دليل تقليل تكلفة السحابة

RATIO

OPÉRATION DINDON

دليل تقني عملي · Opération Dindon · يونيو 2026 · بالإنجليزية أصلًا

◆◆◆

دليل تقليل تكلفة السحابة
AWS · GCP · Azure — النسخة العالمية

◆

◆ ملاحظة على النسخة العربية

هذه الدراسة دليل تقني عملي — سلاسل أوامر دقيقة، وملفّات تهيئة، وقيم مُحدَّدة (عناوين IP، وأسماء الحزم، ومعاملات سطر الأوامر). تُبقيها هذه الطبعة العربية بالإنجليزية الأصلية عمدًا: ترجمة أمر طرفية أو اسم معامل تُفقِده دقّته وتُخاطِر بجعله غير قابل للتنفيذ حرفيًا. النصّ الأصلي هو الأداة القابلة للاستخدام مباشرة — أي إعادة صياغة تُفقِده وزنه العملي.

◆◆◆

أمين غيتي — مهندس بنية تحتية و SRE
أستاذ سابق بمدرسة هندسة · مدرّب بنية تحتية
وثيقة عامة · CC BY-NC-SA 4.0 · Opération Dindon

OPS

SRE/OPS GUIDE · TECHNICAL DOCUMENT · GLOBAL MARKET · JUNE 2026

◆◆◆

CLOUD COST REDUCTION
& PROGRESSIVE MIGRATION
AWS · Google Cloud · Microsoft Azure

◆

Operational guide for SRE, Ops and Infrastructure Architect teams
33 technical blocks · 3 cleaning sections · Reference open-source stack
Summary tables per hyperscaler

◆◆◆

33technical blocks

3cleaning sections

6open-source tables

3Summary tables

◆

TABLE OF CONTENTS — 12 PAGES

◆ Page 2 — AWS · Compute & Network (5 blocks)

◆ Page 3 — AWS · Storage · DB · Serverless · Cleanup (5 blocks)

◆ Page 4 — AWS · Progressive Migration

◆ Page 5 — GCP · BigQuery · Compute · Network (5 blocks)

◆ Page 6 — GCP · Storage · DB · AI · Cleanup (5 blocks)

◆ Page 7 — GCP · Progressive Migration

◆ Page 8 — Azure · Compute · Network · Storage (5 blocks)

◆ Page 9 — Azure · DB · AI · Data · Cleanup (5 blocks)

◆ Page 10 — Azure · Progressive Migration

◆ Page 11 — Reference Open-Source Stack (synthesis)

◆ Page 12 — Quick Reference Tables AWS · GCP · Azure

◆◆◆

This document is purely technical. It contains no legal references or political dimension.
Cost reduction estimates are indicative and based on common use cases.
Actual savings vary depending on architecture, region, volume and utilisation level.

◆◆◆

Amine RAITI · Infrastructure Architect & SRE · Paris
Public document · CC BY-NC-SA 4.0 · AI Powered by Amine

OPS

AWS · PAGE 1/3 · COMPUTE & NETWORK

AMAZON WEB SERVICES — COMPUTE · NETWORK · EKS · VPN

Indicative estimates based on common use cases · Actual savings vary by architecture and volume

◆

◆ BLOCK 1 · NAT GATEWAY — The most overestimated line itemHIGH IMPACT

Real cost: $0.045/h × 8,760h = ~$390/yr + $0.045/GB processed → up to $7,500/yr for a standard multi-AZ setup

✅ ALT 1 · NAT Instance (t3.nano/t3.small) → same function · cost ~$130–220/yr · reduction 70–85% · trade-off: manual HA management

✅ ALT 2 · VPC Endpoints (S3 · DynamoDB · SQS · SSM) → transfers to internal AWS services are free · zero refactoring if already on AWS SDK

✅ ALT 3 · PrivateLink for managed services → avoids NAT Gateway transit · billed per connection only

🔧 When to keep NAT Gateway: critical workloads requiring managed high availability without human intervention

◆ BLOCK 2 · EKS — Elastic Kubernetes ServiceAVERAGE IMPACT

Real cost: $0.10/h per cluster = $876/yr per cluster · independent of node count · 5 clusters = $4,380/yr for control plane alone

✅ ALT 1 · Cluster consolidation: 5 env. clusters → 2 (prod + non-prod) with namespaces + Network Policies → direct control plane savings

✅ ALT 2 · Karpenter vs Cluster Autoscaler → provision at exact granularity required by pods · reduces over-provisioning 20–40%

✅ ALT 3 · Spot instances for non-critical nodes (CI/CD · batch · dev) → 60–90% reduction on node cost

✅ ALT 4 · k3s or k0s self-managed on EC2 → eliminates control plane fees · for non-critical and edge environments

◆ BLOCK 3 · EC2 RESERVED INSTANCES & SAVINGS PLANSHIGH IMPACT

Trap: 1–3 year non-cancellable commitment · frequent over-commitment during growth peaks → fixed cost persists even if workloads disappear

✅ ALT 1 · Compute Savings Plans vs EC2 Instance SP → more flexibility on instance type and region · preferred for evolving architectures

✅ ALT 2 · Spot Instances for interruptible workloads → 60–90% reduction · ideal: ML training · batch · CI/CD · rendering

✅ ALT 3 · Graviton (ARM) vs x86 → same performance · 20% cheaper on equivalent instances · available on EC2 · Lambda · RDS · ElastiCache

🔧 AWS Compute Optimizer → rightsizing recommendations based on actual 14-day usage · free

◆ BLOCK 4 · VPN & CONNECTIVITYAVERAGE IMPACT

Direct Connect: billed by reserved capacity (1G ports = ~$1,500/mo) · often over-sized for light hybrid connections

✅ VPN Site-to-Site: $50–100/mo · sufficient for <500 Mbps · 80–90% reduction vs Direct Connect

✅ Direct VPC Peering (2–5 VPCs) → free at the service level · only inter-region egress is billed · avoids Transit Gateway fees ($0.05/GB)

◆ BLOCK 5 · ECS FARGATEAVERAGE IMPACT

Fargate bills per vCPU + allocated memory = 3–4× more expensive than EC2 for stable, predictable workloads

✅ ECS on EC2 Spot for predictable workloads → 50–70% reduction

✅ Fargate Spot for interruptible workloads → 50–70% reduction vs Fargate on-demand

✅ Migrate to Kubernetes (EKS/k3s) for organisations with K8s expertise → better cost control

OPS

AWS · PAGE 2/3 · STORAGE · DB · SERVERLESS · OBSERVABILITY · CLEANUP

AMAZON WEB SERVICES — S3 · RDS · LAMBDA · CLOUDWATCH · CLEANUP

Indicative estimates based on common use cases

◆

◆ BLOCK 6 · S3 & DATA TRANSFERHIGH IMPACT

Internet egress: $0.09/GB · Standard storage: $0.023/GB · storage tiers often misused

✅ S3 Intelligent-Tiering → auto-migration to IA/Glacier based on access · 40–70% reduction on cold data

✅ S3 Select → transfer only filtered data, not the entire object

✅ CloudFront in front of S3 → reduces direct requests and Internet egress

✅ MinIO S3-compatible on-premise for non-critical data → zero egress

◆ BLOCK 7 · RDS & DATABASESHIGH IMPACT

RDS Multi-AZ db.r5.large: ~$440/mo · often over-sized by 2–3× vs actual usage

✅ Aurora Serverless v2 → billed per ACU · suited for variable workloads · scale-to-zero possible

✅ RDS Proxy → reduces connections · optimises utilisation · improves resilience

✅ PostgreSQL/MySQL on EC2 Spot + PgBouncer → 40–60% reduction · trade-off: manual maintenance

✅ Rightsizing via Performance Insights → identify over-provisioned instances

◆ BLOCK 8 · LAMBDA & SERVERLESSAVERAGE IMPACT

Cold starts + duration billing + Provisioned Concurrency = frequent sources of unanticipated overspend

✅ Lambda Power Tuning → find the memory/cost optimum per function · open-source AWS tool

✅ Graviton2 for Lambda → 20% cheaper · 19% faster on average

✅ Continuous workloads → EC2 or Fargate often cheaper than Lambda beyond 15min/invocation

✅ SnapStart (Java) → eliminates cold starts on JVM functions

◆ BLOCK 9 · CLOUDWATCH · LOGS · OBSERVABILITYAVERAGE IMPACT

Log ingestion: $0.50/GB · storage: $0.03/GB · can represent 15–20% of the total bill without monitoring

✅ OpenTelemetry + Grafana + Loki self-managed → 70–90% reduction on observability

✅ Subscription filters before CloudWatch ingestion → log only what is useful

✅ S3 + Athena for long-term log archiving vs CloudWatch Logs Insights

✅ Metric Streams to Grafana Cloud → cheaper than CloudWatch custom metrics

◆ BLOCK 10 · CLEANUP — AMIs · EBS · Snapshots · Elastic IPs · ECR · Stopped InstancesIMMEDIATE SAVINGS

🔵 Stale AMIs: each AMI stores underlying EBS snapshots ($0.05/GB/mo) · a CI/CD pipeline without cleanup accumulates hundreds of AMIs → monthly audit + retention policy (e.g. keep last 5)

🔵 Detached EBS Volumes: billed at $0.08–0.10/GB/mo even without an attached instance · identify via: aws ec2 describe-volumes --filters Name=status,Values=available

🔵 Orphaned EBS Snapshots: snapshots whose source volume has been deleted · no operational value · often hundreds of GB accumulated

🔵 Unassigned Elastic IPs: $0.005/h (~$44/yr each) · an EIP not attached to a running instance is billed

🔵 Untagged / old ECR images: no pull cost but storage at $0.10/GB/mo · ECR lifecycle policy for automatic cleanup

🔵 Stopped EC2 instances: attached EBS volumes continue to be billed · associated Elastic IPs too

🔧 Tools: AWS Cost Explorer (Idle Resources) · AWS Trusted Advisor · aws-nuke (full non-prod cleanup) · CloudCustodian · Infracost

OPS

AWS · PAGE 3/3 · PROGRESSIVE MIGRATION TO ALTERNATIVES

AMAZON WEB SERVICES — STEP-BY-STEP MIGRATION STRATEGY

Indicative estimates · Effort: L=Low (days) · M=Medium (weeks) · H=High (months)

◆

◆ PROGRESSIVE MIGRATION PRINCIPLE

Do not migrate all at once. Identify high-egress or high contractual cost services first. Abstract SDK dependencies before physically migrating. Each service migrated to an open standard is an additional negotiation lever against AWS.

◆ BLOCK 11 · AWS MIGRATION STRATEGY → ALTERNATIVES

AWS service

Alternative

Method

Effort

Saving

S3

MinIO · Backblaze B2 · Cloudflare R2

Standard S3 SDK → zero refactoring

L

Egress →0

Lambda

OpenFaaS · Knative · Fission

Same Python/Node/Go runtime · adapt handler

M

60–80%

RDS PostgreSQL

Aiven · Neon · Render · Supabase

pg_dump / pg_restore · zero application refactoring

L

30–50%

EKS

k3s · k0s · EKS anywhere

Velero backup/restore · identical manifests

M

Free Control plane

CloudWatch

OpenTelemetry + Grafana + Loki

OTel agent → replaces CloudWatch agent

M

70–90%

Bedrock / SageMaker

vLLM + Llama 3 · Hugging Face self-hosted

OpenAI-compatible API → change endpoint URL only

M

60–90%

MSK (Kafka)

Apache Kafka self-managed · Redpanda

Same Kafka clients → zero refactoring

M

40–60%

ElastiCache Redis

Redis self-managed · KeyDB · Dragonfly

Same Redis protocol → zero refactoring

L

40–60%

NAT Gateway

NAT Instance t3.nano + VPC Endpoints

Replace in route tables · iptables

L

70–85%

Indicative estimates · L=days · M=weeks · H=months · discounts vs. AWS managed on-demand service

◆ RECOMMENDED MIGRATION ORDER

1. Observability (CloudWatch → OTel stack) — immediate bill impact
2. Object storage (S3 → MinIO/Backblaze B2) — zero refactoring
3. Relational databases (RDS → Aiven/Supabase) — pg_dump/restore
4. Serverless (Lambda → OpenFaaS) — adapt the handler
5. Kubernetes (EKS → k3s) — Velero for workloads
6. AI/ML (Bedrock → vLLM) — adapt the API endpoint

◆ AWS MIGRATION TOOLS

✅ Velero → K8s workload backup/restore cross-cluster

✅ AWS DataSync → S3 data transfer to external destination

✅ OpenTofu (Terraform BSL fork) → portable multi-cloud IaC

✅ Infracost → cost estimates before Terraform deployment

✅ aws-nuke → full non-prod environment cleanup

OPS

GCP · PAGE 1/3 · BIGQUERY · COMPUTE · NETWORK · CLOUD RUN

GOOGLE CLOUD PLATFORM — BIGQUERY · GKE · COMPUTE · NETWORK

Indicative estimates based on common use cases

◆

◆ BLOCK 1 · BIGQUERY — The most explosive line itemVERY HIGH IMPACT

Real on-demand cost: $5/TB scanned · a poorly written query on a 10TB table = $50 per run · a dashboard refreshing hourly = $1,200/mo

✅ Date/column partitioning → reduces scans by 80–95% on time-based queries · schema change only

✅ Clustering on frequently filtered columns → reduces scans on filter queries · stackable with partitioning

✅ Materialized views → pre-compute frequent aggregations · auto-updated · billed on delta only

✅ BI Engine → in-memory cache for dashboards · fixed monthly cost vs variable per-query cost · worthwhile from 3+ active dashboards

✅ Flat-rate slots vs on-demand → cost-effective from ~$2,200/mo in queries · 1-year or flexible monthly commitment

✅ Require partition filter → force queries to filter on partition · prevents accidental full scans

✅ Open-source alternatives: ClickHouse (OLAP columnar · 10–100× faster on aggregations) · DuckDB (local analytics · zero infra) · Apache Druid (real-time) · export BigQuery to Parquet via BQ Storage API → direct ClickHouse import with no transformation

◆ BLOCK 2 · GKEAVERAGE IMPACT

Free control plane (1 Autopilot cluster) · nodes billed on requested resource

✅ Autopilot vs Standard → Autopilot bills pods · Standard bills full nodes · Autopilot better for variable workloads

✅ Spot VMs for non-critical nodes → 60–91% reduction

✅ Vertical Pod Autoscaler → auto-adjusts requests/limits · reduces over-provisioning

✅ KEDA → scale-to-zero for event-driven workloads

◆ BLOCK 3 · COMPUTE ENGINE & CUDsHIGH IMPACT

CUD trap: 1–3 year commitment · no refund on termination (GCP s.8.8)

✅ Spot VMs → 60–91% reduction · ideal: ML training · batch · CI/CD

✅ Custom Machine Types → pay exactly the vCPU/RAM needed without rounding up

✅ Sole-tenant nodes for workloads requiring physical isolation · cheaper than over-specification

🔧 GCP Recommender → automatic rightsizing based on actual usage

◆ BLOCK 4 · VPC & NETWORK GCPAVERAGE IMPACT

Cloud NAT: IP address + processed volume · Cloud Interconnect: reserved capacity (k$/mo)

✅ NAT on Spot instance → same approach as AWS · 70–85% reduction

✅ Cloud VPN (~$50/mo) vs Cloud Interconnect for hybrid connections <500 Mbps

✅ Direct VPC Peering vs Cloud Interconnect for simple architectures

◆ BLOCK 5 · CLOUD RUN & FUNCTIONSAVERAGE IMPACT

Min-instances >0 = billed even without traffic · watch for idle costs

✅ Min instances = 0 for non-critical workloads → billed on usage only

✅ Cloud Run Spot → 60–91% reduction for batch processing

✅ Knative self-managed on GKE → removes Cloud Run dependency · same API

✅ CPU always allocated = false → CPU throttled outside requests = cheaper

OPS

GCP · PAGE 2/3 · STORAGE · DB · AI · OBSERVABILITY · CLEANUP

GOOGLE CLOUD — CLOUD STORAGE · CLOUD SQL · VERTEX AI · CLEANUP

Indicative estimates based on common use cases

◆

◆ BLOCK 6 · CLOUD STORAGE & DATA TRANSFERHIGH IMPACT

Internet egress: $0.08–0.12/GB by region · inter-region transfers: $0.01–0.08/GB

✅ Nearline / Coldline / Archive for cold data → 60–90% storage reduction

✅ Cloud CDN in front of GCS → reduces direct requests and egress

✅ Storage Transfer Service for migrations → reduced rate vs standard egress

✅ S3-compatible providers (Backblaze B2 · Cloudflare R2 · Wasabi) for non-critical data

◆ BLOCK 7 · CLOUD SQLHIGH IMPACT

HA instance db-n1-standard-4: ~$660/mo · storage and replicas billed separately

✅ AlloyDB for PostgreSQL-compatible workloads → better perf/cost on intensive loads

✅ Cloud SQL + pgBouncer → reduces connections · optimises resources

✅ PostgreSQL on GCE Spot + PgBouncer → 50–70% reduction · trade-off: maintenance

✅ Aiven for PostgreSQL · Neon · Supabase → managed · cheaper than Cloud SQL

◆ BLOCK 8 · VERTEX AI & MANAGED AIHIGH IMPACT

A100 80GB GPU: ~$3/h · V100: ~$2/h · training and inference cost highly variable by usage

✅ Spot VMs for training (interruptible by nature) → 60–91% reduction

✅ Model quantization + pruning → reduces model size and inference cost by 50–80%

✅ Hugging Face + vLLM self-hosted on Spot GPU → 60–90% reduction vs Vertex AI Prediction

✅ Ollama for local development → eliminates all inference costs in dev/test

◆ BLOCK 9 · GCP OBSERVABILITYAVERAGE IMPACT

Log ingestion: $0.50/GB · custom metrics: $0.10/metric/mo · can explode with many services

✅ OpenTelemetry + Prometheus + Grafana + Loki → full replacement · 70–90% reduction

✅ Log exclusion via Log Router → filter low-value system logs before ingestion

✅ GCP Managed Prometheus → cheaper than Cloud Monitoring custom metrics for K8s metrics

◆ BLOCK 10 · CLEANUP — Persistent Disks · Snapshots · Images · Stopped Instances · Abandoned BucketsIMMEDIATE SAVINGS

🔵 Detached Persistent Disks: billed $0.04–0.17/GB/mo by type (Standard/SSD/Extreme) even without attached VM · identify via: gcloud compute disks list --filter="NOT users:*"

🔵 Orphaned snapshots: snapshots of deleted disks · no operational value · regular audit required · cost $0.026/GB/mo

🔵 Unused Compute Engine images: custom images unused for 90+ days → automatic archiving policy to Cloud Storage (Coldline)

🔵 Stopped instances: attached persistent disks continue to be billed · reserved unassigned IPs too ($0.010/h)

🔵 Abandoned Cloud Storage buckets: buckets created for completed projects · often forgotten · Object Lifecycle Management for automatic cleanup

🔵 Unmanaged Instance Groups: instances in unmanaged groups often forgotten after migrations · quarterly audit recommended

🔧 Tools: GCP Recommender (idle VMs · underused disks) · gcpdiag · Infracost · Active Assist · Cloud Asset Inventory for full resource audit

OPS

GCP · PAGE 3/3 · PROGRESSIVE MIGRATION TO ALTERNATIVES

GOOGLE CLOUD — STEP-BY-STEP MIGRATION STRATEGY

Indicative estimates · Effort: L=Low (days) · M=Medium (weeks) · H=High (months)

◆

◆ GCP-SPECIFIC WARNING — COMMITTED USE DISCOUNTS

GCP CUDs offer no refund on termination (s.8.8). Migration must be planned so commitments are consumed naturally during transition. Never terminate a CUD early — migration savings never offset the total loss of the committed amount.

◆ BLOCK 11 · GCP MIGRATION STRATEGY → ALTERNATIVES

GCP Service

Alternative

Method

Effort

Saving

BigQuery

ClickHouse · DuckDB

Export Parquet via BQ Storage API · direct import

M

70–90%

Cloud Storage (GCS)

MinIO · Scaleway Object

gsutil rsync → zero refactoring if S3 SDK abstracted

L

Egress →0

Cloud SQL PostgreSQL

Aiven · Clever Cloud

pg_dump / pg_restore · zero application refactoring

L

30–50%

GKE

k3s · k0s self-managed

Velero backup/restore · identical K8s standard manifests

M

Free Control plane

Cloud Functions

OpenFaaS · Knative

Same runtime · adapt the input handler

M

60–80%

Cloud Monitoring/Logging

OTel + Prometheus + Grafana + Loki

OTel Collector replaces GCP agent

M

70–90%

Vertex AI (inférence)

vLLM · Hugging Face

OpenAI-compatible API → change endpoint URL only

M

60–90%

Cloud Pub/Sub

Apache Kafka · Redpanda

Adapt publishers/consumers · same semantics

M

40–60%

Memorystore Redis

Redis self-managed · KeyDB

Same Redis protocol → zero refactoring

L

40–60%

Indicative estimates · L=days · M=weeks · discounts vs. managed on-demand GCP service

◆ RECOMMENDED GCP MIGRATION ORDER

1. Observability (Cloud Monitoring → OTel stack) — immediate impact
2. Object storage (GCS → MinIO/Backblaze B2) — zero refactoring
3. BigQuery → ClickHouse — strong ROI if volume > 1TB/mo
4. Cloud SQL → Aiven/Neon — pg_dump/restore
5. GKE → k3s — Velero
6. Cloud Functions → Knative
7. Vertex AI → vLLM self-hosted

◆ GCP MIGRATION TOOLS

✅ Velero → backup/restore K8s cross-cluster

✅ gsutil rsync → GCS transfer to S3-compatible destination

✅ BQ Storage API → export BigQuery to Parquet/Arrow

✅ Migrate for Anthos → GCE VM migration to containers

✅ OpenTofu → IaC portable multi-cloud

✅ gcpdiag → audit ressources orphelines GCP

OPS

AZURE · PAGE 1/3 · COMPUTE · NETWORK · STORAGE · SERVERLESS

MICROSOFT AZURE — AKS · NETWORK · BLOB · FUNCTIONS · RESERVATIONS

Indicative estimates based on common use cases

◆

◆ BLOCK 1 · AZURE BLOB STORAGE & DATA TRANSFERHIGH IMPACT

LRS: $0.018/GB · Internet egress: $0.07–0.09/GB · inter-region transfers: $0.02/GB

✅ Archive tier → $0.001/GB storage · rehydration $0.02/GB · 94% reduction vs Hot

✅ Azure CDN in front of Blob → reduces direct requests and Internet egress

✅ Lifecycle Management → automatic archiving after X days without access

✅ MinIO S3-compatible on-premise for non-critical data → zero egress

◆ BLOCK 2 · AKS — Azure Kubernetes ServiceAVERAGE IMPACT

Free control plane · Windows nodes 40% more expensive than Linux · billed on node uptime

✅ Spot Node Pools → 60–90% reduction for non-critical workloads

✅ KEDA → scale-to-zero for event-driven workloads · évite les coûts idle

✅ Linux everywhere possible vs Windows → direct 40% saving

✅ Virtual Nodes (ACI) for bursting → no idle nodes · pay-per-pod

◆ BLOCK 3 · VPN & EXPRESSROUTEAVERAGE IMPACT

ExpressRoute: billed by reserved bandwidth · often over-sized for light hybrid connections

✅ VPN Gateway Standard (~$140/mo) for hybrid connections <500 Mbps · 80–90% reduction vs ExpressRoute

✅ Direct VNet Peering vs Virtual WAN Hub → for simple architectures (2–5 VNets) · avoids Hub fees

✅ Azure Bastion vs VPN for admin access → fixed cost · more secure than VM-based VPN

◆ BLOCK 4 · AZURE RESERVATIONS & SAVINGS PLANSWARNING CONTRACT

⚠ Non-cancellable Savings Plans · early termination fee: 12% + $50,000/yr cap · Capacity Blocks non-refundable

✅ Reserve only on the most stable and predictable resources · keep on-demand for variable resources

✅ Azure Reservations vs Savings Plans → Reservations more flexible on instance types for the same service

✅ Azure Hybrid Benefit → use existing Windows Server / SQL Server licences on Azure · 40–55% reduction

◆ BLOCK 5 · AZURE FUNCTIONS & APP SERVICEAVERAGE IMPACT

Consumption Plan · Premium Plan · Dedicated Plan → often sub-optimally chosen based on load pattern

✅ Consumption Plan → pay-per-execution · ideal for sporadic workloads (<1M executions/mo)

✅ Premium Plan → pre-warmed instances · avoids cold starts · fixed billing · worthwhile for >8h/day activity

✅ Containers on AKS Spot for continuous workloads → more control · cheaper

✅ Knative on AKS → removes Azure Functions dependency · same scale-to-zero model

◆ AZURE HYBRID BENEFIT — OFTEN UNDER-USED

If your organisation holds Windows Server or SQL Server licences with Software Assurance, Azure Hybrid Benefit allows reuse on Azure VMs or AKS. Reduction: 40–55% on Windows VM cost · up to 80% on Azure SQL database. Check systematically before any Windows workload deployment.

OPS

AZURE · PAGE 2/3 · DB · AI · DATA · OBSERVABILITY · CLEANUP

MICROSOFT AZURE — SQL · COSMOS · OPENAI · SYNAPSE · MONITOR · CLEANUP

Indicative estimates based on common use cases

◆

◆ BLOCK 6 · AZURE SQL & COSMOS DBHIGH IMPACT

Azure SQL DTU: difficult to size · Cosmos DB RU: unpredictable cost without precise monitoring

✅ vCore vs DTU → vCore more predictable · better cost/performance for stable workloads

✅ Cosmos DB Autoscale throughput → suited for variable workloads · avoids over-provisioning

✅ Serverless Cosmos DB → for sporadic workloads · billed on usage only

✅ PostgreSQL + Citus (multi-tenant) → complete open-source alternative to Azure SQL

✅ FerretDB (MongoDB on PostgreSQL) → Cosmos DB alternative for MongoDB APIs

◆ BLOCK 7 · AZURE SYNAPSE & DATA FACTORYAVERAGE IMPACT

Data Factory: billed per activity run · can explode on frequent pipelines without optimisation

✅ Pipeline consolidation → group activities to reduce the number of billed runs

✅ Self-hosted Integration Runtime → run on existing machines · eliminates Azure IR fees

✅ Serverless Synapse SQL Pool → billed on queries only · no permanent nodes

✅ Full open-source stack: Apache Airflow (orchestration) + dbt (transformation) + ClickHouse (storage)

◆ BLOCK 8 · AZURE OPENAI & COGNITIVE SERVICESHIGH IMPACT

GPT-4: $30/M input tokens · GPT-4o: $5/M tokens · thousands of $/mo for intensive usage

✅ Caching frequent responses (Redis/Memcached) → 30–60% reduction on repetitive queries

✅ Model distillation → fine-tune a smaller model on outputs of the large one → 90% inference cost reduction

✅ Prompt optimisation → reducing prompt size directly reduces the bill

✅ Llama 3 / Mistral / Phi-3 on Spot GPU + vLLM → 60–90% reduction vs Azure OpenAI

◆ BLOCK 9 · AZURE MONITOR & LOG ANALYTICSAVERAGE IMPACT

Log Analytics ingestion: $2.76/GB · retention beyond 31 days billed · Application Insights expensive at high volume

✅ OpenTelemetry + Grafana + Loki self-managed → 70–90% reduction

✅ Data Collection Rules (DCR) → filter Windows/Linux events before ingestion

✅ Basic Logs tier for low-value logs → $0.60/GB vs $2.76/GB · 8-day retention

✅ Log Analytics Archive tier → $0.03/GB/mo after 90 days · on-demand queries

◆ BLOCK 10 · CLEANUP — Managed Disks · Snapshots · VM Images · Public IPs · NSGs · App RegistrationsIMMEDIATE SAVINGS

🔵 Unattached Managed Disks: billed by type (Standard HDD $0.04/GB · Premium SSD $0.17/GB/mo) even without a VM · identify via: az disk list --query "[?diskState=='Unattached']"

🔵 Orphaned snapshots: snapshots of deleted disks · cost $0.04/GB/mo · monthly audit via Azure Cost Management

🔵 Stale VM images: generalised images unused for 90+ days → retention policy · archive to Azure Shared Image Gallery

🔵 Unassigned Public IPs: $0.0036/h (~$32/yr each) · an IP not attached to an active resource is billed

🔵 Unused NSGs and routes: no direct cost but increase attack surface and maintenance complexity · quarterly audit recommended

🔵 Expired Entra App Registrations: expired certificates and secrets · no direct cost but security risk and technical debt · Microsoft Entra Recommendations for audit

🔵 Stopped (deallocated) VMs: Managed disks continue to be billed · static public IPs too · delete or release attached resources

🔧 Tools: Azure Advisor (Cost tab) · Azure Cost Management + Budgets · AzureHound (IAM/resource audit) · Azure Resource Graph for cross-subscription audit queries

OPS

AZURE · PAGE 3/3 · PROGRESSIVE MIGRATION TO ALTERNATIVES

MICROSOFT AZURE — STEP-BY-STEP MIGRATION STRATEGY

Indicative estimates · Effort: L=Low (days) · M=Medium (weeks) · H=High (months)

◆

◆ AZURE-SPECIFIC WARNING — SAVINGS PLANS & CAPACITY BLOCKS

Azure Savings Plans are non-cancellable (12% termination fee + $50,000/yr cap). Capacity Blocks are non-refundable. Plan migration so commitments are consumed naturally. Never terminate early without first calculating whether termination cost is lower than migration gain.

◆ BLOCK 11 · AZURE MIGRATION STRATEGY → ALTERNATIVES

Service Azure

Alternative

Méthode

Effort

SAVING

Azure Blob Storage

MinIO · Backblaze B2 · Cloudflare R2

azcopy → rsync · S3-compatible SDK → zero refactoring

L

Egress →0

Azure Functions

Knative · OpenFaaS

Same Python/Node/Java runtime · adapt the handler

M

60–80%

Azure SQL

PostgreSQL managed (Aiven · Neon · Supabase)

pg_dump / pg_restore · zero application refactoring

L

30–50%

AKS

k3s · k0s · RKE2 self-managed

Velero backup/restore · identical K8s manifests

M

Free Control plane

Cosmos DB (MongoDB API)

MongoDB Community · FerretDB

Same MongoDB driver → zero refactoring if API compatible

L

40–70%

Azure Monitor / Log Analytics

OTel + Prometheus + Grafana + Loki

OTel Collector replaces Azure Monitor agent

M

70–90%

Azure OpenAI

Llama 3 · Mistral · Phi-3 + vLLM

OpenAI-compatible API → change endpoint URL only

M

60–90%

Azure Synapse / Data Factory

Apache Airflow + dbt + ClickHouse

Airflow DAG rewriting · dbt models

H

50–70%

Azure Cache for Redis

Redis self-managed · KeyDB · Dragonfly

Same Redis protocol → zero refactoring

L

40–60%

Indicative estimates · L=days · M=weeks · H=months · discounts vs. Azure managed on-demand service

◆ RECOMMENDED AZURE MIGRATION ORDER

1. Observability (Azure Monitor → OTel stack) — immediate impact
2. Blob Storage → MinIO/Backblaze B2 — zero refactoring
3. Azure SQL → managed PostgreSQL (Aiven/Neon) — pg_dump/restore
4. Azure Functions → Knative — adapt the handler
5. AKS → k3s — Velero for workloads
6. Cosmos DB → MongoDB Community — compatible API
7. Azure OpenAI → vLLM self-hosted — adapt endpoint

◆ AZURE MIGRATION TOOLS

✅ Velero → backup/restore K8s cross-cluster

✅ azcopy → Blob transfer to S3-compatible

✅ Azure Migrate → VM assessment and migration

✅ AzureHound → IAM audit and orphaned resources

✅ OpenTofu → IaC portable multi-cloud

✅ Azure Resource Graph → cross-subscription audit queries

OPS

REFERENCE OPEN-SOURCE STACK — CROSS-CUTTING SYNTHESIS AWS · GCP · AZURE

OPEN-SOURCE ALTERNATIVES BY CATEGORY — PORTABILITY & INDEPENDENCE

Indicative estimates · Portability: F=Full · P=Partial · Effort: L=Low · M=Medium · H=High

◆

◆ TABLE 1 · OBSERVABILITY — Full replacement of CloudWatch / Cloud Monitoring / Azure MonitorSAVING 70–90%

Composant

Role

Replaces

Portability

OpenTelemetry

Instrumentation & collect

CloudWatch Agent · GCP Agent · Azure Monitor Agent

T

Prometheus

Metrics

CloudWatch Metrics · Cloud Monitoring · Azure Metrics

T

Grafana

Dashboards & alerts

CloudWatch Dashboards · Google Cloud Dashboards · Azure Workbooks

T

Loki

Logs

CloudWatch Logs · Cloud Logging · Log Analytics

T

Tempo

Distributed traces

AWS X-Ray · Cloud Trace · Azure Application Insights

T

◆ TABLE 2 · OBJECT STORAGETOTAL PORTABILITY

Tool

Replaces

Effort

MinIO

S3 · GCS · Blob

L

Ceph / Rook

S3 · GCS · Blob (production)

M

Scaleway / OVHcloud S3

S3 · GCS · Blob (managé EU)

L

Standard S3 SDK (boto3 · minio-py) → zero application refactoring

◆ TABLE 3 · DATABASESTOTAL PORTABILITY

Tool

Replaces

Effort

PostgreSQL + PgBouncer

RDS · Cloud SQL · Azure SQL

L

ClickHouse

BigQuery · Redshift · Synapse

M

MongoDB Community

DocumentDB · Cosmos DB

L

Redis / KeyDB / Dragonfly

ElastiCache · Memorystore · Cache for Redis

L

◆ TABLE 4 · ORCHESTRATION & K8s PORTABILITY

Tool

Role

Effort

k3s / k0s

Lightweight K8s · replaces EKS/GKE/AKS

M

Velero

K8s cross-cluster backup/restore

L

OpenTofu

Portable IaC (Terraform BSL fork)

L

Karpenter / KEDA

Intelligent Autoscaling K8s

M

◆ TABLE 5 · OPEN-SOURCE AI & MLSAVING 60–90%

Tool

Replaces

Effort

vLLM

Bedrock · Vertex AI · Azure OpenAI

M

Ollama

Dev/test local · zéro coût cloud

L

Hugging Face + Ray Serve

SageMaker · Vertex AI · ML Studio

H

MLflow

SageMaker Experiments · Vertex Experiments

L

◆ TABLE 6 · SERVERLESS & EVENT-DRIVEN

Tool

Replaces

Portability

Effort

OpenFaaS · Knative

Lambda · Cloud Functions · Azure Functions

Total after migration

M

Apache Kafka · Redpanda

MSK · Pub/Sub · Event Hub · Event Bridge

Total (same protocol)

M

Apache Airflow

Step Functions · Cloud Workflows · Logic Apps

Total

M

OPS

QUICK REFERENCE TABLES — AWS · GCP · AZURE

SUMMARY BY HYPERSCALER — LINE ITEMS · ALTERNATIVES · REDUCTIONS

Indicative estimates based on common use cases · Économies réelles variables selon architecture et volume

◆

◆ AWS — SUMMARY BY IMPACT ORDER

Item

Main alternative

SAVING est.

Effort

NAT Gateway

NAT Instance + VPC Endpoints

70–85%

L

EC2 Savings Plans

Spot + Graviton + Compute Optimizer

20–90%

M

RDS

Aurora Serverless v2 · PostgreSQL EC2 Spot

30–60%

M

CloudWatch Logs

OTel + Grafana + Loki

70–90%

M

EKS (multi-clusters)

Consolidation + Karpenter + Spot nodes

40–80%

M

Cleaning (EBS · AMIs · EIPs)

aws-nuke · Trusted Advisor

Immediate

L

◆ GCP — SUMMARY BY IMPACT ORDER

Item

Main alternative

SAVING est.

Effort

BigQuery (queries)

Partitioning + Clustering + ClickHouse

80–95% (scans)

L–M

Compute Engine CUDs

Spot VMs + Custom Machine Types

60–91%

L

Cloud SQL

AlloyDB · PostgreSQL on Spot + PgBouncer

30–70%

M

Cloud Monitoring/Logging

OTel + Prometheus + Grafana + Loki

70–90%

M

Vertex AI (inference)

vLLM + open-source models on Spot

60–90%

M

Cleaning (Disks · Snapshots · Images)

GCP Recommender · gcpdiag

Immediate

L

◆ AZURE — SUMMARY BY IMPACT ORDER

Item

Main alternative

SAVING est.

Effort

Azure OpenAI

vLLM + Llama 3 / Mistral on Spot

60–90%

M

Log Analytics

OTel + Grafana + Loki + Basic Logs tier

70–90%

M

ExpressRoute

Standard VPN Gateway

80–90%

L

Azure SQL / Cosmos DB

PostgreSQL + Citus · FerretDB

30–70%

M

Azure Hybrid Benefit

Use existing Windows/SQL licenses

40–55%

L

Cleaning (Disks · IPs · VM Images)

Azure Advisor · AzureHound

Immediate

L

All estimates are indicative and based on common use cases. Actual savings vary depending on architecture, region, volume, and usage level.
This document is purely technical. It contains no legal references or political dimension.
Amine RAITI · Infrastructure Architect & SRE · Paris · Public document · CC BY-NC-SA 4.0
AI Powered by Amine — AI is an amplifier of ideas and forms, not the source.