgraph
      maxParallelism: 1024 # Maximum number of tasks executed in parallel in the flowgraph
  segment:
    insertBufSize: 16777216 # Max buffer size to flush for a single segment.
    deleteBufBytes: 67108864 # Max buffer size to flush del for a single channel
    syncPeriod: 600 # The period to sync segments if buffer is not empty.
  port: 21124
  grpc:
    serverMaxSendSize: 536870912
    serverMaxRecvSize: 536870912
    clientMaxSendSize: 268435456
    clientMaxRecvSize: 268435456
  memory:
    forceSyncEnable: true # `true` to force sync if memory usage is too high
    forceSyncThreshold: 0.6 # forceSync only take effects when memory usage ratio > forceSyncThreshold. note that should set it a small value like 0.2 for standalone deployment 
    forceSyncSegmentRatio: 0.3 # ratio of segments to sync, top largest forceSyncSegmentRatio segments will be synced

# Configures the system log output.
log:
  level: debug # Only supports debug, info, warn, error, panic, or fatal. Default 'info'.
  file:
    rootPath:  # root dir path to put logs, default "" means no log file will print. please adjust in embedded Milvus: /tmp/milvus/logs
    maxSize: 300 # MB
    maxAge: 10 # Maximum time for log retention in day.
    maxBackups: 20
  format: text # text or json
  stdout: true # Stdout enable or not

grpc:
  log:
    level: WARNING
  serverMaxSendSize: 536870912
  serverMaxRecvSize: 536870912
  client:
    compressionEnabled: false
    dialTimeout: 5000
    keepAliveTime: 10000
    keepAliveTimeout: 20000
    maxMaxAttempts: 5
    initialBackoff: 1
    maxBackoff: 60
    backoffMultiplier: 2
  clientMaxSendSize: 268435456
  clientMaxRecvSize: 268435456

# Configure the proxy tls enable.
tls:
  serverPemPath: configs/cert/server.pem
  serverKeyPath: configs/cert/server.key
  caPemPath: configs/cert/ca.pem

common:
  chanNamePrefix:
    cluster: by-dev
    rootCoordTimeTick: rootcoord-timetick
    rootCoordStatistics: rootcoord-statistics
    rootCoordDml: rootcoord-dml
    rootCoordDelta: rootcoord-delta
    search: search
    searchResult: searchResult
    queryTimeTick: queryTimeTick
    dataCoordStatistic: datacoord-statistics-channel
    dataCoordTimeTick: datacoord-timetick-channel
    dataCoordSegmentInfo: segment-info-channel
  subNamePrefix:
    proxySubNamePrefix: proxy
    rootCoordSubNamePrefix: rootCoord
    queryNodeSubNamePrefix: queryNode
    dataCoordSubNamePrefix: dataCoord
    dataNodeSubNamePrefix: dataNode
  defaultPartitionName: _default # default partition name for a collection
  defaultIndexName: _default_idx # default index name
  retentionDuration: 0 # time travel reserved time, insert/delete will not be cleaned in this period. disable it by default
  entityExpiration: -1 # Entity expiration in seconds, CAUTION make sure entityExpiration >= retentionDuration and -1 means never expire
  indexSliceSize: 16 # MB
  threadCoreCoefficient: 10 # This parameter specify how many times the number of threads is the number of cores
  DiskIndex:
    MaxDegree: 56
    SearchListSize: 100
    PQCodeBudgetGBRatio: 0.125
    BuildNumThreadsRatio: 1
    SearchCacheBudgetGBRatio: 0.1
    LoadNumThreadRatio: 8
    BeamWidthRatio: 4
  gracefulTime: 5000 # milliseconds. it represents the interval (in ms) by which the request arrival time needs to be subtracted in the case of Bounded Consistency.
  gracefulStopTimeout: 30 # seconds. it will force quit the server if the graceful stop process is not completed during this time.
  storageType: minio # please adjust in embedded Milvus: local
  # Default value: auto
  # Valid values: [auto, avx512, avx2, avx, sse4_2]
  # This configuration is only used by querynode and indexnode, it selects CPU instruction set for Searching and Index-building.
  simdType: auto
  security:
    authorizationEnabled: false
    # The superusers will ignore some system check processes,
    # like the old password verification when updating the credential
    superUsers: root
    tlsMode: 0
  session:
    ttl: 20 # ttl value when session granting a lease to register service
    retryTimes: 30 # retry times when session sending etcd requests

# QuotaConfig, configurations of Milvus quota and limits.
# By default, we enable:
#   1. TT protection;
#   2. Memory protection.
#   3. Disk quota protection.
# You can enable:
#   1. DML throughput limitation;
#   2. DDL, DQL qps/rps limitation;
#   3. DQL Queue length/latency protection;
#   4. DQL result rate protection;
# If necessary, you can also manually force to deny RW requests.
quotaAndLimits:
  enabled: true # `true` to enable quota and limits, `false` to disable.
  # quotaCenterCollectInterval is the time interval that quotaCenter
  # collects metrics from Proxies, Query cluster and Data cluster.
  # seconds, (0 ~ 65536)
  quotaCenterCollectInterval: 3
  ddl:
    enabled: false
    collectionRate: -1 # qps, default no limit, rate for CreateCollection, DropCollection, LoadCollection, ReleaseCollection
    partitionRate: -1 # qps, default no limit, rate for CreatePartition, DropPartition, LoadPartition, ReleasePartition
  indexRate:
    enabled: false
    max: -1 # qps, default no limit, rate for CreateIndex, DropIndex
  flushRate:
    enabled: false
    max: -1 # qps, default no limit, rate for flush
  compactionRate:
    enabled: false
    max: -1 # qps, default no limit, rate for manualCompaction
  dml:
    # dml limit rates, default no limit.
    # The maximum rate will not be greater than max.
    enabled: false
    insertRate:
      max: -1 # MB/s, default no limit
    deleteRate:
      max: -1 # MB/s, default no limit
    bulkLoadRate:
      max: -1 # MB/s, default no limit, not support yet. TODO: limit bulkLoad rate
  dql:
    # dql limit rates, default no limit.
    # The maximum rate will not be greater than max.
    enabled: false
    searchRate:
      max: -1 # vps (vectors per second), default no limit
    queryRate:
      max: -1 # qps, default no limit
  limitWriting:
    # forceDeny false means dml requests are allowed (except for some
    # specific conditions, such as memory of nodes to water marker), true means always reject all dml requests.
    forceDeny: false
    ttProtection:
      enabled: false
      # maxTimeTickDelay indicates the backpressure for DML Operations.
      # DML rates would be reduced according to the ratio of time tick delay to maxTimeTickDelay,
      # if time tick delay is greater than maxTimeTickDelay, all DML requests would be rejected.
      # seconds
      maxTimeTickDelay: 300
    memProtection:
      # When memory usage > memoryHighWaterLevel, all dml requests would be rejected;
      # When memoryLowWaterLevel < memory usage < memoryHighWaterLevel, reduce the dml rate;
      # When memory usage < memoryLowWaterLevel, no action.
      enabled: true
      dataNodeMemoryLowWaterLevel: 0.85 # (0, 1], memoryLowWaterLevel in DataNodes
      dataNodeMemoryHighWaterLevel: 0.95 # (0, 1], memoryHighWaterLevel in DataNodes
      queryNodeMemoryLowWaterLevel: 0.85 # (0, 1], memoryLowWaterLevel in QueryNodes
      queryNodeMemoryHighWaterLevel: 0.95 # (0, 1], memoryHighWaterLevel in QueryNodes
    diskProtection:
      enabled: true # When the total file size of object storage is greater than `diskQuota`, all dml requests would be rejected;
      diskQuota: -1 # MB, (0, +inf), default no limit
  limitReading:
    # forceDeny false means dql requests are allowed (except for some
    # specific conditions, such as collection has been dropped), true means always reject all dql requests.
    forceDeny: false
    queueProtection:
      enabled: false
      # nqInQueueThreshold indicated that the system was under backpressure for Search/Query path.
      # If NQ in any QueryNode's queue is greater than nqInQueueThreshold, search&query rates would gradually cool off
      # until the NQ in queue no longer exceeds nqInQueueThreshold. We think of the NQ of query request as 1.
      # int, default no limit
      nqInQueueThreshold: -1
      # queueLatencyThreshold indicated that the system was under backpressure for Search/Query path.
      # If dql latency of queuing is greater than queueLatencyThreshold, search&query rates would gradually cool off
      # until the latency of queuing no longer exceeds queueLatencyThreshold.
      # The latency here refers to the averaged latency over a period of time.
      # milliseconds, default no limit
      queueLatencyThreshold: -1
    resultProtection:
      enabled: false
      # maxReadResultRate indicated that the system was under backpressure for Search/Query path.
      # If dql result rate is greater than maxReadResultRate, search&query rates would gradually cool off
      # until the read result rate no longer exceeds maxReadResultRate.
      # MB/s, default no limit
      maxReadResultRate: -1
    # colOffSpeed is the speed of search&query rates cool off.
    # (0, 1]
    coolOffSpeed: 0.9

trace:
  # trace exporter type, default is stdout,
  # optional values: ['stdout', 'jaeger']
  exporter: stdout
  # fraction of traceID based sampler,
  # optional values: [0, 1]
  # Fractions >= 1 will always sample. Fractions < 0 are treated as zero.
  sampleFraction: 0
  jaeger:
    url:  # when exporter is jaeger should set the jaeger's URL