Skip to content

COSTS.md

Latest commit

 

History

History
88 lines (70 loc) · 4 KB

COSTS.md

File metadata and controls

88 lines (70 loc) · 4 KB
Raw

Infrastructure Cost Analysis

Current stack — idle cost (no images processed, no requests)

Service Monthly cost Notes
AOSS ~$700 4 OCUs minimum (2 indexing + 2 search, HA enabled by default) × $0.24/OCU-hr × 730 hrs
App Runner ~$2.50 Memory ($0.007/GB-hr × 0.5 GB) charged continuously; vCPU only billed during active requests
ECR ~$0.10 Docker image storage at $0.10/GB
Lambda $0 Pay per invocation only
API Gateway v2 $0 Pay per request only
DynamoDB (on-demand) $0 Pay per request only
S3 $0 Empty bucket
Total ~$700 Almost entirely AOSS

AOSS is the only AWS-native serverless vector search service, but its minimum capacity pricing makes it impractical for development or low-traffic workloads.

Setting standby_replicas = "DISABLED" on the collection halves the OCU count to 2 (single-AZ, no HA), reducing AOSS to ~$350/month — still expensive.

Alternatives for vector storage + k-NN search

OpenSearch Service managed — easiest migration

  • Cost: ~$25/month (t3.small.search, single node)
  • Serverless: No — fixed instance
  • Migration effort: Minimal — same opensearch-py API, endpoint swap only
  • Downside: Manual instance sizing, no auto-scaling

RDS PostgreSQL + pgvector — cheapest managed DB

  • Cost: ~$13/month (db.t3.micro) + ~$32/month NAT Gateway if Lambdas need VPC access
  • Serverless: No
  • Migration effort: Medium — swap opensearch-py for psycopg2, rewrite index/search queries
  • Downside: VPC complexity and NAT cost erode the savings unless using RDS Data API

Aurora Serverless v2 + pgvector

  • Cost: ~$43/month minimum (0.5 ACU × $0.12/ACU-hr, does not scale to zero) + VPC/NAT
  • Serverless: Yes (but no true scale-to-zero)
  • Migration effort: Medium — same as RDS pgvector
  • Downside: More expensive than plain RDS once VPC costs are included

DynamoDB + brute-force cosine similarity in Lambda

  • Cost: ~$0 (DynamoDB on-demand, no requests = no cost)
  • Serverless: Yes
  • Migration effort: Medium — store vectors as DynamoDB attributes, compute cosine similarity in Lambda by scanning all items
  • Downside: Full table scan on every search — degrades beyond a few thousand images; not a real vector index

Qdrant on EC2

  • Cost: ~$8/month (t3.micro)
  • Serverless: No
  • Migration effort: Medium — swap opensearch-py for qdrant-client
  • Downside: You manage the instance, persistence, and restarts

Pinecone (free tier)

  • Cost: $0 (1 index, 100k vectors)
  • Serverless: Yes (managed by Pinecone)
  • Migration effort: Medium — swap client library, auth via API key instead of IAM
  • Downside: Not AWS-native; paid plans start at ~$70/month beyond free tier

Qdrant Cloud (free tier) — best fit for this kata

  • Cost: $0 (1 cluster, 1 GB RAM, 0.5 vCPU, persistent storage)
  • Serverless: Yes (managed by Qdrant)
  • Migration effort: Medium — swap opensearch-py for qdrant-client, auth via API key
  • Downside: Not AWS-native; API key must be stored in Lambda env vars or Secrets Manager

Summary

Option Idle cost Serverless AWS-native
AOSS (current) ~$700/mo Yes Yes
OpenSearch managed ~$25/mo No Yes
RDS + pgvector ~$13/mo (+VPC) No Yes
Aurora Serverless v2 + pgvector ~$43/mo (+VPC) Yes* Yes
DynamoDB + Lambda brute-force ~$0 Yes Yes
Qdrant on EC2 ~$8/mo No Yes
Pinecone free tier $0 Yes No
Qdrant Cloud free tier $0 Yes No

* Aurora Serverless v2 does not scale to zero — minimum 0.5 ACU always running.

For a kata or low-traffic demo, Qdrant Cloud free tier offers the best combination of zero cost, serverless operation, and purpose-built vector search performance. For a production AWS workload requiring IAM auth and no external dependencies, OpenSearch managed on t3.small.search is the pragmatic step down from AOSS.