Skip to content

feat: USE-method host and container panels, lighter agent and leaner repo - #41

Merged
simonvanlierde merged 3 commits into
mainfrom
perf/cadvisor-enabled-metrics
Oct 2, 2026
Merged

simonvanlierde merged 3 commits into
mainfrom
perf/cadvisor-enabled-metrics

Conversation

@simonvanlierde

@simonvanlierde simonvanlierde commented Oct 2, 2026 •

Copy link
Copy Markdown
Contributor

Host & Containers now answers the USE questions for hosts and containers, the spoke agent does less work for the same data, and the repo drops code and config that only restated defaults.

  • Fix: the host network panel showed the agent container's own veth. netdev now reads PID 1's procfs and leaves out bridges and veths.
  • Host panels for the collectors the agent already ran, and container panels for PSI pressure, memory vs limit, network and disk I/O.
  • cAdvisor enables only the kinds those panels read, which skips its per-container filesystem walk (3.4–5.2% → 2.2–2.7% of a core on a 12-container host). The agent ships 8 of its own series instead of ~415 and batches for 5s.
  • One ProjectTelemetrySilent rule covers every project, still one alert instance per project/env. retired.yaml drops the old per-project rules on the next Grafana restart; delete it afterwards.
  • Breaking: just ps|logs|pull|down|tail are gone in favour of docker compose.
  • Smaller cleanups: healthchecks exit-status pings, MIG/fabric GPU panels, default-valued config, pre-rename import branches, a slimmer demo image, and ADR 0002's GPU alert claim.

cAdvisor gathered disk and network stats for every container each second, then the cadvisor_scope keep-list discarded them. enabled_metrics limits it to cpu, memory and oom_event, which cover the four shipped families.
- fix(templates): read netdev from PID 1's procfs, so host network panels show
  the host's interfaces instead of the agent container's veth
- feat(dashboards): host CPU by mode, load per core, disk utilisation and
  throughput, transmit traffic, temperature and scheduler panels; container
  CPU/memory/IO pressure, memory vs limit, network and disk I/O
- feat(templates): enable cAdvisor network, diskIO and pressure kinds and keep
  the families those panels read
- perf(templates): ship only up, export failures and queue gauges from the
  agent's own metrics; batch for 5s instead of 200ms
- perf(config): drop Prometheus's own histogram buckets like the other stack jobs
- perf(dashboards): stop requesting exemplars on GPU panels
@simonvanlierde simonvanlierde changed the title perf(templates): collect only the cAdvisor kinds the keep-list ships feat: USE-method host and container panels, lighter spoke agent Oct 2, 2026
…onfig

- feat(alerting): render ProjectTelemetrySilent and the coverage backstop into
  one projects.yaml; absent() per pair keeps one alert instance per project/env,
  and retired.yaml drops the per-project rules once
- refactor(just): drop the ps/logs/pull/down/tail delegates for docker compose;
  read the alpine image from compose.yml; mount infra files directly
- refactor(templates): ping healthchecks.io with the exit status; reuse the
  project-scoped container discovery for the GPU exporter; drop default relabel
  actions
- refactor(dashboards): drop the MIG and NVLink fabric panels
- refactor(config): drop Loki, Prometheus, datasource and dashboard-provider
  settings that restate defaults, Grafana's sign-up default, and the empty
  Cloudflare provider block
- refactor(infra): drop the pre-rename and zone-scoped branches of the import
  generator
- build(demo): OTLP HTTP exporter only, single-stage image, no unused route
- fix(bootstrap): no early exit on a checkout without .env
- docs(adr): 0002 no longer claims three GPU alert rules
@simonvanlierde simonvanlierde changed the title feat: USE-method host and container panels, lighter spoke agent feat: USE-method host and container panels, lighter agent and leaner repo Oct 2, 2026
@simonvanlierde
simonvanlierde merged commit 56b1702 into main Oct 2, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant