Public Cloud Basis:

Public cloud known for its ubiquity, cost-effectiveness and pay-as-you-go model is appealing to host an analytical framework that can collect traffic from anywhere in the world. We chose Azure in our case study but the use of an public cloud is not only feasible but also recommended for replicating the study and most public clouds offer parity in the features used with our analytical framework. Our choice of Azure was based on the <10 ms latency for resources with connections over the Azure high-speed backbone.

The resource types and the cost-calculations are presented here as basis for our studies for cost-effectiveness of drone video sensing analytics that follows next.

Our Pipeline Cost Estimates:

Component Assumption Monthly Estimate $

AKS Cluster 3-node (Standard_D4s_v5) w/ airflow ~0.10/hr x 730 hrs = $73.00

VM Instances (3 x D4s_v5) Bursty ~150/month each = $450

Storage/Data Volume 12GB Hot Tier ~1.80 per month

Backup (AKS Snapshots) Daily ~$5.00 per month

Network Egress 50GB Central US region ~$3.50 per month

Monitoring and Logs Centralized ~$15.00 per month

Azure Data Factory Orchestration + 1 DIU x 1hr/day self-hosted IR ~8.00 per month

MySQL Flexible Server 2 vCores, 8GB RAM ~$124.83 / month

MySQL storage 20GB ~$0.115 x 20 = $2.30/month

MySQL Backup Daily 7-day retention ~$1.00 per month

Application Gateway 1 instance ~$300 per month

Azure Databricks Premium Tier, 2-node DS13_v2 cluster

VM: 3 x $0.598/hr x 730 hrs

DBU: 2 nodes x 2 DBUs / hr x $0.55/DBU x 730 hrs w/ airflow VM: ~$120/month

DBU Cost: ~$160/month

Azure Cognitive Search 1 index 1 GB 1 semantic ranker $249.98/month

Total Estimated Cost All of the above ~$1514.43/month

Typical End-User Resource-Type Cost Basis

Resource Type Monthly $ Quantity

Application Gateway 300/Unit 1

MySQL 30/unit 1

AKS 50/unit 1

Databricks 12/unit 1

Storage Account 0 2

Key Vault 0 2

ADF 8/unit 1

Cognitive Search 1 index 1 GB 1 semantic ranker 1

External commodity model or Large Language-Model usage costs:

 Unit Quantity Price

Storage 12 GB Hot Tier 1 $1.80 per month

Vector Store Image + vector + metadata 26 $0.36 per month

Compute Serverless #Number of Agents ~0.10/hr in burst mode x number of queries per hour as 1 x number of effective hours as 10 = $1.00

Network 1 Virtual Network (egress/dns/tls certificates) 1 $12.00 per month

LLM Tokens 1 token 202629 $0.40 to $30+ per million output tokens

Training+Tuning+Deployment Commodity $0.65/month

Streaming Stack cost:

 Size Quantity Price

Storage 12GB Hot tier 1 $1.80 per month

Vector Store Image + vector + metadata 17833 $249 per month

Compute 3-node (Standard_D4s_v5) AKS instance 1 ~0.10/hr x 730 hrs = $73.00

Network 1 Virtual Network (egress/dns/tls certificates) 1 $12.00 per month

LLM Tokens 1 token 100 Million tokens $0.40 to $30+ per million output tokens

Training+Tuning+Deployment $200 per month

The above costs are inclusive of both CapEx (initial) and OpEx (recurring) costs for realizing a fully functional drone video sensing analytics framework. However, which most of these costs are similar between operational and analytical frameworks stemming from the use of the same resource-types, it must be noted that Operational frameworks lean more on computation power and consumption versus analytical frameworks. With importance-based sampling, the total cost of ownership reduces compute time by a factor of 2 at least as compared to operation-only workloads. Furthermore, analytics frameworks leverage commodity models, commodity compute and fine-grained task-library to leverage only those necessary for a query. Analytical frameworks are also easier to build focusing on narrow tasks and leverage multiple and cheaper compute as opposed to doubling down on expensive compute for everything from training, testing, deploying and predictions.