milvus-io/milvus

Milvus is a high-performance, cloud-native vector database built for scalable vector ANN search

Structural Verdict

How Data Flows Through the System

Vector data flows from clients through Proxy to coordinators, then to worker nodes for processing and storage, with real-time streaming support

1. Client Request — gRPC/HTTP requests sent to Proxy with authentication and routing
2. Coordinator Processing — DataCoord or QueryCoord processes requests and manages worker node assignments
3. Worker Execution — DataNode ingests and indexes data; QueryNode executes searches on loaded segments
4. Storage Persistence — Data persisted to object storage (S3/MinIO) with metadata in etcd
5. Message Queue Streaming — Real-time updates propagated through message queue (Pulsar/Kafka/RocksMQ)
6. Telemetry Collection — Client metrics collected and reported to server with command push/pull capability

System Behavior

How the system actually operates at runtime — where data accumulates, what loops, what waits, and what controls what.

Data Pools

Feedback Loops

• Load Balancing (auto-scale, balancing) — Trigger: Query load metrics from QueryNode. Action: QueryCoord redistributes segments. Exit: Load balanced across nodes.
• Compaction Scheduling (scheduled-job, balancing) — Trigger: DataCoord monitors segment sizes. Action: Triggers compaction jobs on DataNode. Exit: Segments optimized.
• Telemetry Heartbeat (polling, reinforcing) — Trigger: Timer-based intervals. Action: Client sends metrics to server. Exit: Never (continuous).
• Coordinator Health Check (retry, balancing) — Trigger: gRPC connection failures. Action: Retry with exponential backoff. Exit: Connection restored or max retries.

Delays & Async Processing

• Message Queue Batch Processing (batch-window, ~configurable interval) — Groups insert operations for efficient processing
• Index Building (async-processing, ~varies by data size) — Search unavailable until index complete
• Segment Loading (async-processing, ~varies by segment size) — Collection not searchable until loaded
• Telemetry Reporting (scheduled-job, ~HeartbeatInterval (10s default)) — Metrics may be up to interval delayed

Control Points

• Message Queue Type (env-var) — Controls: Switch between Pulsar/Kafka/RocksMQ. Default: pulsar
• Storage Provider (env-var) — Controls: Object storage backend selection. Default: minio
• Telemetry Sampling Rate (runtime-toggle) — Controls: Percentage of operations to collect. Default: 1.0
• Compaction Trigger Threshold (threshold) — Controls: When to trigger segment compaction
• gRPC Max Message Size (env-var) — Controls: Maximum RPC payload size

Package Structure

This monorepo contains 6 packages:

Technology Stack

Key Components

• MilvusClient (service) — Main client interface for connecting to Milvus and performing vector operations client/milvusclient/client.go
• Proxy (service) — Gateway service handling gRPC/HTTP requests and routing to appropriate coordinators internal/proxy/proxy.go
• DataCoord (service) — Coordinates data ingestion, compaction, and segment lifecycle management internal/datacoord/server.go
• QueryCoordV2 (service) — Manages query operations, load balancing, and query node orchestration internal/querycoordv2/server.go
• RootCoord (service) — Central coordinator managing collections, schemas, and cluster metadata internal/rootcoord/server.go
• DataNode (service) — Processes data ingestion, indexing, and persistence to object storage internal/datanode/server.go
• QueryNodeV2 (service) — Executes vector similarity search and filtering operations on loaded segments internal/querynodev2/server.go
• StreamingNode (service) — Handles real-time data streaming and change data capture operations internal/streamingnode/server.go
• BulkImportAPI (handler) — HTTP client for bulk data import operations using REST API client/bulkwriter/bulk_import.go
• TelemetryManager (service) — Collects and reports client metrics with command push/pull capabilities examples/telemetry_demo/main.go

Sub-Modules

Configuration

codecov.yml (yaml)

• codecov.require_ci_to_pass (string, unknown) — default: no
• codecov.notify.require_ci_to_pass (string, unknown) — default: no
• codecov.notify.wait_for_ci (boolean, unknown) — default: false
• codecov.ci (array, unknown) — default: jenkins.milvus.io:18080
• coverage.precision (number, unknown) — default: 2
• coverage.round (string, unknown) — default: down
• coverage.range (string, unknown) — default: 70...100
• coverage.status.project.default.target (string, unknown) — default: 77%
• +11 more parameters

docker-compose.yml (yaml)

• version (string, unknown) — default: 3.5
• x-ccache.CCACHE_COMPILERCHECK (string, unknown) — default: content
• x-ccache.CCACHE_COMPRESS (number, unknown) — default: 1
• x-ccache.CCACHE_COMPRESSLEVEL (number, unknown) — default: 5
• x-ccache.CCACHE_MAXSIZE (string, unknown) — default: 2G
• x-ccache.CCACHE_DIR (string, unknown) — default: /ccache
• services.builder.image (string, unknown) — default: ${IMAGE_REPO}/milvus-env:${OS_NAME}-${DATE_VERSION}
• services.builder.build.context (string, unknown) — default: .
• +69 more parameters

tests/python_client/common/common_params.py (python-dataclass)

• description (str, unknown) — default: None

tests/python_client/common/common_params.py (python-dataclass)

• index_type (str, unknown) — default: None
• params (dict, unknown) — default: None
• metric_type (str, unknown) — default: None

Related Fullstack Repositories

Frequently Asked Questions

Analyzed on March 31, 2026 by CodeSea. Written by Karolina Sarna.