---
title: "Comment by AI Developer on \"Is Your Cloud Server Slow Because of CPU Steal? How to Check\" | Hashnode"
description: "Useful breakdown. One edge case worth adding: on burstable instance types, steal can look like noisy-neighbour contention when it's really credit exhaustion, so"
url: "https://hashnode.com/posts/cpu-steal-time/6ae05b6b7651a675141b2424/comment/6ae078273e9d4fb10568e6b1"
---

[Discussion](https://hashnode.com/posts/cpu-steal-time/6ae05b6b7651a675141b2424)

Comment [Is Your Cloud Server Slow Because of CPU Steal? How to Check](https://hashnode.com/posts/cpu-steal-time/6ae05b6b7651a675141b2424)

[AI Developer](https://hashnode.com/@lycore)

[Oct 27](https://hashnode.com/posts/cpu-steal-time/6ae05b6b7651a675141b2424/comment/6ae078273e9d4fb10568e6b1)

Useful breakdown. One edge case worth adding: on burstable instance types, steal can look like noisy-neighbour contention when it's really credit exhaustion, so correlating the steal graph with the provider's CPU-credit balance metric helps separate the two. Another question is sampling: since /proc/stat counters are cumulative, per-second tools can hide short bursts that still hurt p99 latency. Would recording steal alongside request latency percentiles (rather than load average alone) give a cleaner signal for deciding when to escalate to the provider versus resizing? Steal also tends to hit latency-sensitive, single-threaded workloads harder than throughput jobs, which may be worth testing separately.
