Performance Optimization

Resource Sizing by File Type

Source	CPU	Memory	Ephemeral Disk	Processing Time
2K EXR (8MB)	500m	512Mi	500Mi	~5s
4K EXR (50MB)	1000m	1Gi	1Gi	~15-30s
4K DPX (25MB)	1000m	1Gi	1Gi	~10-20s
8K EXR (200MB)	2000m	2Gi	3Gi	~60-120s

Ephemeral disk usage per file: source + 2 OCIO intermediates + PNG intermediate + outputs. Formula: ~3x source size peak.

Setting	Value	Rationale
RPS factor	1	CPU-bound work; don't share a pod
minScale	1	Avoid cold start (oiiotool + ffmpeg binary load)
maxScale	5	Cap total cluster memory (5 x 2GB = 10GB)
Scale-down delay	60s	Avoid pod thrashing during burst ingestion

S3 download -- Full file download required (no Range GET like exr-inspector). Mitigated by boto3 adaptive retries and connection pooling (25 connections).
ffmpeg H.264 encoding -- CPU-bound, single-threaded with -preset fast. Consider -preset ultrafast for 2x speed at cost of ~30% larger files.
Color transform -- CPU-bound, two transforms per file (sRGB + Rec709). Most time spent on the resize operation, not the colorconvert.
S3 upload -- Two uploads per file (thumbnail + proxy). Proxy upload is the larger one (typically 1-5MB for a single-frame proxy).

Set delivery to unordered -- each file is independent, no sequencing needed
Use the same S3 endpoint for download and upload to minimize network hops
The function cleans up all /tmp files after each request, so ephemeral disk does not accumulate between invocations
If processing 8K+ files, increase OIIO_TIMEOUT beyond the default 300s

Change	Impact	Status
S3 ContentType + tags on upload	Browser delivery ready	Done (v2.0.0)
ffmpeg `-movflags +faststart`	Browser streaming without full download	Done (v2.0.0)
Batch color transform (single intermediate)	-1 transform per file	Planned
ffmpeg `-preset ultrafast` option	2x encode speed	Planned
Sequence-aware processing (detect frame range)	Skip duplicate transforms	Planned