Durable AI ResearchTemporal on Azure Kubernetes Service checking…

Generate a durable AI research report

Submit a topic and Temporal orchestrates a multi-step workflow — resilient to pod crashes, retries and transient failures — running on Azure Kubernetes Service.

⚡
Try the Temporal features playground Interactively explore durable timers, automatic retries, queries, signals, cancellation and the event-history log — no API key needed.
Open playground →
① Run a research request

New research request

Recent requests

② Watch it process — live metrics

📊 Temporal message processing Open full dashboard ↗

Real-time view of activity execution, task-queue polling and gRPC traffic, scraped from Temporal and the worker by Prometheus and visualised in Grafana. Auto-refreshes every 10s.

③ Understand the architecture

🏗️ Solution design — how it works

This is a durable AI research pipeline. A request flows through a Temporal workflow that orchestrates four activities. Because Temporal persists every step, the process survives pod crashes, restarts, and transient API failures — it resumes exactly where it left off instead of starting over.

Architecture

🌐 Browser UIthis page
▼  HTTP (LoadBalancer · port 80)
⚙️ API serverFlask · serves UI + REST
▼  gRPC (port 7233) — start workflow
⏱️ Temporal serverschedules & persists state
▼ ▲  polls task queue · records every step
🛠️ Workerruns workflow + activities
▼
🧠 OpenAILLM content
📄 ReportLabPDF
🗄️ PostgreSQLmetadata
🔔 Notifycompletion

The workflow (4 durable activities)

1
LLM call — sends your prompt to OpenAI (GPT-4o) to generate research content. Retries 3× with exponential backoff · 60s timeout
2
Create PDF — renders the content into a formatted PDF report with ReportLab. Retries 2× · 30s timeout
3
Store metadata — writes report ID, user, prompt & file path to PostgreSQL. Retries 2× · non-fatal if it fails
4
Send notification — signals completion (e.g. email hook). Lenient retry · optional

Why Temporal (durable execution)

  • Crash-proof: if the worker pod dies mid-run, Temporal replays history and continues from the last completed activity — no lost work, no duplicate LLM calls.
  • Automatic retries: transient errors (timeouts, 5xx) are retried per-activity with backoff, without custom retry code.
  • Visibility: every step is recorded; you can query status & results by workflow_id.
  • Separation of concerns: the workflow defines what happens and in what order; activities do the actual I/O (LLM, DB, files).

Deployment (Azure Kubernetes Service)

APIDeployment + LoadBalancer (public IP)
WorkerDeployment — executes the workflow
TemporalDeployment — orchestration engine
PostgreSQLStatefulSet + persistent disk
ConfigConfigMap (hosts) + Secret (keys)
RegistryAzure Container Registry image
MonitoringPrometheus scrape + Grafana dashboard

REST API: POST /api/workflows/generate-report · GET /api/workflows/<id>/status · GET /api/workflows/<id>/result · GET /health