🔔 I’m thrilled to announce the release of my new book, "Generative AI for the Rest of Us: Your Future, Decoded". Grab your copy now and join the AI revolution before your toaster starts making memes!. Paperback: Kindle:

VarBear #486 is out now!. Devs are in the cockpit of a coding universe shaped by relentless AI innovation and shadowy pitfalls. Highlights: 🚀 NestJS secrets / 📉 AI slows devs 19% / ⚙️ Asynchrony ≠ Concurrency. Read:

I just published a 𝐦𝐢𝐧𝐢 𝐠𝐮𝐢𝐝𝐞 that highlights 𝟏𝟎 𝐜𝐨𝐦𝐦𝐨𝐧 𝐩𝐢𝐭𝐟𝐚𝐥𝐥𝐬 I’ve seen/experienced with K8s autoscaling These 𝐫𝐞𝐚𝐥-𝐰𝐨𝐫𝐥𝐝 𝐢𝐬𝐬𝐮𝐞𝐬 may affect performance, reliability, and cost. 👉

📕 Big thanks to our sponsor PerfectScale — download their free "𝐌𝐚𝐬𝐭𝐞𝐫𝐢𝐧𝐠 𝐊𝐮𝐛𝐞𝐫𝐧𝐞𝐭𝐞𝐬 𝐀𝐮𝐭𝐨𝐬𝐜𝐚𝐥𝐢𝐧𝐠" to learn the most effective autoscaling techniques ✉️ Become the smartest K8s engineer in the room:

👉 𝐀𝐮𝐭𝐨𝐬𝐜𝐚𝐥𝐢𝐧𝐠 𝐂𝐚𝐧 𝐇𝐮𝐫𝐭 𝐒𝐋𝐎𝐬 𝐈𝐟 𝐘𝐨𝐮 𝐃𝐨𝐧’𝐭 𝐏𝐫𝐞𝐝𝐢𝐜𝐭 If traffic spikes at :00 every hour and scale‑up takes two minutes, you will drop requests. To avoid that, you can scale up ahead of time, either with a scheduled job or a simple kubectl .👇.

👉 𝐑𝐞𝐪𝐮𝐞𝐬𝐭𝐬 𝐌𝐚𝐭𝐭𝐞𝐫 HPA scales based on usage ÷ request, not raw usage. Too-high requests = under-scaling, even under load. 👇.

👉 𝐇𝐨𝐩𝐞 𝐢𝐬 𝐧𝐨𝐭 𝐚 𝐒𝐭𝐫𝐚𝐭𝐞𝐠𝐲 Feed synthetic metrics into HPA to simulate Black Friday at 2a.m. Validate that scale‑out time meets your SLO before production traffic does. 👇.

👉 𝐂𝐨𝐬𝐭‑𝐀𝐰𝐚𝐫𝐞 𝐀𝐮𝐭𝐨𝐬𝐜𝐚𝐥𝐢𝐧𝐠 𝐈𝐬 𝐒𝐭𝐢𝐥𝐥 𝐍𝐞𝐰 Classic autoscalers optimise for utilisation, not $/req. Some tools can help you pick cheaper instance types, spot capacity, or consolidate nodes. 👇.

👉 𝐂𝐥𝐮𝐬𝐭𝐞𝐫 𝐀𝐮𝐭𝐨𝐬𝐜𝐚𝐥𝐞𝐫 𝐈𝐬𝐧'𝐭 𝐈𝐧𝐬𝐭𝐚𝐧𝐭 Provisioning a VM, attaching storage, and registering the node often takes 60‑120s. Buffer for that delay or pre‑scale ahead of known spikes. 👇.

👉 𝐍𝐨𝐝𝐞 𝐒𝐜𝐚𝐥𝐞‑𝐃𝐨𝐰𝐧 𝐂𝐚𝐧 𝐂𝐮𝐭 𝐂𝐚𝐩𝐚𝐜𝐢𝐭𝐲 Cluster Autoscaler may drain a node before the replacement pod is fully ready. Use PDB, readiness & preStop hooks, and multiple replicas to avoid a traffic brown‑out. 👇.

👉 𝐂𝐨𝐥𝐝‑𝐒𝐭𝐚𝐫𝐭 𝐓𝐢𝐦𝐞 𝐂𝐚𝐧 𝐖𝐢𝐩𝐞 𝐎𝐮𝐭 𝐭𝐡𝐞 𝐁𝐞𝐧𝐞𝐟𝐢𝐭 If a new pod takes 90 seconds to start, and the autoscaler reacts 15 seconds after the traffic spike, users will wait about 105 seconds before that pod is ready to handle traffic. 👇.

👉 𝐇𝐏𝐀 + 𝐕𝐏𝐀: 𝐌𝐢𝐱 𝐂𝐚𝐫𝐞𝐟𝐮𝐥𝐥𝐲 Running VPA in Auto and HPA on the same metric will cause thrashing. Instead, use VPA in Off mode to get right‑sized requests, while HPA scales replicas on custom or external metrics. 👇.

👉 𝐘𝐨𝐮 𝐂𝐚𝐧 𝐀𝐮𝐭𝐨𝐬𝐜𝐚𝐥𝐞 𝐨𝐧 𝐁𝐮𝐬𝐢𝐧𝐞𝐬𝐬 𝐌𝐞𝐭𝐫𝐢𝐜𝐬 HPA lets you use custom, external, or object metrics—QPS, queue depth, SLO latency, even revenue per pod. Most clusters already have Prom; add its Adapter and unlock a smarter scaling. 👇.

👉 𝐂𝐏𝐔 % ≠ 𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 HPA's default CPU metric works for CPU‑bound apps, but async‑I/O, GPU, or net‑bound services can stay ~0% CPU while they're actually overloaded. Always pick a metric that tracks user exp, not just CPU. 👇.

𝐌𝐨𝐬𝐭 𝐊𝟖𝐬 𝐚𝐮𝐭𝐨𝐬𝐜𝐚𝐥𝐢𝐧𝐠 𝐬𝐞𝐭𝐮𝐩𝐬 𝐚𝐫𝐞 𝐬𝐢𝐥𝐞𝐧𝐭𝐥𝐲 𝐛𝐫𝐨𝐤𝐞𝐧: Misleading metrics, poor confs, and more. and no one notices. Here are 10 gotchas that can break your autoscaling 👇

RT @eon01: Issue 486 of Kaptain is out:. Edge computing: nimble vs ironclad security—Chick-fil-A on K8s, USAF at scale. Secrets run wild, K….

RT @_FAUNVarbear_: VarBear #486 is out now!. Devs are in the cockpit of a coding universe shaped by relentless AI innovation and shadowy pi….

🚀 Kala 486 is out. From Google’s GenAI to Meta’s mega datacenters, dive into 10 paradigm shifts—from GenAI processors & economic chaos to self-evolving graphs & Netflix’s generative AI. Read →

RT @_FAUNDOL_: DevOpsLinks #485 is out! 📬. Data center costs are rising faster than your code after a hasty npm install—AI to blame? Dynamo….

RT @_FAUNDOL_: DevOpsLinks 486 is out :AI’s a wild card—speed & quality up, stability rattled. Infra adventures—server hikes to Kubernetes….

RT @_FAUNkala_: 🚀 Kala 486 is out. From Google’s GenAI to Meta’s mega datacenters, dive into 10 paradigm shifts—from GenAI processors & eco….