As a product manager in a previous life, I learned how bloated software used to eat server capacity when the code produced extra fields, leftover buffers, and copies that were redundant. The boxes ran hotter, pages slowed, and customers felt it and complained, and IT had to keep adding capacity. I learned to treat wasted memory as a customer experience problem, not a cleanup task for later.

Cloudflare wrote about that same waste this week, only the numbers are network-sized. Bloated records in its public DNS cache were eating memory across hundreds of billions of entries. The company published how it packed those records tighter, and how much capacity that freed, without adding servers.
On August 27, systems engineer Sebastiaan Neuteboom wrote that Big Pineapple, the platform behind the 1.1.1.1 resolver, Gateway DNS, DNS Firewall, and other Cloudflare DNS services, holds more than 250 billion cached records at once. Five changes in how those records sit in memory cut a typical entry from 953 bytes to 420 bytes. Across the fleet that freed about 100 terabytes, the RAM in roughly 130 of Cloudflare’s Gen 13 servers (Cloudflare, 2026).

Memory is expensive and servers are hard to get. Freeing 100 terabytes without buying boxes is a large saving.

I have used Cloudflare’s 1.1.1.1 on my own devices and home network for a short stretch, and I will point them at it again. On my list it sits next to Google’s public DNS at 8.8.8.8. This memory work is inside Cloudflare’s resolver.

The DNS memory change is already in production

Cloudflare started putting the smaller DNS-record design on its servers on May 18, 2026. Every DNS service that uses Big Pineapple had the change by July 6, 2026. Neuteboom’s post walks through the boring parts: drop spare capacity on strings that never grow, pack three record lists into one buffer, stop storing the owner name twice when it already lives in the cache key, and keep common records as raw bytes so the machine copies them instead of rebuilding them. Inserts rose from 625,000 to 893,000 entries per second. Lookup time fell from 828 nanoseconds to 670. At the 99th percentile, a Big Pineapple instance dropped from 9.3 gigabytes of resident memory to 5.3 (Cloudflare, 2026).

Cloudflare says one wasted byte per entry costs more than 250 gigabytes across that cache. That is why this work shows up in a public post. 

The company plans to put the freed memory back into a larger cache, which should raise hit rates and send fewer queries upstream. That is the operational close. You get more useful DNS answers out of the same racks.

At 250 billion records, even a few wasted bytes per record means buying more memory.

The disk compression is still a prototype.

On September 1, intern Aashi Patel described Cache Transcoding, built inside Pingora, Cloudflare’s proxy, during the 1.1.1.1 Intern Program. On a cache miss, eligible text is encoded with Zstandard at compression level 3 before it is written to disk. The compressed form stays on disk and moves between data centers in Tiered Cache. The edge decodes it before the client sees the page (Cloudflare, 2026).

Eligible means a 200 response, no existing Content-Encoding, a known length of at least 4 kibibytes, and a compressible type such as HTML, JSON, CSS, or JavaScript. Images, video, and fonts stay out because they are already packed. In the test corpus, those text files shrank to about one third of their original on-disk size, roughly 2.8 times denser. Patel wrote that the extra CPU cost stayed at a few percent under the traffic and reuse assumptions the intern team tested. The encode is paid once. The storage and the cross-data-center bandwidth savings keep showing up every time that object is reused.

The disk work is a prototype.I hope Cloudflare puts this disk compression into production soon.

Companies just buy more servers before asking engineers how to fix waste in the software.

I saw that as a product manager. Bloated code ate capacity, customers complained, and IT added boxes.

Cloudflare has made this argument about its fleet before. In August it said FedRAMP High runs on the same software stack as the commercial network, not a separate government island. I covered that authorization on August 28. The memory and disk posts are the private version of the same choice: squeeze the machines that are already in the building.

The freed memory will go back into a larger DNS cache. That can mean more answers served from memory, fewer lookups sent on to the domain’s own name servers, and a slightly faster 1.1.1.1. You would feel that as performance.

The disk prototype, when it ships, would save Cloudflare storage and the bandwidth it spends moving cache files between its own data centers.
Other operators can do the same work on their own fleets. Mail systems, product catalogs, session stores, and log buffers waste memory the same way. The job is to find the leftover fields and stop buying boxes to hold them.

About Shashi Bellamkonda

Loading bio…

LinkedIn Profile

Get new posts by email:

Subscribe

Powered by