[Feature][P1]: Investigate and Implement Zstd Compression for Docker Images

[rzabarazesh]

🚀 The feature, motivation and pitch

Description

Docker images currently use gzip compression by default. Zstd (Zstandard) is a modern compression algorithm developed by Facebook/Meta that offers better compression ratios and significantly faster decompression speeds compared to gzip. This could reduce image sizes and pull times, especially for the large vLLM images (~30GB). This task investigates whether switching to zstd compression provides measurable benefits.

What You'll Do

1. Research Phase :

   • Understand Docker's compression options and BuildKit support
   • Review zstd compression levels (1-19, with 3 as default)
   • Research ECR compatibility with zstd-compressed layers
   • Check if containerd/Docker runtime supports zstd decompression

2. Benchmarking Phase :

   • Build vLLM image with default gzip compression
   • Rebuild with zstd at various compression levels (3, 9, 15)
   • Measure: image size, build time, push time, pull time, decompression time
   • Test with both cold cache and warm cache scenarios
   • Benchmark on actual CI instances (g6.4xlarge)

3. Implementation Phase (if justified):

   • Update Dockerfile and BuildKit configuration
   • Modify CI pipeline to use zstd compression
   • Document compression settings
   • Deploy to staging and monitor

Deliverables

• Research document on zstd vs gzip for Docker images
• Comprehensive benchmark results (table with all metrics)
• Recommendation: which compression level to use (or stay with gzip)
• Updated Dockerfile with zstd configuration (if implementing)
• CI pipeline changes (if implementing)
• Before/after comparison in production

Alternatives

No response

Additional context

No response

Before submitting a new issue...

• Make sure you already searched for relevant issues, and asked the chatbot living at the bottom right corner of the documentation page, which can answer lots of frequently asked questions.