Cloudflare Big Pineapple shrinks DNS entries to free 100TB RAM

Cloudflare has optimized its Big Pineapple DNS platform to reclaim 100TB of RAM across its global server fleet. This software update allows the company to expand its cache capacity without buying new hardware, which matters for anyone relying on fast, reliable DNS resolution. The optimization directly addresses memory efficiency at a scale where a single wasted byte costs 250GB of total overhead.

Software update reclaims 100TB RAM across global server fleet

The company achieved these gains by redesigning how DNS cache entries are stored in memory. Big Pineapple now handles 250 billion cached DNS entries simultaneously, a volume that demands extreme precision in memory management. Cloudflare implemented five specific changes at the Rust level to reduce the footprint of each individual entry.

• DNS Cache Entry Size: 420 bytes (reduced from 953 bytes)
• Total RAM Freed: 100TB
• Cached DNS Entries: 250 billion
• Insert Throughput: 893,000 entries per second (up from 625,000)
• Lookup Latency: 670 nanoseconds (down from 828)

Each cached entry shrank from 953 bytes to 420 bytes. The platform replaced Rust Vec and String structures with fixed-size boxed slices and collapsed DNS record lists into single buffers. These changes reduced the p99 resident memory per instance from 9.3GB to 5.3GB. The update runs on Gen 13 servers equipped with 768GB of DDR5-6400 RAM.

The memory reduction drove a 43% increase in insert throughput to 893,000 entries per second. Lookup latency also dropped by 19%, falling from 828 to 670 nanoseconds. Cloudflare plans to use the reclaimed memory to expand the DNS cache size rather than reducing its hardware configuration. This follows a previous memory reclamation project involving the FL2 rewrite.

Cloudflare stated that it freed up roughly 100TB of RAM across its global fleet without reconfiguring any physical RAM modules in its servers. The company confirmed that it achieved this by redesigning how each DNS cache entry is laid out in memory. The optimization coincides with high DDR5 pricing trends, making the software efficiency gains particularly relevant for infrastructure costs.