SuperAPI AI Directory

canva-performance-tuning

Coding & development

View repository

Updated 10-15 22:42

About this skill

Canva Measured Performance Tuning

Overview

Change one controlled variable at a time and compare logical-operation outcomes, not just raw request latency. Never cache credentials or assume fixed Canva URL lifetimes and performance guarantees.

Prerequisites

  • Baseline window, normalized endpoint, workload, and local SLO
  • Current OpenAPI/response contract and endpoint rate metadata
  • Cache data class, freshness/invalidation policy, feature flag, and rollback

Instructions

Step 1: Establish a baseline

Use Read and Grep to measure logical operations, provider calls, latency distribution, job completion, retries, cache hits, errors, and queue depth with synthetic or approved data.

Step 2: Identify the bottleneck

Separate network/connection latency, unnecessary fields, pagination, duplicate reads, write retries, polling cadence, token locks, worker capacity, and local storage.

Step 3: Design one experiment

Use Write or Edit to change one cache, pagination, connection, queue, or poll control behind a feature flag with explicit success and rollback thresholds.

Step 4: Protect caches

Cache only approved metadata, key by authorization boundary, encrypt where policy requires, and invalidate on writes, consent/ownership changes, or freshness expiry.

Step 5: Tune asynchronous polling

Persist job ID, begin with a short local interval, apply bounded exponential backoff, and stop at the application budget while reconciliation continues asynchronously.

Step 6: Validate and roll out

Compare correctness, freshness, duplicate prevention, resource use, and latency. Roll back on stale authorization, missed completion, increased errors, or throttling.

Authentication

Canva Connect calls use Bearer access tokens obtained by a backend through OAuth 2.0 Authorization Code with SHA-256 PKCE. Request explicit least-privilege scopes, keep client secrets and tokens out of browser-visible state, and serialize refresh so the replacement single-use refresh token is stored atomically.

Tool Discipline

Use Read and Grep for discovery and evidence. Use Write or Edit only for the approved artifact, code, configuration, test, or receipt described by this workflow; do not make an unapproved Canva-side change.

Output

  • Scoped decision or implementation artifact
  • Redacted operation and validation receipt
  • Failure, rollback, and follow-up ownership record

Examples

A service reduces design-list calls with a tenant- and user-authorized metadata cache. The experiment proves freshness and invalidation before rollout and never stores thumbnail or export URLs durably.

Error Handling

Failure Response
No baseline exists Instrument before optimizing
Cache serves stale authorization Disable it and fix ownership/consent invalidation
Polling increases 429s Lower scoped concurrency and widen bounded intervals
Latency improves but correctness regresses Roll back the experiment

Resources