Speeding up DNF and RPM using Copy on Write
summary
Facebook runs a production DNF/RPM plugin that uses reflink copy-on-write on btrfs (and XFS support planned) to avoid rewriting package data: rpm data sections are reflinked instead of copied so installs take roughly constant time per file. A transcoder writes decompressed streams into a digest-addressable cache and a repo index of file digests lets clients request only missing ranges. Prototype hits ~95% rollout internally and an index for 100k files is ~4 MB, reducing IO and download work.
AI-written summary · about
chapters (10)
- 0:00Start
- 8:28How dnf and Rpm Work
- 12:27Ref Linking
- 14:50Installing
- 17:10Table of Contents
- 19:17Reuse Local Extents
- 23:01Repos
- 32:33Docker Images
- 34:31Dnf with Rpm Copy and Write in Production at Facebook
- 41:11Building an Image Pipeline
related talks
15mrebuilddb at Facebook scale
20mFuture of the DNF package management stack
37mDNF5 - Next Generation of Software Management
26mLocal package layering on bootc systems with DNF5
16mlightning talk, CentOS Dojo, DevConf.us
54mStill packaging like it's 1999?
01h 00mRPMing Your Apps And Tools
43mThe Forefront of the Development for NVDIMM on Linux Kernel