798 articles on cost optimization — what drives the cost, how it is priced, and where the savings actually are.
Inf2 (Inferentia) and Trn1 (Trainium) instances are AWS custom ML accelerators priced per instance-hour, often far cheaper per inference or training step than GPU instances. The catch is model compatibility. Here is when they cut cost.
Timestream bills writes per million, storage in a fast memory tier (per GB-hour) and a cheap magnetic tier (per GB-month), and queries per GB scanned. Tuning retention between tiers and scanning less are the levers. Here is the model.
Cosmos DB serverless bills per request unit (RU) consumed plus storage, with no provisioned throughput to pay for when idle. It suits spiky or low-traffic workloads; steady high traffic is cheaper on provisioned throughput. Here is how to choose.
Synapse bills dedicated SQL pools per DWU-hour, serverless SQL per TB scanned, and Spark pools per vCore-hour, plus data integration. Pausing dedicated pools and using serverless for ad-hoc queries are the big levers. Here is the breakdown.
Data Factory bills orchestration per activity run, data movement per DIU-hour, and mapping data flows per vCore-hour, plus pipeline monitoring. Chatty pipelines and always-on data flow clusters are the usual overspend. Here is the breakdown.
Managed disks bill by type and provisioned size (Premium SSD, Standard SSD, Standard HDD, Ultra), so you pay for capacity whether or not you use it. Ultra and Premium v2 add provisioned IOPS and throughput. Here is how to pick the right tier.
Databricks bills a DBU (Databricks Unit) rate per workload and tier, on top of the Azure VM cost of the cluster. Job clusters, Spot VMs, and autoscaling with autotermination are the big levers. Here is how the two-part model works.
Vertex AI bills training per node-hour, online prediction per node-hour for the endpoint you keep deployed, plus AutoML and model API usage. Idle deployed endpoints are the classic surprise. Here is where the money goes and how to cut it.
Cloud Logging charges about $0.50 per GB ingested after a 50 GB monthly free allowance, plus retention beyond the default. Exclusion filters and sampling are the main levers. Here is how the ingestion-based model works and where cost hides.
Cloud Build bills per build-minute by machine type, with a daily free allowance. Slow builds and oversized machines drive the cost. Caching, parallelism, and right-sized machines are the levers. Here is how the per-minute model works.
Cloud Tasks bills per operation (task creation and dispatch), with the first million operations free each month and about $0.40 per million after. It is inexpensive; the cost to watch is the work the tasks trigger. Here is the model.
GCP offers two network service tiers: Premium routes over Google's backbone for best performance, Standard routes over the public internet for cheaper egress. Choosing Standard where performance allows cuts egress cost. Here is the tradeoff.
A read replica adds a full instance cost to offload reads and improve read latency, but it does not help writes and adds replication lag. Knowing when a replica pays off, versus caching or a bigger primary, is the decision. Here is the framework.
Every database connection consumes memory on the server, so without pooling you scale the database up just to hold connections. A pooler like PgBouncer or RDS Proxy lets a small database serve many clients. Here is the cost-performance case.
Spot instances cost up to 90% less but can be reclaimed with little warning, so the saving comes with an interruption risk. Matching Spot to fault-tolerant work, and blending with on-demand, is how you get the discount without the outage. Here is how.
Forecasting cloud cost means projecting steady run-rate, growth, and known changes, then accounting for the variable and committed portions differently. Pre-deploy estimates make the forecast far more accurate. Here is a practical approach.
Showback reports each team's cloud cost for visibility; chargeback bills it back to their budget for accountability. Both need clean cost attribution. Here is how they differ, when to use each, and what they require to work.
Total cloud spend says little; cost per unit, per customer, per request, per transaction, tells you whether the business scales profitably. Here is how to build unit-cost metrics and why they beat watching the total bill.
Non-production environments often run 24/7 at production-like size, spending most of their money while nobody uses them. Scheduling, right-sizing, and single-AZ are the big levers. Here is how to cut non-prod cost without slowing teams.
Since February 2024, AWS charges about $0.005 per hour (roughly $3.60/month) for every public IPv4 address, whether attached or idle. Idle Elastic IPs and sprawling public IPs add up fast. Here is what changed and how to cut it.
EBS snapshots bill about $0.05 per GB-month for changed blocks (they are incremental), or about $0.0125 in the Archive tier for rarely-restored snapshots. Snapshot sprawl from unmanaged schedules is the usual culprit. Here is how to control it.
Intelligent-Tiering moves objects between access tiers automatically based on usage, with no retrieval fees, for a small monitoring charge per object. It is ideal for unpredictable access patterns. Here is when it saves money and when it does not.
Multi-AZ roughly doubles instance cost by running a standby replica, and mirrors storage, for automatic failover. Multi-AZ cluster deployments add readable standbys. Here is what you pay for high availability and when it is worth it.
Simple Email Service charges roughly $0.10 per 1,000 emails sent, plus a small charge for attachments and inbound, and standard data transfer. It is one of the cheapest ways to send email at scale. Here is the breakdown and the few gotchas.