Leo’s Build Log
PostsProjectsAbout한국어

PinLog

An AI place journal that finds places again by experience and feeling, even when you forget the name

Screens

Overview

Period
Jul – Aug 2026 (5 weeks)
Team
6 people
Role
Infra lead

The problem

Saved places are hard to find again once you forget the name, and why you saved them is lost. On the infra side, several services had to be deployed and run on a single given server without cloud API access.

The approach

  1. 01
    Natural-language search over your own contextPlaces are saved with the reason and experience behind them, and found again by a sentence instead of a name.
  2. 02
    Discovery through anonymous collectionsPublic collections that hide identity and original notes let you discover others' taste, then add your own context to what you find.
  3. 03
    GitOps on a single serverEvery service ships through a shared Helm chart, and Argo CD applies the state declared in Git to k3s.

Architecture

ArchitectureServicesPlatformDeliveryObserve · alertLogsscrapechanged by a person
Frontend
Backend
AI Service
k3s1 VM · 4 vCPU · 15 GiB
containerdBuilt-in runtime
Helm · ApplicationSetConfig → deploy unit
GitHub ActionsBuild · pin SHA
Infra repo PRRequired checks
Argo CDGitOps sync
PrometheusMetrics
AlertmanagerAlerts
LokiLogs
SentinelLLM summary
MattermostTeam channel
PersonDecides fixes
Single VM (Linux)
  1. 01
    Runtimek3s embedded containerd (Docker and cri-dockerd removed)
  2. 02
    DeliveryGitHub Actions → infra repo PR → Argo CD → Helm and ApplicationSet
  3. 03
    ObservabilityPrometheus, Alertmanager, Loki → Sentinel (in-house alert summarizer) → Mattermost → a person

User scenario

User scenarioUserSystemRetry pathnoyes
User
Pick a spotDrop a pin
Add a placeMemo · memories
My saved placesBrowse saved list
Search in words“A quiet café for reading on a rainy day”
Open recordRevisit · add
Backend savesPin · record
AI searchMatches your records
Highlight pinsOn the map
Right place?
RephraseChange terms, search again
  1. 01
    User flowPick a spot → record → list → natural-language search → open the record
  2. 02
    System flowSave → AI search over your records → highlighted results
  3. 03
    Retry pathNot the place you meant? Rephrase and search again

My work

  1. 01
    Only verified images shipDeploy PRs are opened only after CI success and the registry digest are verified, the commit is rechecked before merge, and rollbacks are recorded as Git reverts.GitHub Actions · private GHCR · pinned full commit SHA + image digest · Helm · Argo CD ApplicationSet
  2. 02
    Splitting one server's resourcesSeparate dev and prod areas, resource ceilings and capped observability storage keep services from starving each other.Namespaces · resource budgets · NetworkPolicy · Pod Security Admission
  3. 03
    AI-assisted alertsAlerts are summarized in Korean from a limited set of related metrics and logs. If the AI fails or evidence is thin, a rule-based alert goes out instead, and nothing in the cluster is changed automatically.Alertmanager → Sentinel Receiver (redaction · allow-listed JSON) → AI API → output validation → Mattermost
  4. 04
    Observability and external probesMetrics, logs and dashboards, plus a separate external HTTPS/TLS probe so a full server outage is still reported.Prometheus · Loki · Alloy · Grafana · GitHub-hosted external probe
  5. 05
    Secrets and accessSecrets are encrypted in Git, and deploy automation tokens are scoped per repository with least privilege.Sealed Secrets · least-privilege tokens · Cloudflare Tunnel · Traefik

Why these tools

k3sFitting several services on one 4 vCPU / 15 GiB VMA lightweight cluster runs three services plus observability
containerdCPU throttling on the Docker / cri-dockerd pathSwitched to the k3s built-in runtime; call path cut from 5 hops to 3
ArgoCD · HelmThe same deploy config repeated for every serviceShared chart and ApplicationSet; a new deploy unit is one config file
GitHub ActionsUnreviewed changes reaching prod; untraceable imagesPinned commit SHA and digest, applied via infra PR after required checks
Prometheus · Alertmanager · LokiNoticing trouble only after the factMetric-based alerts, pulling only the related logs as evidence
LLM API · MattermostTriaging repeated alerts by handAutomated alert and log summaries; server changes stay a human decision

Improvements

  1. 01
    Container runtime eating CPUKubelet polling through Docker and cri-dockerd kept the CPU busy; switching to K3s embedded containerd removed that layer, using a procedure with DB backup, preserved config and a rollback path.
  2. 02
    A broken backup could become the latest restore pointDumps are now checked as archives, and only verified files are atomically promoted to latest.dump.

Related posts