Skip to content

Actionable Cost Audit Checklist

This document provides a hands-on checklist for auditing and optimizing the costs associated with the InsightHub project. Use it periodically to identify potential savings.

Cloud Infrastructure (Supabase)

  • [ ] Database Compute & Usage:
    • Action: Go to Dashboard > Reports > Database and check the CPU, Memory, and I/O usage patterns over the last month.
    • Review Question: Is our current compute add-on appropriate for our traffic? Are we consistently below 50% CPU usage? Could we downgrade to a smaller instance without impacting performance?
    • Review Question: Are there any unused development/staging database instances that can be paused or deleted?
  • [ ] Storage Optimization:
    • Action: Review Supabase Storage buckets. Are there large, old, or unreferenced files that can be deleted?
    • Action: Implement a lifecycle policy for storage buckets to automatically move old or infrequently accessed files to cheaper storage tiers (if available) or delete them.
  • [ ] Database & Table Structure:
    • Action: Run ANALYZE on your tables. Check for unused indexes that consume storage and slow down writes.
    • Review Question: Are we using data types that are larger than necessary (e.g., text for a field that will only ever hold 10 characters)?
  • [ ] Network Egress:
    • Review Question: Are we making large data transfers out of Supabase? Could any data processing be moved into a database function or an Edge Function to reduce egress costs?

AI/LLM Usage

  • [ ] Model Selection:
    • Review Question: Are we using the most cost-effective model for each specific task? For simple classification or summarization, could we use a smaller, cheaper model instead of a large one? (e.g., using deepseek/deepseek-chat instead of deepseek/deepseek-r1 for simple tasks).
  • [ ] Prompt & Completion Optimization:
    • Review Question: Are our prompts as concise as possible? Are we minimizing the number of tokens sent in each API request?
    • Action: Set a max_tokens limit on completions to prevent unexpectedly long and expensive responses from the LLM.
  • [ ] API Call Caching:
    • Action (Code Review): Identify LLM API calls that are likely to be repeated with the same input. Implement a caching layer (e.g., using Redis or a simple database table) to store and reuse results, avoiding redundant API calls.
  • [ ] Usage Monitoring:
    • Action: Implement logging to track the number of tokens used per API call and attribute costs to specific features or users.

CI/CD (GitHub Actions)

  • [ ] Runner Optimization:
    • Review Question: Are our CI jobs using larger-than-necessary GitHub-hosted runners? Could we switch to smaller runners for simple linting or testing jobs?
  • [ ] Workflow Efficiency:
    • Action: Review the runtime of your most frequent workflows. Are there slow steps that could be parallelized or optimized?
    • Action: Implement caching for dependencies (npm, pip, poetry) to significantly speed up job setup times and reduce runner minutes.
  • [ ] Trigger Conditions:
    • Review Question: Are workflows being triggered unnecessarily? For example, does a documentation change need to trigger a full backend deployment pipeline? Refine on: triggers and paths filters.

Data Processing & External APIs

  • [ ] Third-Party API Usage:
    • Action (Code Review): Are we making redundant calls to external APIs? Can results be cached?
    • Review Question: Are we subscribed to the optimal pricing tier for our usage volume on third-party services?
  • [ ] Data Transfer:
    • Review Question: Are we efficiently batching data sent to external services?

General Architecture & Review Process

  • [ ] Regular Audits:
    • Action: Schedule a recurring calendar event (e.g., quarterly) for the team to run through this entire checklist.
  • [ ] Cost-Benefit Analysis:
    • Review Question: For each third-party service or managed tool we use, is the cost justified by the time and effort it saves our team? Could we replace it with a cheaper or open-source alternative?