Skip to content

package-performance.md

Latest commit

 

History

History
201 lines (166 loc) · 11.7 KB

package-performance.md

File metadata and controls

201 lines (166 loc) · 11.7 KB
Raw

Package size and installation benchmark

Release adoption (2.5.0)

The owner selected the tested smaller Android profile for release. Starting at 2.5.0, release builds emit a compressed ARM64 APK plus a compressed universal compatibility APK under its existing filename. The updater prefers ARM64 on 64-bit ARM processes and retains universal compatibility for older clients and other architectures. Windows uses application cleanup and the tested .chm removal while retaining high CAB compression. Historical 2.4.1 measurements below remain the baseline; new release sizes are verified from the built assets.

Scope and provenance

The experiment keeps the UTHelper 2.4.1 application, dependencies, native plugins, signing checks, notifications and update boundaries intact. No production asset is replaced. APKs are debug-signed diagnostics; device IPAs are unsigned and require user signing. Neither is a production update.

The initial matrix runs use identical application source and dependency constraints; candidates change packaging options. Baseline and candidate are rebuilt rather than comparing different application releases. pubspec.lock, source revision, package hash, manifest and measurements accompany the diagnostic artifacts. The first macOS baseline failed on Bash's empty-array handling; the corrected iOS run replaces that measurement. The first Windows measurement stopped before installation because the environment-scoped WiX acceptance was unavailable; the corrected run receives the already accepted wix7 marker explicitly.

Android results

MiB means 1,048,576 bytes. Expanded archive bytes are not an Android storage report.

Variant APK MiB Expanded archive MiB Reduction from universal
Universal, modern native packaging 155.936 174.060 baseline
Universal, application cleanup 155.925 174.037 0.007%
ARM64, modern native packaging, cleanup 64.228 83.309 58.81%
ARM64, legacy native compression, cleanup 34.610 83.309 77.81%

The universal APK carries 48.004 MiB of ARM64 native libraries, 40.315 MiB of ARM32 libraries and 50.413 MiB of x86-64 libraries. In the ARM64 set alone, libapp.so is 14.376 MiB, libflutter.so is 11.045 MiB and libpython3.14.so is 5.559 MiB, before Python packages, stdlib and supporting native dependencies. Python has a real runtime cost, but shipping three architectures is the largest measured opportunity here. A stack rewrite is not required for the 59% reduction. Any future Dart-only rewrite needs a feature-parity prototype and the same notification, background-sync, authentication and update checks before its size or speed can be compared fairly; this experiment does not claim a measured win for an unbuilt replacement.

All four builds passed APK signature verification, exact application/version/build checks and the five notification/background receiver checks. This is native build and package evidence, not a device execution test. The initial experiment used host measurements at the owner's request. A subsequently connected Samsung tablet is covered in the Android device report, which separates authenticated runtime checks from isolated fresh-install measurements. Battery and low-end-phone acceptance remain outside these samples.

The package reader also performs three complete sequential reads with CRC checks. These are warm-cache host decompression measurements on separate hosted VMs, not Android installation times or a controlled cross-VM speed comparison. They must not decide startup or frame-performance acceptance.

ARM64-only packages omit ARM32 and x86-64 support. Keep universal as the default and as the compatibility option. Modern packaging is the initial balance choice: it cuts the download substantially while preserving direct native-library mapping. Legacy packaging is the smaller-download experiment; it extracts native libraries at install time and can increase installed storage and loading cost. It is not the default without device evidence.

iOS results

Variant IPA MiB Expanded archive MiB Entries
Baseline 35.0042 85.2821 3,471
Application cleanup 34.9916 85.2600 3,467

Both device IPAs passed the repository's native bundle identity, ARM64, version, build-number and unsigned-sideload-container verification. Cleanup saves about 13 KiB in this clean CI build; it is not a substantial iOS size optimization. The largest compressed components include Dart App.framework (about 7 MiB), Python packages (about 5.4 MiB), Python stdlib (about 5.2 MiB), and Flutter.framework (about 4.2 MiB). Removing required runtime libraries would change the application contract, so no such removal is proposed. iPhone installation/startup/frame timing is outside this Windows-host experiment.

Windows method

Baseline and --cleanup-app build on separate disposable Windows runners. Each runner creates private MSI candidates with the existing high CAB compression and medium CAB compression, then repeats real quiet installation/uninstallation three times per candidate. Every installed bundle is verified and runs the existing isolated single-instance/autostart/hidden-window/activation smoke matrix before uninstall. It never installs over a maintainer's application or uses their profile.

The first sample includes Windows Installer initialization; subsequent samples reuse OS/file caches. Compare repeat distributions, not a single best value. The EXE bootstrapper is not timed separately. Filesystem bytes and file count are measured after install. Build time and ZIP read time are not installer time.

Bundle CAB compression MSI MiB Median install s Median uninstall s Installed MiB / files
Baseline high 47.970 12.398 18.083 147.849 / 3,471
Baseline medium 50.912 12.174 17.724 147.849 / 3,471
Application cleanup high 47.990 10.710 16.190 147.827 / 3,468
Application cleanup medium 50.908 14.510 19.608 147.827 / 3,468
Application + help-file cleanup high 45.693 15.713 26.716 145.306 / 3,467
Application + help-file cleanup medium 48.603 13.427 21.154 145.306 / 3,467
Source-config confirmation high 45.681 13.585 21.438 145.306 / 3,467
Source-config confirmation medium 48.595 13.804 21.580 145.306 / 3,467

All 24 installation/uninstallation cycles and their native activation smoke checks passed. Application cleanup removes only about 23 KiB from a clean hosted build. Its faster high-compression installation on another VM is not evidence that removing those 23 KiB caused a speedup. Medium compression costs about 2.94 MiB more download and does not consistently improve installation. Keep high compression in production; do not claim an installer speed improvement from these samples. The help-file trial removes one PyWin32 compiled help (.chm) file, retaining all runtime modules. It saves 2.52 MiB installed and about 2.30 MiB in the high-compression MSI compared with application cleanup alone. The high-compression MSI is 4.75% smaller than baseline. Its slower install samples also came from another VM; the results do not establish a causal speed improvement or regression. Help-file cleanup is retained for its measured size reduction and passing native activation checks. The final source-config build repeated all six install/activation checks on Windows build 26100, an AMD EPYC 7763 host with four logical CPUs allocated to the runner. Its staged and authored configurations are semantically identical; the selected defaults reproduce the measured profile. Separate builds vary by small packaging overhead, so the high-compression result is about 45.68–45.69 MiB.

The machine-readable results retain all samples, artifact hashes, source commits and run links. Refresh them from the diagnostic artifacts after a new workflow run; keep the method and limits in this document in sync. Full package/lock/manifest reports remain in the short-lived workflow artifacts and the operator's private baseline directory.

Changes retained

  • Enable application cleanup through the supported Flet source configuration; all three platforms built successfully with it. Compilation remains enabled.
  • Remove only Windows package help files with the tested **.chm glob in [tool.flet.windows.cleanup]; keep all runtime dependencies.
  • Expose explicit Android architecture/legacy/output-directory options in the existing patched build script. The subsequently selected release defaults are described above; diagnostic modern profiles explicitly disable compression.
  • Keep Windows high CAB compression. Do not remove Pillow, lxml, SSL, Pydantic, Python runtime or notification components to manufacture a smaller result.
  • Do not change the production updater or publish benchmark artifacts as releases.

Reproduce

The benchmark workflow is manual and read-only with respect to repository contents, release assets and secrets. It accepts platform=android|ios|windows|all; Windows also requires the owner's existing WiX EULA acceptance marker. It publishes only short-lived diagnostic artifacts, never a GitHub release.

The workflow stages explicit cleanup settings in its disposable checkout with prepare_package_benchmark.py. Baseline disables application cleanup and retains help files; cleanup enables application cleanup only; the Windows documentation-only profile enables both. Original and staged authored TOML are saved as evidence. This avoids inheriting newly adopted defaults into the baseline. Flet 0.86.5 does not support --no-cleanup-app; source configuration is used instead. No generated Dart source is edited.

Local Android candidates use the same required two-pass notification patch path:

./scripts/build_android.ps1 -Architecture arm64-v8a -LegacyPackaging:$false -OutputDir build/apk-arm64-modern
./scripts/build_android.ps1 -Architecture arm64-v8a -LegacyPackaging -OutputDir build/apk-arm64-legacy

Measure an existing artifact without installing or rewriting it:

python scripts/measure_package.py build/apk-arm64-modern/UTHelper.apk --output build/arm64-report.json

Keep application/package compilation enabled. Cleanup removes redundant cache directories after producing adjacent runtime bytecode; never recursively delete all .pyc files. Do not edit generated Dart code or re-zip a production APK to implement these changes: build from the authored configuration and verify again.

The 2.5.0 release adoption coordinates the additional ARM64 asset with inventory, runtime selection, native ABI/signature verification and universal compatibility. No older published asset is replaced.

Supporting sources