Google's ABFS Cuts an Android Checkout From 36m19s to 7.4s, but a Cache-Off Build Runs 23% Slower
Software / news
Google's ABFS Cuts an Android Checkout From 36m19s to 7.4s, but a Cache-Off Build Runs 23% Slower
Google open-sourced its Android build filesystem on Nov. 17. The README table behind the 8-second headline shows what happens when the cache is switched off.

Google has open-sourced ABFS, a content-addressed filesystem and build cache for Android, and claims it cuts a checkout of Android 17 from about 36 minutes to under 8 seconds. The figures are Google's own, and the README table behind them includes a result the headline numbers do not show: a build with caching switched off is slower than the old way.
The announcement, "Accelerating development for massive Git repositories", was published on Nov. 17, 2026 by the Android ABFS Team. The source is at github.com/google/abfs under the Apache 2.0 licence. The repository showed 78 stars when read.
What ABFS does
ABFS stands for Another Build File Service. A Linux kernel module called CASFS shows the full directory tree of a checkout while file contents download on first read, which is why the checkout is fast. A Ninja-compatible tool, Ninja-go, replaces timestamp checks with file-content hashes and shares build results through a remote cache served by a gRPC CAS server.
The post puts Android 17 AOSP at 1,084 Git repositories, 1.62 million files and a 300 GB built workspace. A typical tree with vendor code is 2 to 3 times that size, over 3,000 repositories. The Cuttlefish build graph has 192,249 Ninja actions. Google also says an agent checkout takes under 10 seconds, with no table behind that line.
The README table, vendor-supplied
Google ran each case three times on a GCE n2d-standard-64 VM with 64 vCPUs and 256 GB of RAM, with the ABFS server in the same VPC. The traditional baseline was repo init --partial-clone and repo sync -j8. The README says higher -j values hit Gerrit rate limiting on android.googlesource.com.
| Case | Mean time | README speedup |
|---|---|---|
| Traditional checkout | 36m19s | 1.0x |
| ABFS checkout, empty local cache | 7.4s | 294.5x |
| ABFS checkout, warm local cache | 3.6s | 605.3x |
| Traditional build | 1h20m35s | 1.0x |
| ABFS build, cache use disabled | 1h39m11s | 0.81x |
| ABFS build, empty local, full remote cache | 2m11s | 36.9x |
| ABFS build, full local cache | 1m19s | 61.2x |
The cache-disabled build took 5,951 seconds against 4,835 for the traditional one, which is 23% longer by my arithmetic from the README's figures. The README says this is expected: CASFS adds overhead, and so does "tapping", the recording of every read, write and delete so build steps can be cached.
- Traditional build4835 s
- ABFS, cache use disabled5951 s
- ABFS, empty local cache, full remote cache131 s
- ABFS, full local cache79 s
Source: google/abfs README on GitHub, mean of 3 runs on a GCE n2d-standard-64, accessed 2026-11-19
The 2m11s case is the one the README calls the common developer experience: a first build after a sync, with a remote cache filled by CI or colleagues. No figure covers a team without such a cache.
What it asks of you
The README lists who will struggle. Clients must be online with strong network performance. The remote cache runs several services that a system administrator has to maintain, and setup needs additional compute and a large amount of cache storage. Builds with dates, times or build numbers in their output will see a lower hit rate, and changing environment variables visible to build steps reduces reuse.
The client side is a kernel module. The README says CASFS has been built "against a limited number of kernel versions and tested it on fewer," and that insmod does not persist across reboots. The FUSE implementation is described as not feature complete or ready for use. The post adds that AOSP releases before Android 17 need patching, Bazel has a proof-of-concept integration, and Chromium and the Android Common Kernel are exploring compatibility.
How this compares with other speedup claims
Each vendor speedup needs the same check: baseline, cold case, cached case. We ran it on StarSling's YC launch claiming 10x faster CI, where its own pages gave three different speedups. Faster checkouts also matter for coding agents, whose review costs we covered in the Harvard paper on coding-agent review bottlenecks.
I did not build or run ABFS, and the independent coverage so far is a Hacker News submission with 5 points when checked. If you maintain an Android fork on Linux with a CI that already builds every commit: this is worth a trial on a spare machine. The client is a Linux kernel module, so it does not apply elsewhere.
Sources
More in Software
- 01OpenAI's WebRTC Voice Stack Draws a 30 ms Objection That Its Own Engineers Never AddressedLuke Curley says WebRTC drops audio older than 30 ms. OpenAI's engineers wrote about port ranges and routing, which is a different problem.
- 02Superpowers Passes 322,000 GitHub Stars as a Set of 15 Prompted Skills for 16 Coding AgentsJesse Vincent's MIT-licensed framework tells agents to brainstorm, plan and test first. Its clone count is large, and its evidence for better results is one author's view.
- 03GitHub Disables 144 Picofly Repositories After a Nintendo Notice Names OneThe November 6 notice is the first on GitHub aimed at Switch modchip firmware, and it asked for forks to be taken down without listing a single one.
- 04Chrome's Gemini Nano Download: What Google's Docs Say, and the Numbers They Leave OutGoogle's page on built-in model management gives triggers, a 30-day rule and full-download updates, but no file size, no disk threshold and no consumer off switch.