When you say it doesn't maintain them, do you mean it flattens them out? Or do you mean it doesn't have them at all?

I think the former?

In the above example in which out/Debug/md5sum_bin_host is a symlink to out/Debug/clang_x64/md5sum_bin on the builder, it'd wind up instead being a copy of the binary on the tester.

Can you link to builder that doesn't get symlinks into its resulting zip? Some quick local testing indicates that the vanilla zip command appropriately packs the targets of symlinks into the archive. We must be doing something quirky.

Here's the builder that seems to be giving us issues: https://build.chromium.org/p/chromium.fyi/builders/Android%20Builder%20%28dbg%29?numbuilds=200

And the associated test builder having issues when trying to run the component build: https://build.chromium.org/p/chromium.fyi/builders/Android%20VR%20Tests?numbuilds=200

Thanks for the links. I see what's going on now. On zipping, the symlink gets replaced by the file it points to. This can cause problems when it has run time dependencies on libraries it expects to be nearby. (The failure on the VR bot is due to the library being there but of the device's arch, not the host's)

I can confirm that simply passing the '--symlinks' arg to zip keeps the symlink, which can then be used to checksum files on the host from the host. I'm weary of adding that option however, since it seems fairly intuitive to me that it would simply replace the symlink at zip-time. And since a lot of bots use zip_build.py to transfer their binaries/tests, I'm also weary of breaking any bots out there that depend on the current behavior.

I wonder why this doesn't show up anywhere else. FWIU, all of our android tester bots get a host and device md5sum binary shipped to them.

(Doesn't devil's dependency system pin a bunch of host/device binaries, md5sum included, and fetch them from gs?)

Moving Infra>Client>Android -> Infra>Client>Chrome+OS=Android