Workers

We use Temporal to run workflows and activities in the background. If you have not used Temporal before, their docs have a comprehensive introduction. We recommend reading that to get used to the concepts that we use on a daily basis.

Temporal Workers

Temporal provides different types of workers to handle different workloads. Our setup includes:

• Default Worker: Handles I/O bound workflows and activities (API calls, database operations)
• Document parser: Handles document parsing, image extraction, and selective PDF OCR

Default Worker

The default worker uses a thread pool executor to handle I/O bound operations. This worker:

• Runs workflows and activities that are primarily I/O bound
• Uses ThreadPoolExecutor for concurrency
• Handles most of our business logic workflows

Document parser

The document parser is the single ingestion service for supported document formats. It:

• Parses Office files through AnyDoc
• Parses PDFs and selectively OCRs image-only pages through pdf-inspector
• Runs on the document-parser task queue

Our Setup

Our backend is highly concurrent and, for the most part, I/O bound. Most of our time is spent waiting for API calls to LLMs and other external services.

Production Setup

We run multiple Temporal workers:

• Default workers: Handle I/O bound workflows and activities
• Document parser workers: Handle document parsing and OCR with resource limits
• Temporal Server: Manages workflow state and scheduling

Worker Configuration

Workers are configured in our Docker setup:

worker:
  image: slidespeak-backend-code
  command: python -m server.temporal.worker
  depends_on:
    - temporal

document-parser:
  image: slidespeak-document-parser
  command: bun run start
  depends_on:
    - temporal