Сравнение векторных баз данных¶

~6 минут чтения

URL: JishuLabs, DataCamp, Medium, LiquidMetal AI, Liveblocks Тип: vector-database / pinecone / qdrant / weaviate / milvus Дата: Февраль 2026 Сбор: Ralph Research ФАЗА 5

Предварительно: RAG-техники и векторные БД

Зачем это нужно¶

Выбор vector DB определяет latency, стоимость и операционную сложность RAG-системы на годы вперёд. Pinecone (managed) -- нулевой ops, но vendor lock-in и ~$5K/мес на 100M векторов. Qdrant (Rust, open-source) -- p50 8ms и 1500 QPS, лучший по throughput. Milvus масштабируется до 10B+ векторов с GPU. pgvector -- бесплатный, но p50 50ms и нет hybrid search из коробки. Ошибка выбора на старте -- это миграция миллионов векторов позже.

Part 1: Overview¶

Executive Summary¶

Key Insight:

In 2026, all major vector databases (Pinecone, Weaviate, Qdrant) have rolled out major search speed and live scale-up improvements. Pinecone leads on simplicity, Qdrant on open-source flexibility, Weaviate on hybrid search, and Milvus on scale. pgvector is the budget option for existing Postgres users.

2026 Vector Database Leaders:

Database	Best For	License	Pricing Model
Pinecone	Simplicity, managed	Proprietary	Serverless
Qdrant	Open-source, Rust	Apache 2.0	Hybrid
Weaviate	Hybrid search, GraphQL	BSD-3	Hybrid
Milvus	Enterprise scale	Apache 2.0	Self-host/Cloud
pgvector	Postgres extension	PostgreSQL	Free (self-host)

Part 2: Feature Comparison¶

Core Features Matrix¶

Feature	Pinecone	Qdrant	Weaviate	Milvus	pgvector
Open Source	❌	✅	✅	✅	✅
Managed Cloud	✅	✅	✅	✅	✅ (Supabase)
Self-hosted	❌	✅	✅	✅	✅
Hybrid Search	⚠️ Limited	✅	✅ Best	✅	✅
Multi-tenancy	✅	✅	✅	✅	⚠️ Manual
Distributed	✅	✅	✅	✅ Best	❌
GPU Acceleration	❌	✅	✅	✅	❌
Real-time Updates	✅	✅	✅	✅	✅

Index Types¶

Database	HNSW	IVF	Flat	Custom
Pinecone	✅	❌	❌	✅ Serverless
Qdrant	✅	❌	✅	✅ Quantization
Weaviate	✅	❌	✅	✅
Milvus	✅	✅	✅	✅ GPU indexes
pgvector	✅	✅	✅	❌

Search Capabilities¶

Capability	Pinecone	Qdrant	Weaviate	Milvus
Vector search	✅	✅	✅	✅
Keyword search	⚠️	✅	✅ Native	✅
Hybrid search	⚠️ Basic	✅	✅ Best	✅
Filtered search	✅	✅	✅	✅
Faceted search	❌	✅	✅	✅
Geo search	❌	✅	✅	✅
Graph traversal	❌	❌	✅	❌

Part 3: Database Deep Dives¶

1. Pinecone¶

Aspect	Details
Founded	2019
Language	Go (closed source)
Focus	Simplicity, managed service
Best for	Teams wanting zero ops

Pinecone Strengths:

Strength	Details
Zero ops	Fully managed
Serverless	Pay per query
Simple API	Minimal configuration
Consistent performance	Predictable latency

Pinecone Limitations:

Limitation	Impact
No self-hosting	Vendor lock-in
Limited hybrid	Basic keyword search
Expensive at scale	$0.096/GB + query costs
No GPU	CPU only

2. Qdrant¶

Aspect	Details
Founded	2021
Language	Rust
License	Apache 2.0
Focus	Performance, open-source

Qdrant Strengths:

Strength	Details
Rust performance	Memory-safe, fast
Rich filtering	Complex queries
Quantization	4x-32x compression
GPU support	CUDA acceleration

Qdrant 2026 Roadmap:

Feature	Status
Agent-native retrieval	In progress
Multi-vector	Planned
Sparse vectors	Beta

3. Weaviate¶

Aspect	Details
Founded	2019
Language	Go
License	BSD-3-Clause
Focus	Hybrid search, GraphQL

Weaviate Strengths:

Strength	Details
Best hybrid search	Native BM25 + vector
GraphQL API	Flexible queries
Modules system	Embeddings, generative
Schema-first	Strong typing

Weaviate Unique Features:

Feature	Description
Generative modules	Built-in LLM integration
Vectorization modules	Auto-embed text/images
Cross-references	Link objects
Multi-tenancy	Native isolation

4. Milvus¶

Aspect	Details
Founded	2019
Language	Go
License	Apache 2.0
Focus	Enterprise scale

Milvus Strengths:

Strength	Details
Massive scale	Billions of vectors
GPU indexes	CUDA, GPU quantization
Multiple index types	HNSW, IVF, ScaNN, DiskANN
Cloud-native	Kubernetes native

Milvus Scale Benchmarks:

Scale	Latency	QPS
1M vectors	1-5ms	10,000+
100M vectors	5-15ms	5,000+
1B vectors	10-30ms	2,000+

5. pgvector¶

Aspect	Details
Launched	2022
Language	C (Postgres extension)
License	PostgreSQL
Focus	Simple, free

pgvector Strengths:

Strength	Details
Free	No extra cost
Postgres native	Single database
SQL queries	Familiar interface
ACID compliance	Transactional

pgvector Limitations:

Limitation	Impact
No distributed	Single node
Limited indexes	HNSW, IVFFlat only
No GPU	CPU only
Scale limits	<100M vectors

Part 4: Performance Benchmarks¶

Latency Comparison (1M vectors, 1536d)¶

Database	P50	P95	P99
Pinecone	5ms	12ms	25ms
Qdrant	3ms	8ms	15ms
Weaviate	4ms	10ms	20ms
Milvus	2ms	6ms	12ms
pgvector	8ms	20ms	40ms

Recall@10 (same embedding model)¶

Database	Recall@10	Notes
Pinecone	98-99%	Optimized index
Qdrant	97-99%	HNSW with quantization
Weaviate	97-99%	HNSW default
Milvus	96-99%	Index-dependent
pgvector	95-98%	HNSW

QPS (Queries Per Second)¶

Database	Single Node	Cluster (3 nodes)
Pinecone	N/A (managed)	5,000-20,000
Qdrant	2,000-5,000	6,000-15,000
Weaviate	1,500-4,000	5,000-12,000
Milvus	3,000-8,000	10,000-30,000
pgvector	500-1,500	N/A

Part 5: Pricing Comparison¶

Managed Cloud Pricing¶

Database	Storage	Query Cost	Notes
Pinecone	$0.096/GB/mo	$0.0001-0.0002/query	Serverless
Qdrant Cloud	$0.025/GB/mo	Free (compute-based)	Hybrid
Weaviate Cloud	$0.025/GB/mo	Free (compute-based)	Hybrid
Milvus (Zilliz)	$0.095/GB/mo	Free (compute-based)	Dedicated
Supabase (pgvector)	$0.125/GB/mo	Free	Postgres included

TCO Comparison (10M vectors)¶

Scenario	Pinecone	Qdrant Cloud	Milvus Cloud	pgvector
1M queries/day	$1,500/mo	$400-800/mo	$500-1,000/mo	$200-400/mo
10M queries/day	$15,000/mo	$800-1,500/mo	$1,000-2,000/mo	$400-800/mo
100M vectors	$9,600/mo	$2,500/mo	$3,000/mo	$1,250/mo

Self-Hosting Costs¶

Database	Min Hardware	Monthly Cost
Qdrant	4 CPU, 8GB RAM	$50-100
Weaviate	4 CPU, 16GB RAM	$100-200
Milvus	8 CPU, 32GB RAM	$200-500
pgvector	Existing Postgres	$0 extra

Part 6: Selection Guide¶

Decision Tree¶

Start → What's your priority?
         │
         ├──► Zero ops, managed service?
         │    └──► YES → Pinecone
         │
         ├──► Open source + performance?
         │    └──► YES → Qdrant
         │
         ├──► Best hybrid search?
         │    └──► YES → Weaviate
         │
         ├──► Billions of vectors?
         │    └──► YES → Milvus
         │
         ├──► Already using Postgres?
         │    └──► YES → pgvector
         │
         └──► Budget constraint?
              └──► YES → pgvector or Qdrant self-hosted

Use Case Matrix¶

Use Case	Recommended	Reason
Startup MVP	pgvector	Free, simple
Production RAG	Qdrant	OSS + managed option
Enterprise search	Weaviate	Hybrid search
Billion-scale	Milvus	Distributed native
Zero ops	Pinecone	Fully managed
Cost optimization	pgvector	No extra cost

Part 7: Interview-Relevant Numbers¶

Database	Market Share	Growth YoY
Pinecone	25-30%	+15%
Milvus	20-25%	+25%
Qdrant	15-20%	+40%
Weaviate	10-15%	+20%
pgvector	10-15%	+50%

Scale Limits¶

Database	Max Vectors	Max Dimensions
Pinecone	100M+	20,000
Qdrant	1B+	65,535
Weaviate	1B+	65,535
Milvus	10B+	32,768
pgvector	~100M	2,000 (practical)

Performance Numbers to Remember¶

Metric	Best	Typical
Search latency	1-5ms	5-15ms
Index build (1M)	1-5 min	5-15 min
Memory/1M vectors	2-4GB	4-8GB
Recall@10 target	>95%	97-99%

Interview Questions¶

Conceptual:

"Pinecone vs Qdrant: когда что?" -- Pinecone: zero ops, стартапы без DevOps, быстрый MVP. Qdrant: нужен контроль (self-hosted), latency <10ms, budget-conscious (open-source). Pinecone p50=20ms vs Qdrant p50=8ms.
"Почему pgvector не подходит для серьёзного RAG?" -- p50 latency 50ms (в 5x хуже Qdrant), нет нативного hybrid search, recall деградирует при >10M векторов без тюнинга, нет GPU acceleration. Подходит только если уже есть PostgreSQL и <1M векторов.
"HNSW vs IVF: trade-offs?" -- HNSW: recall 95-99%, latency <5ms, но высокое потребление RAM (полный граф в памяти). IVF: recall 90-95%, медленнее build, но меньше RAM. DiskANN: recall 90-95% с disk-based storage для >1B векторов.

System Design:

"100M векторов, 1536 dims, <15ms latency, бюджет $3K/мес -- какую БД выберете?" -- Qdrant Cloud (~$2.5K/мес) или self-hosted Qdrant ($1K infra). Pinecone $5K -- выходит за бюджет. Milvus потребует GPU infra. INT8 quantization для 4x memory reduction с <1% recall loss.

Частые ошибки

"Выбираю vector DB по benchmark latency" -- Benchmarks измеряют на идеальных условиях. В production: сетевой overhead, filtering, metadata joins, concurrent queries. Реальная latency в 2-5x выше benchmark. Всегда тестируй на своих данных.

"Больше dimensions = лучше качество" -- OpenAI text-embedding-3-large (3072 dims) лишь на 1-2% лучше text-embedding-3-small (1536 dims) по MTEB, но потребляет 2x storage и 2x latency. Matryoshka embeddings позволяют уменьшить dims без retraining.

"Managed = дорого, self-hosted = дёшево" -- Self-hosted Qdrant/Milvus требует DevOps: мониторинг, бэкапы, scaling, upgrades. TCO self-hosted часто = managed + $2-3K/мес на ops engineer time.

Sources¶

JishuLabs — "Vector Database Comparison 2026: Pinecone vs Weaviate vs Qdrant vs Milvus"
DataCamp — "7 Best Vector Databases in 2026"
Medium — "Vector Database Comparison: Pinecone vs Weaviate vs Qdrant vs FAISS vs Milvus vs Chroma"
LiquidMetal AI — "Vector Comparison"
Liveblocks — "Best Vector Database for AI Products"
Reintech — "Vector Database Comparison 2026"
LinkedIn — "Vector Database Updates 2026"