book

The Architecture Handbook for Milvus Vector Database

Name: The Architecture Handbook for Milvus Vector Database
ISBN: 9781835881705

by Yudong Cai, Jeremy Zhu, Xuan Yang, Bang Fu

March 2026

Intermediate

502 pages

13h

English

Packt Publishing

Read now

Unlock full access

Preface
Free benefits with your book
Part 1: Getting Started with Milvus
Chapter 1: Introduction to Milvus
Technical requirementsUnderstanding a vector databaseTypes of databasesUnstructured dataExploring the encoding processOne-hot encodingTF-IDF encodingSimilarity search algorithmsLSHIntroducing MilvusSummary
Chapter 2: Deploying Milvus in Multiple Ways
Technical requirementsUnderstanding common service deployment methodsIntroducing Docker and Docker ComposeIntroducing Docker ComposeIntroducing the Kubernetes operatorIntroducing Helm chartsComparing Helm charts and Kubernetes operatorsDeploying MilvusDeploying Milvus with Docker ComposeDeploying Milvus through KubernetesUsing operators to deploy MilvusUsing Helm charts to deploy MilvusSummary
Chapter 3: Interacting with Milvus
Technical requirementsIntroducing the core objects of MilvusFieldsSchemasCollectionIndexInteracting with Milvus via the Python SDKStep 1: Connecting to MilvusStep 2: Creating a collectionStep 3: Viewing collectionsStep 4: Inserting dataStep 5: Querying the data with an ID of 2Step 6: Upserting the data with an ID of 2Step 7: Running a query with filteringStep 8: Deleting an entityStep 9: Carrying out vector searchStep 10: Deleting the collectionInteracting with Milvus via REST APIsStep 1: Creating a collectionStep 2: Describing the collectionStep 3: Listing all collectionsStep 4: Checking the index information and load status of the collectionStep 5: Carrying out a data query or searchStep 6: Dropping the collectionSummary
Chapter 4: Configuring the Milvus System
Technical requirementsUnderstanding dependency services and startup configurations for MilvusManaging core system and service configurations in MilvusAdjusting configurations before starting MilvusModifying configurationMonitoring access and system logsDeploying Prometheus and GrafanaDeploying Loki and PromtailSummary
Part 2: System Architecture and Data Lifecycle Management
Chapter 5: Understanding the Milvus Data Model and Architecture
Technical requirementsMilvus data modelCollection, entity, and schemaEntity compositionPartitionPartition APIsPartition key fieldShardsTypical number of shardsMilvus' layered architectureAccess layerCoordination layerWorker layerStorage layerMQ – the lifeline of streaming dataGlobal event orderingTimestamp in MilvusTimetick mechanismSummary
Chapter 6: Data Modification and Maintenance in Milvus
Technical requirementsThe core of data organization – the segmentData states – streaming vs historicalInternal structure – the file composition of a segmentThe lifecycle of a segmentPhase one – the growing statePhase two – the sealed statePhase three – the flushed statePhase four – the dropped stateThe journey behind DML requestsInsert requestSynchronous processing (during the API call)Asynchronous processing (after the API returns)Visibility and persistence pointsDelete requestSynchronous processing (during the API call)Asynchronous processing (after the API returns)Upsert requestChannel checkpointsSummary
Chapter 7: Reading Data in Milvus
The query scatter-gather patternQuery participantsThe scatter phase – distributing the workloadThe gather phaseLoading and balancingReplicasLoading and releasing replicasDynamic load balancingHandling deletesThe delegator's role and Bloom filtersEnsuring multi-level consistencyHow Milvus ensures consistencySummary

Chapter 8: Compaction and Garbage Collection
Technical requirementsWhy compaction matters in Milvus?Segment fragmentationHow fragmentation impacts query performanceMix compaction's performance optimizationClustering compaction enables intra-partition segment pruningStorage and performance overhead of soft deletionsL0 and mix compaction address soft delete overheadCompaction in detailThe common processTaking a closer look at the scheduling processMix compactionL0 compactionClustering compactionTriggersManaging obsolete data with garbage collectionGC working principlesSafety first – multi-layer protection ensuring data integrityPerformance priority – intelligent scheduling and query protectionFine-tuning GC behaviorSummary
Part 3: Vector Computing Engine, Indexing, and Advanced Query Mechanisms
Chapter 9: Exploring Milvus' Vector Engine
Technical requirementsIntroducing Milvus' core engine: KnowhereUnderstanding the functional scope of KnowhereMilvus data modelKnowhere architectureWorking with Knowhere's main APIsKnowhere parametersTrain(), Add(), and Build()Serialize() and Deserialize()Search() and RangeSearch()HasRawData() and GetVectorByIds()Understanding Knowhere's functionalityThread managementVersion managementPrometheus monitorSummary
Chapter 10: How to Select a Vector Index
Introducing performance metricsRecall and accuracyVector per second and latencyUnderstanding vector types and metric typesFloatVector seriesBinaryVectorSparseFloatVectorTheory behind indexesMeasuring the performance of indexesFLAT index and BruteForce seriesIVF series indexesSCANNHNSW series indexesDiskANNRAFT series indexesSPARSE series indexesRecommendations for index selectionSummary
Chapter 11: Handling Complicated Search Requests
Handling attribute filteringBoolean or logical expressionASTPlan AST generationAST executionUsing bitsets to soft delete vectorsWhat's a bitset?How to calculate a bitsetUsing SIMD to accelerate distance calculationsSIMD APIs' dynamic hookingConnecting the dots – an overview of Milvus query executionSummary
Part 4: System Evaluation, Tuning, and Ecosystem Application
Chapter 12: Getting Started with Milvus Performance Benchmarking
Technical requirementsUnderstanding ANN search and Milvus performance metricsWhat is ANN search?How ANN search works in MilvusWhy Milvus performance mattersCommon performance challenges in vector searchMeasuring Milvus performanceFactors influencing Milvus' performanceBenchmarking tools overviewStandardized benchmarks vs.real-world workloadsCommon benchmarking toolsRunning your benchmark with VectorDBBench and LocustUsing Locust for custom performance testingSummary
Chapter 13: Stability and Reliability Evaluation for the Milvus Vector Database
Technical requirementsUnderstanding Milvus stability and reliabilityUnderstanding the relationship between stability and reliabilitySetting up the Milvus reliability testing environmentConfiguring Milvus for high-availability testingKey Grafana monitoring panels for reliability testingHow to interpret Grafana metrics for cluster healthLoad testing for Milvus reliabilityDesigning load scenariosExample scenario: e-commerce product recommendation systemExecuting load testsStep 1: preparing test dataStep 2: writing load scripts with LocustStep 3: writing your own custom load shapeAnalyzing the results and deriving action itemsLong-term running (soak) testMethods and goalsE-commerce soak test scenarioComplex workload testing simulating real-world scenariosSimulating Milvus component failures and verifying recoveryPrerequisites: installing Chaos MeshSimulating active MixCoord failureTest stepsVerification pointsSimulating worker node failureUnderstanding the Milvus WAL mechanismTest stepsTesting WAL data durabilityVerification points for QueryNode failureExample test outputSummary
Chapter 14: Scalability Evaluation for the Milvus Vector Database
Technical requirementsScaling strategies in MilvusVertical scaling (scale-up)Horizontal scaling (scale-out)Worker nodes and their scaling benefitsMaking the right scaling decisionIntroduction to scalability testing methodologyDefine clear test objectivesPrepare test environment and dataDataset requirementsDesign basic test workloadsFocusing on key performance indicatorsUtilizing basic testing and monitoring toolsHands-on lab: vertical scaling with testing methodologiesHands-on lab: horizontal scaling (performance improvement type)Scenario – single in-memory replica test (verify basic horizontal scaling effects)Scenario – multiple in-memory replicas testHands-on lab: capacity improvement testing with methodologyScenario – verifying the cluster's ability to linearly handle larger datasets after scalingUsing scaling to resolve Milvus performance bottlenecksMonitoring scaling effectivenessSummary
Chapter 15: Getting Started with Milvus Performance Tuning
Technical requirementsUnderstanding performance trade-offsGoal 1: making searches fasterSearch speed vs.accuracySearch speed vs.memory usageGoal 2: speeding up data ingestionWrite throughput vs.search performanceWrite throughput vs.memory usageWrite throughput vs.search freshnessGoal 3: maximizing load capacityCapacity vs.search accuracyCapacity vs.search speedMaking the right trade-off decisionsWhen search speed is your primary goalWhen data ingestion is your primary goalWhen load capacity is your primary goalMaking your searches fasterThe vector search execution pathOptimization technique 1: choosing and tuning the right indexOptimization technique 2: leveraging batch search processingOptimization technique 3: scaling with multiple replicasCombining techniques for maximum performanceSpeeding up your data ingestionUnderstanding data ingestion performance challengesOptimization technique 1: batch insertion for immediate impactOptimization technique 2: multi-shard parallelizationOptimization technique 3: bulk import for massive datasetsCombining ingestion techniques for maximum throughputMaximizing queryable data capacityUnderstanding queryable capacity bottlenecksOptimization technique 1: IVF_SQ8 quantization for substantial capacity expansionOptimization technique 2: mmap for intelligent hot/cold data managementOptimization technique 3: DiskANN for breaking the billion-vector barrierCombining capacity techniques for maximum efficiencySummary
Chapter 16: Implementing Multi-Tenancy in Milvus
Technical requirementsHow Milvus enables multi-tenancyAvailable multi-tenancy strategiesStrategy 1: Database-level isolationStrategy 2: Collection-level isolationStrategy 3: Partition-based multi-tenancyStrategy 4: Resource group-based performance isolationMulti-tenancy strategy comparisonChoosing the right multi-tenancy strategySummary
Chapter 17: How Milvus Works in AI
Technical requirementsIdentifying Milvus' role in AI applicationsMilvus with unstructured dataMilvus with RAGMilvus with multi-modal AIImage search with MilvusBuilding a chatbot with Milvus and LlamaIndexMetadata filteringSummary
Chapter 18: Unlock Your Exclusive Benefits
Unlock this Book's Free Benefits in 3 Easy Steps
Appendix
Introducing KNNS and ANNSKNNSANNSUnderstanding ANNS indexing algorithmsIVF algorithmHNSW algorithmUnderstanding ANNS quantization techniquesScalar quantizationProduct quantizationSubscribe for a free e-book
Index

Content preview from The Architecture Handbook for Milvus Vector Database

Appendix

Throughout this book, we referred to the K-nearest neighbor search (KNNS) and approximate nearest neighbor search (ANNS). If you are new to these terms, we have specifically added this appendix to provide a more in-depth explanation of these two concepts to help you gain a deeper understanding of vector search.

This appendix will cover the following main topics:

Introducing KNNS and ANNS
Understanding ANNS-based algorithms

Introducing KNNS and ANNS

Let's start with KNNS.

KNNS

KNNS stands for K-nearest neighbor search, which is a general similarity search or classification algorithm, not specifically designed for the vector search domain. Its core idea is to "find the K most similar samples," which can be applied to various data types ...

Become an O’Reilly member and get unlimited access to this title plus top books and audiobooks from O’Reilly and nearly 200 top publishers, thousands of courses curated by job role, 150+ live events each month,
and much more.

Read now

Unlock full access

More than 5,000 organizations count on O’Reilly

O’Reilly covers everything we've got, with content to help us build a world-class technology community, upgrade the capabilities and competencies of our teams, and improve overall team performance as well as their engagement.

Julian F.

Head of Cybersecurity

I wanted to learn C and C++, but it didn't click for me until I picked up an O'Reilly book. When I went on the O’Reilly platform, I was astonished to find all the books there, plus live events and sandboxes so you could play around with the technology.

Addison B.

Field Engineer

I’ve been on the O’Reilly platform for more than eight years. I use a couple of learning platforms, but I'm on O'Reilly more than anybody else. When you're there, you start learning. I'm never disappointed.

Amir M.

Data Platform Tech Lead

I'm always learning. So when I got on to O'Reilly, I was like a kid in a candy store. There are playlists. There are answers. There's on-demand training. It's worth its weight in gold, in terms of what it allows me to do.

Mark W.

Embedded Software Engineer

Publisher Resources

ISBN: 9781835881705

Cloud Computing

Data Engineering

Data Science

AI & ML

Programming Languages

Software Architecture

IT/Ops

Security

Design