Serverless mimari nasıl verimlilik sağlar?

05 Ağu 03:41

In serverless you’re basically paying for “invocations × duration × memory”, so the cost metric is a combination of request count and execution time (rounded to the nearest 100 ms). To get a clear efficiency picture, I always instrument the function with a lightweight timer (e.g., `process.hrtime` in Node) and push the latency, memory usage, and cold‑start flag to CloudWatch/Datadog. Then I compare the per‑invocation cost to the baseline (e.g., an EC2 instance) and look for patterns: a spike in duration usually means a cold start or an external service latency, while a steady rise in memory usage indicates you’ve over‑provisioned. For performance tuning, focus on three things: (1) keep the function “warm” for the most common paths—either by using provisioned concurrency (AWS) or a scheduled ping; (2) size the memory just enough to meet the CPU needs—doubling memory often cuts execution time more than proportionally, lowering the total bill; and (3) minimize cold‑start overhead by bundling only the required dependencies and lazy‑loading heavy libraries. With those metrics in place you can continuously iterate: adjust memory, add warm‑up, and watch the cost per‑invocation drop while keeping latency under your SLA.