Skip to content

STANDALONE.md

Latest commit

 

History

History
81 lines (66 loc) · 4.04 KB

STANDALONE.md

File metadata and controls

81 lines (66 loc) · 4.04 KB
Raw

Can AutoAB ship as a standalone app?

Short answer: not without real work, and the obstacle is R, not Python. This records what was actually tested, so the question does not have to be re-investigated from scratch.

What the pipeline needs at runtime

Component Bundling difficulty
IgBLAST binaries Trivial, already bundled via extraResources
Reference databases Trivial, already in the repository
Python + changeo/presto/pandas/… Easy, see below
R + alakazam/shazam/ape Hard, see below

Python: solved problem

python-build-standalone publishes relocatable CPython builds for macOS on both architectures (cpython-3.11.x-aarch64-apple-darwin-install_only.tar.gz and the x86_64 equivalent, current as of release 20260728). Unpack it into the app bundle, pip install -r backend/requirements.txt into it, point AUTOAB_PYTHON at it. Nothing is patched, so no code-signature issues arise. Roughly 250 MB.

R: tested, and it is the sticking point

R installs to a fixed location and bakes that path into its own files. A relocation attempt was carried out and got most of the way:

  1. bin/R is a shell script with R_HOME_DIR=/Library/Frameworks/... hardcoded. Patchable with sed. ✅
  2. bin/Rscript is a Mach-O binary containing the absolute path as a string. Not patchable in place, but avoidable by invoking bin/R -f script.R instead. ✅
  3. The shared libraries carry absolute install names. Rewriting them to @loader_path with install_name_tool works. Afterwards otool -L showed zero absolute references, and base R ran correctly from an arbitrary directory, reporting the relocated R_HOME and computing normally. ✅
  4. Apple Silicon kills modified binaries. install_name_tool invalidates the code signature, and the process is then terminated with SIGKILL (observed: exit 137, no output at all). Every touched Mach-O must be re-signed ad hoc with codesign --force --sign -. ⚠️
  5. Doing this across the package tree is where it broke. The 109 required packages contain 107 .so files. After rewriting their references and re-signing, R segfaulted (invalid permissions). The naive @rpath + add_rpath approach is not sufficient; getting this right needs a more careful, tested bundling script. ❌

Size, measured: base R is ~48 MB, the 109 required packages are ~459 MB (alakazam pulls in a substantial Bioconductor subtree: Biostrings, GenomicRanges, SummarizedExperiment and friends). So an R bundle is roughly 500 MB, and a fully standalone app would land near 1 GB.

The pragmatic answer

conda env create -f backend/environment.yml installs Python, R, all R packages, Change-O, pRESTO and IgBLAST in one command. That is the normal distribution mechanism for bioinformatics tooling, it is cross-platform, and it sidesteps every problem above.

Note that this file previously listed bioconductor-alakazam and bioconductor-shazam, which do not exist on any channel. The correct package names are r-alakazam and r-shazam. The environment could never have been created as specified; it has been corrected.

If a true standalone is wanted later

In rough order of effort:

  1. Bundle Python only (easy) and keep conda or a single install.packages() line for R. Removes most of the setup burden for little risk.
  2. Add a first-run bootstrap: detect missing dependencies on launch and install them, rather than failing with a blank backend.
  3. Full bundling: relocate R with a proper, tested script that rewrites install names and re-signs every binary, then sign and notarise the whole app with an Apple Developer ID. Expect days rather than hours, and expect it to break again on the next R version.
  4. Remove R altogether by porting calculateDistribution.R, the tree-building scripts and the visualisation to Python. This is the only option that makes the app genuinely self-contained, but it changes scientific code and would need its results re-validated against the current ones.