756 articles on cost optimization — what drives the cost, how it is priced, and where the savings actually are.
A function on a one minute schedule runs 43,200 times a month whether or not there is work to do. Polling schedules are the quietest source of waste in a serverless estate. Here is the arithmetic.
Resizing images in functions looks trivially cheap per image until you count storage, egress, ephemeral disk, and the variants you generate. Here is the full pipeline cost for a real workload.
Serverless wins decisively at low and medium scale. There is a point where sustained, predictable load makes containers or instances cheaper, and it is more specific than most rules of thumb suggest.
Retries are free to configure and expensive to run. Duplicate processing multiplies compute, downstream calls, and side effects. Here is what retry policy really costs and how idempotency changes it.
Archiving events is cheap. Replaying them runs your entire downstream pipeline a second time at full price. Here is what archives, retention, and a replay actually cost.
Serverless pricing is per request, which makes per-tenant attribution possible in a way shared servers never allow. Doing it well requires instrumenting the right dimensions. Here is the method.
Putting a function in a VPC is free in itself. What it pulls in is not: a NAT gateway for outbound access, data processing charges, and the choice between endpoints and gateways. Here is the bill.
FIFO queues cost 25 percent more per request, but the real price difference is throughput, batching behaviour, and the downstream design ordering forces. Here is the full comparison.
A managed proxy bills per vCPU-hour of the database it fronts. A self-hosted pooler bills as a small container. Both let you run a smaller database instance, which is where the real money is. Here is the arithmetic.
A read replica is a full copy of your database with a full copy of its price. A cache node is a fraction of that. The choice comes down to hit ratio, data freshness, and how many distinct queries you are serving. Here is the cost-per-read math.
Google prices Cloud SQL by vCPU and GB separately; AWS prices RDS by fixed instance shape. That single difference changes which cloud is cheaper for your particular ratio of CPU to memory. Here is the comparison at real prices.
The per-hour rates for managed MySQL and PostgreSQL are nearly identical on every major cloud. The cost difference shows up somewhere else entirely: in how each engine uses storage, connections, and replication. Here is where it actually lands.
Atlas M30 costs about $394 per month. The same three-node replica set on EC2 costs about $230 in instances. The $164 gap is not the whole story, and for most teams it is not even the important part. Here is the full accounting.
Most Redis clusters are sized by guess and run at 20 percent memory utilization. Sizing from working set, eviction policy, and replication overhead usually lets you drop two node sizes and save several hundred dollars a month.
Storage autoscaling grows your database disk automatically and never shrinks it. A one-off bulk load can leave you paying for terabytes you deleted years ago. Here is how the ratchet works and what it costs.
io2 Block Express charges $0.065 per provisioned IOPS per month on top of storage. Provision 40,000 IOPS and you have added $2,600 to the monthly bill before a single byte is stored. Here is when that is the right call.
PITR is billed on change volume, not database size, which is why a small write-heavy database can cost more to back up than a large static one. Here is how the meters work across AWS, Azure, and GCP.
Manual snapshots never expire. Every pre-migration safety copy taken over five years is still billing. A typical mid-sized account holds hundreds of gigabytes of snapshots nobody could name the purpose of.
HA doubles compute and storage on every major cloud, so the question is not whether availability is good but what outage cost the second node avoids. Here is how to do that arithmetic honestly.
CPU utilization alone is a bad sizing signal for databases, because the binding constraint is usually memory or connections. Here is a four-signal method that produces a defensible instance choice instead of a guess.
Serverless databases promise you pay only for what you use, but most of them have a floor. Understanding the minimum capacity charge is the difference between a $5 development database and a $44 one.
Autovacuum tuning is usually framed as a performance topic. It is also a storage bill: dead tuples on a ratcheting volume become permanent cost, and a badly tuned vacuum can add 40 percent to your database size.
Adding an index that turns a sequential scan into a lookup can remove more load than doubling your instance size, and it costs one afternoon rather than $378 a month forever. Here is how to find the queries worth fixing.
Indexes routinely occupy more space than the tables they serve, and a bloated or duplicated index set can be half your database volume. Here is how to measure it and what removing it is worth.