DEV Community

Cover image for Rent the Cores, Keep the Phone: Building LineageOS 24 on EC2 and Iterating at Home
AWS Community Builders profile image xbill
xbill Subscriber for AWS Community Builders

Posted on

Rent the Cores, Keep the Phone: Building LineageOS 24 on EC2 and Iterating at Home

#aws #android #lineageos #ec2

This article provides a step by step guide to running the bulk of a LineageOS 24.0 build on an Amazon EC2 instance and finishing it on a home machine, where the phone and its logs are. Claude Code sizes the instance, runs the build, fixes the port errors and moves the tree.

https://github.com/xbill9/android_kernel_google_msm-4.9/tree/main/backport-tools/lineage-24.0/aws

A from-scratch LineageOS 24.0 build for the Pixel 3a stalled on a 32 GB home machine during Soong analysis. On a c7a.48xlarge the tree synced in 15 minutes and the first build reached 156,615 of about 232,000 steps in 19 min 17 s. Building on EC2 at the home machine's own path, with the home user name, makes the result portable: the 266 GB tree came home in 1 h 37 min and the first local m bacon produced the zip in 2 h 30 min, with one Soong analysis re-run.

What is this project trying to Do?

The Pixel 3a (sargo) runs LineageOS 23.2 on its original 4.9 kernel. The next step is 24.0, which is Android 17, and the port is a loop: change the device tree, build, flash, read the boot log, repeat.

That loop has two halves with different needs. The full build wants many cores and a lot of memory once. Every iteration after it wants the phone on a USB cable and a build that redoes only what changed.

The plan: rent the first half, keep the second half at home, and let Claude Code run the hand-off.

At This Point You Should Have…

  • An AWS account with an on-demand vCPU quota of at least 192 for standard instances (this one has 264)
  • A home build machine with room for the whole tree; this one has 278 GB free and is an i7-3770 with 8 threads and 31 GB of RAM, running Debian sid
  • The LineageOS 24.0 local manifest and device-tree forks from the repository
  • Claude Code with the AWS MCP server
  • A clone: git clone https://github.com/xbill9/android_kernel_google_msm-4.9

Step 1 — Size the Instance From the Last Local Build

The port notes record what the home machine saw. Soong analysis held 31 GB resident with 44 GB of swap in use at the peak and took about 1 h 45 min. Compiling then ran 8 threads at 94-95 % user time, and one run finished 130,617 of about 233,000 steps in 7 h 49 min.

So the instance needs at least 75 GB of RAM for analysis and as many cores as the compile can use. Claude Code queried prices and quotas through the AWS MCP server:

Instance vCPU RAM (GiB) On-demand $/h Spot $/h, cheapest zone
m7a.16xlarge 64 256 3.71 0.86
m7a.24xlarge 96 384 5.56 1.28
c7a.48xlarge 192 384 9.85 2.93
m7a.48xlarge 192 768 11.13 3.19

The spot placement score for m7a.48xlarge was 1 out of 10 in every us-east-1 zone, so a 192-core spot request is likely to be refused or reclaimed mid-build. The build runs on-demand on a c7a.48xlarge: the same core count as the m7a, and 384 GiB leaves room well beyond the 75 GB analysis peak.

Step 2 — Launch at the Home Machine's Path

The instance is Ubuntu 24.04 with a 1000 GB gp3 root volume at 16,000 IOPS and 1000 MB/s. SSH is open to the home IP only, with an imported key pair.

Two settings decide whether the finished tree can come home and build incrementally. The tree lives at /home/xbill/android/lineage-24.0, the same absolute path as on the home machine, and the build runs with BUILD_USERNAME=xbill.

The reason is in the build output. The Soong graph uses relative paths, while the Kati graph names the tree by absolute path:

out/soong/build.lineage_sargo.ninja: 0 lines with /home/ubuntu of 3712391
out/build-lineage_sargo.ninja: 124 lines with /home/ubuntu of 5697004
Enter fullscreen mode Exit fullscreen mode

And Soong records BUILD_USERNAME among the environment values that trigger a fresh analysis:

soong.environment.used.lineage_sargo.build BUILD_USERNAME ubuntu
Enter fullscreen mode Exit fullscreen mode

A different user name at home means a new analysis there, the step that needed 75 GB.

Step 3 — Sync From Scratch

ec2-build.sh installs the build packages, runs repo init at depth 1, installs the local manifest and syncs with 32 jobs. On EC2 the platform clang is synced by repo in full, where the home tree uses a sparse checkout.

14:39:26 repo sync
14:53:00 sync attempt 1 exit=1 free= 828G
14:54:09 sync attempt 2 exit=0 free= 828G
Enter fullscreen mode Exit fullscreen mode

Sync took 15 minutes. One repository answered HTTP 429, the AOSP server's rate limit, and the retry loop fetched it a minute later.

Step 4 — Build

lunch lineage_sargo cp2a userdebug, then m bacon at the default parallelism, which is 194 jobs on this instance.

14m50.05s Build Failure: 156615 done, 1 failed, 75511 remaining - 175.96/s
real    19m16.886s
user    2259m47.302s
Enter fullscreen mode Exit fullscreen mode

Analysis and two thirds of the compile took 19 min 17 s, against 1 h 45 min for analysis alone at home. The failure is a port error, covered next.

Step 5 — Fix Port Errors at Cloud Speed

Each fix goes to the fork on GitHub, the instance syncs that one project, and the build resumes from where it stopped. Four Android 17 changes surfaced in order:

Error Fix Rebuild
unknown type logger_app drop the device tree's logger_app.te; Android 17 no longer defines the type 5 min 49 s
unknown type wifi_logging_data_file stop adding wifi_perf_diag policy; Pixel adds it for eng builds only 9 min 51 s
Found conflicting genfscon rules for /class/typec use the platform's sysfs_typec label; keep sysfs_usb_c on usbc0 16 min 36 s
Unresolved symbol ... createSurface ... LayerMetadata in libsecureuisvc_jni.so shim library exporting the Android 12 symbol the build finished with the zip

The fourth is a vendor blob compiled when SurfaceComposerClient::createSurface took LayerMetadata by value. Android 17 takes it by const reference, so the mangled name changed. A shim exports the old name and forwards:

sp<SurfaceControl> SurfaceComposerClient_createSurface_metadataByValue(
        SurfaceComposerClient* thiz, const String8& name, uint32_t w, uint32_t h, PixelFormat format,
        int32_t flags, const sp<IBinder>& parentHandle, gui::LayerMetadata metadata,
        uint32_t* outTransformHint)
        __asm__("_ZN7android21SurfaceComposerClient13createSurfaceERKNS_7String8EjjiiRKNS_2spINS_7IBinderEEENS_3gui13LayerMetadataEPj");
Enter fullscreen mode Exit fullscreen mode

The __asm__ label gives an ordinary C++ function the blob's symbol name, and a free function with the object as its first argument has the member function's calling convention. The body calls thiz->createSurface(...) with the same arguments.

extract-files.py gets a blob_fixup().add_needed('libgui_shim_secureui.so') for the blob, so a fresh extraction reproduces the patch, and proprietary-files.txt pins both the original and the patched SHA-1.

🔎 Tip: Same Commit Hash on Both Machines

The instance has no GitHub credentials. Commits made there travel as git format-patch --binary files and are applied at home with git am --committer-date-is-author-date under the same committer identity. The result has the same hash as the instance's commit, so after the push the instance tree and GitHub agree, and repo sync on the instance changes no file.

Step 6 — Trim the Tree Right After the Sync

Sizes on the instance with the build at 85 %:

Part Size Needed at home
out/soong/.intermediates 90 GB yes, it is the build state
.repo/project-objects 32 GB yes, every project's git data
prebuilts/clang/host/linux-x86 19 GB two of five versions
prebuilts/rust-toolchain 13 GB yes
out/target/product/sargo 14 GB yes; packaging is rebuilt in minutes

The home machine has 278 GB free once its old 24.0 tree is gone. Three clang versions (r547379, r563880c, r584948b) are unused by this device, and the same git sparse-checkout the home tree uses takes the clang directory from 19 GB to 7.4 GB.

Do it before the first build. git sparse-checkout set rewrites every file it keeps, so the compiler gets a time newer than the finished build:

2026-10-23 17:09:39 prebuilts/clang/host/linux-x86/clang-r596125/bin/clang
2026-10-23 17:08:03 out/target/product/sargo/lineage_sargo-ota.zip
Enter fullscreen mode Exit fullscreen mode

The next m bacon redid almost the whole build:

status: ninja: 26m11.74s Build Succeeded: 181084 steps - 115.21/s
Package Complete: out/target/product/sargo/lineage-24.0-20261023-UNOFFICIAL-sargo.zip
Enter fullscreen mode Exit fullscreen mode

Step 7 — Pull the Tree Home

pull-tree.sh runs on the home machine and copies the whole tree with rsync -aHz --partial and zstd compression, keeping file times and hard links.

start 13:43:57
done: 266G
real    96m45.082s
end 15:20:42
Enter fullscreen mode Exit fullscreen mode

266 GB in 1 h 37 min. At $0.09/GB out of us-east-1 that is $23.94 of data transfer (arithmetic).

Step 8 — First Build at Home

The first local m bacon on the copied tree re-ran Soong analysis, then packaged the zip:

ninja: 1h44m06.32s Build Succeeded: 1 steps - 0.00/s
Package Complete: out/target/product/sargo/lineage-24.0-20261023-UNOFFICIAL-sargo.zip
ninja: 42m33.96s Build Succeeded: 246 steps - 0.10/s
real    149m46.233s
Enter fullscreen mode Exit fullscreen mode

Of the 16,869 inputs to the Soong graph, one was newer than the graph:

deps 16869 newer 1 missing 0
15:22:36 out/soong/soong.lineage_sargo.variables
Enter fullscreen mode Exit fullscreen mode

All 124 recorded environment values matched the home machine, so the path and user-name rules held; what rewrites soong.lineage_sargo.variables on the first local run is still open. The analysis took 1 h 44 min with up to 45 GB of swap in use, the same cost as a from-scratch analysis at home. After it, only 246 steps ran.

Step 9 — Orchestrate With Claude Code

Claude Code ran every step from one session on a laptop:

  • Sizing: quotas, on-demand prices, spot prices and spot placement scores through the AWS MCP server.
  • Launch: key pair import, a security group for the home IP, and RunInstances with the volume settings.
  • Watching: each long step runs in a tmux session on the instance; Claude Code waits on the log in a background job and is notified when the build exits.
  • Fixing: reads the failed command from out/siso_output, traces the type or symbol through the tree, commits to the forks and resumes the build.
  • Memory: the instance id, cost per hour and teardown rule go into Claude Code's project memory, so a later session finds a running instance.

The home machine is reached on the local network with its own credentials, used to stop the slow local build and free disk.

And What Does It Cost?

More than a home build. Rates from the AWS Pricing API, us-east-1:

Item Rate This run
c7a.48xlarge on-demand $9.85/h $48.61 for the 4 h 56 min from first sync to termination (arithmetic, a lower bound)
Data transfer out to home $0.09/GB $23.94 for 266 GB (arithmetic)
gp3 1000 GB at 16,000 IOPS and 1000 MB/s billed while the volume exists deleted with the instance

The copy home took 1 h 37 min after the last build. A smaller instance type holds the same volume: stop the instance, change its type, start it, and copy from that.

Compare and Contrast

i7-3770, home c7a.48xlarge, EC2
Threads 8 192
RAM 31 GB + 79 GB swap 384 GiB
Soong analysis about 1 h 45 min, swapping 🥇 included in the 19 min 17 s below
Build steps completed 130,617 in 7 h 49 min 🥇 156,615 in 19 min 17 s, analysis included
First build after the copy 2 h 30 min, analysis re-run —
Phone on USB 🥇 yes no
Cost electricity $9.85/h on-demand

So, Which One?

Both, in sequence. EC2 for the from-scratch build and for any change that re-runs Soong analysis; the home machine for edits that keep the build graph, flashing, and boot logs.

Teardown

Once the tree was home the instance was terminated, which deletes its 1000 GB volume, and the security group and key pair were deleted. Stopping instead keeps the volume, and the synced tree, for the next full rebuild, at the price of the volume.

Summary

The goal of this project was to move the long part of a LineageOS 24.0 build off a 32 GB home machine without losing incremental builds next to the phone. The key to the solution was building on EC2 at the home machine's absolute path with the home user name, so the finished tree is portable. The results were:

  • 🟢 Full sync of the LineageOS 24.0 tree in 15 minutes on a c7a.48xlarge.
  • 🟢 Soong analysis plus 156,615 build steps in 19 min 17 s.
  • 🟢 Four Android 17 port errors fixed between builds of 6 to 17 minutes each.
  • 🟢 Fork commits carry the same hash on the instance and on GitHub.
  • 🟢 The 266 GB tree copied home in 1 h 37 min, and the first local build produced the zip.
  • ⚠️ That first local build re-ran Soong analysis (1 h 44 min); the trigger is one rewritten input file whose cause is still open.
  • ⚠️ Trim prebuilts before the first build; trimming afterwards cost a 181,084-step rebuild.
  • ⚠️ The EC2 build ran without Soong's sandbox; Ubuntu 24.04 restricts the unprivileged user namespaces it needs.
  • ❌ A .bp or .mk change still re-runs Soong analysis at home.

Scope: one c7a.48xlarge on-demand instance in us-east-1a running Ubuntu 24.04, one home machine (i7-3770, 31 GB, Debian sid), LineageOS 24.0 for sargo built 2026-10-23. Home figures come from the 24.0 port notes and the local build log; EC2 figures from the build log on the instance and the build logs copied home with the tree. A device-tree edit rebuilt at home and a boot of this zip on the phone were not run for this article. Parts of the analysis and writing were done with AI assistance (Claude).

The strategy for splitting a LineageOS build between EC2 and a home machine was validated with an incremental step by step approach.

References

Top comments (0)