Skip to content

Personal knowledge base · DevOps

I systematize practices.
I put them to work.

devops-kb is a personal map of the best DevOps practices. Every card follows the same structure: problem → solution → example → adoption status. It works as long as it gets applied.

New practices · October 29, 2026

21 cards: DevOps, Platform/DevSecOps, MLOps/LLMOps and AI tools.

🔄

Renovate for GitLab

Scheduled MRs updating dependencies and images by digest.

Not studied
↗
🧰

gitlab-ci-local

Run .gitlab-ci.yml jobs locally before push.

Not studied
↗
🏅

OpenSSF Scorecard

Assess a repository: dependency pinning, dangerous workflows, maintenance.

Not studied
↗
🧯

agent-scan

Check MCP servers and skills before installing.

Not studied
↗
🧪

Ansible-lint + Molecule

Profile-based linting and role tests: converge, idempotence, verify.

Not studied
↗
🚧

CrowdSec for a VPS

Attack detection from Nginx and sshd logs, blocking via a bouncer.

Not studied
↗
🕵️

zizmor

Find dangerous triggers, unpinned actions and excessive permissions in GitHub Actions.

Not studied
↗
🧬

Grype on an SBOM

Find CVEs in a stored SBOM without rebuilding the image.

Not studied
↗
🚪

LiteLLM

A single OpenAI-compatible gateway: keys, budgets, local and cloud models.

Not studied
↗
🔭

Langfuse

Traces, tokens and cost of LLM requests.

Not studied
↗
🎯

promptfoo + garak

Prompt regression tests in CI and model vulnerability scanning.

Not studied
↗
🎓

MLOps Zoomcamp

A learning track: MLflow, model deployment and monitoring, CI/CD for ML.

Not studied
↗
🧩

Agent Skills

Package a repeatable procedure as a skill for an AI agent.

Not studied
↗
🏛️

Backstage

A service catalog, templates and TechDocs in an internal portal.

Not studied
↗
⚡

vLLM

Your own inference server with an OpenAI-compatible API.

Not studied
↗
☸️

Read-only Kubernetes MCP

Agent access to the cluster: read_only, the view role, Secrets denied.

Not studied
↗
🔎

HolmesGPT

AI analysis of Prometheus and Alertmanager alerts.

Not studied
↗
👥

agency-agents

Role scaffolds for subagents, tailored to your stack.

Not studied
↗
📜

Conftest

Rego policies for YAML and other configs.

Not studied
↗
🐢

Sloth: SLO

SLOs and error budget burn alerts for Prometheus.

Not studied
↗
✍️

Cosign

Sign images and verify the signature before deploy.

Not studied
↗

Practices · October 11, 2026

Studied

🛰️

ss -tulpan instead of netstat

A modern netstat replacement for sockets and processes. Dozens of times faster on busy hosts.

Studied
↗
🐳

Hadolint in pre-commit

Automatic Dockerfile validation before commit: apt without --no-install-recommends, pinned versions, safe USER.

Studied
↗
☸️

ArgoCD app-of-apps pattern

One root Application manages a dozen children. The cluster's GitOps state lives in one repository.

Studied
↗
🧱

Terragrunt for DRY modules

One Terraform module — dozens of environments without copy-paste. terragrunt.hcl + include.

Studied
↗
📊

Four Golden Signals

Latency · Traffic · Errors · Saturation — the minimum mandatory dashboard for any service. From the Google SRE book.

Studied
↗
🛡️

Trivy image scanning in CI

The build fails on HIGH/CRITICAL CVEs before the push to the registry — vulnerabilities get fixed, not accumulated.

Studied
↗
🪝

pre-commit for IaC

tflint · yamllint · detect-secrets in one .pre-commit-config.yaml — locally and in CI.

Studied
↗
📈

Prometheus recording rules

Heavy PromQL is precomputed in the background — dashboards open in milliseconds, not 15 seconds.

Studied
↗
🎛️

The USE method for capacity

Utilization · Saturation · Errors for every resource. The bottleneck shows up in a minute, without "let's add nodes".

Studied
↗
🛡️

kube-bench · CIS Benchmark

Automatic cluster check against 100+ CIS Kubernetes items. A Job in the cluster, a report in Slack.

Studied
↗
🔒

Nginx TLS hardening

One tls.conf — TLS 1.2/1.3, HSTS, OCSP stapling. SSL Labs A+ on every server block.

Studied
↗

Key principles

🎯

Practicality

Only what actually applies at work — no theory for theory's sake.

🧩

Gaps first

Start with the areas where knowledge is most lacking.

✅

Applied = paid off

A card is valuable only once it is put to use, not once it is read.