Page MenuHomePhabricator

[Nix] Support building Android through gradle on Linux
ClosedPublic

Authored by jon on Jul 7 2022, 4:15 PM.
Tags
None
Referenced Files
F3357741: D4483.id14791.diff
Sun, Nov 24, 1:18 AM
Unknown Object (File)
Fri, Nov 22, 1:37 AM
Unknown Object (File)
Fri, Nov 22, 1:17 AM
Unknown Object (File)
Fri, Nov 22, 12:42 AM
Unknown Object (File)
Fri, Nov 22, 12:32 AM
Unknown Object (File)
Thu, Nov 21, 7:37 PM
Unknown Object (File)
Thu, Nov 21, 9:50 AM
Unknown Object (File)
Tue, Nov 12, 11:22 AM

Details

Summary

Pick some of the relevant work from
https://phab.comm.dev/D4433 which can be used to
build android

Test Plan

From a x86_64-linux machine:
Follow existing directions to download android sdk and ndk

nix develop
yarn cleaninstall # download node_modules
cd native/android && ./gradlew bundleRelease --no-daemon

Diff Detail

Repository
rCOMM Comm
Lint
Lint Not Applicable
Unit
Tests Not Applicable

Event Timeline

There are a very large number of changes, so older changes are hidden. Show Older Changes

Wasn't able to run this locally since my Nix setup is broken (beyond the zshrc stuff), but running on NixOS via Buildkite here: https://buildkite.com/comm/android-build-nix/builds/2#018217ad-a103-48e5-998e-f56e2007d39d

This revision is now accepted and ready to land.Jul 19 2022, 11:55 AM
atul requested changes to this revision.EditedJul 19 2022, 12:46 PM
nix develop 
cd native/android && ./gradlew bundleRelease --no-daemon

Could you specify what architectures/operating systems this was tested on?

I can't get it to work on

  • aarch64 macOS
  • x86 nixOS (emulated via qemu via docker on aarch64 Darwin host):
error: unable to load seccomp BPF program: Invalid argument
  • aarch64 nixOS (emulated via qemu via docker on aarch64 Darwin host):
error: flake 'git+file:///comm' does not provide attribute 'devShells.aarch64-linux.default', 'devShell.aarch64-linux', 'packages.aarch64-linux.default' or 'defaultPackage.aarch64-linux'
       Did you mean devShell?
This revision now requires changes to proceed.Jul 19 2022, 12:46 PM

Could you specify what architectures/operating systems this was tested on?

x86_64-linux NixOS, after doing the SDK + NDK installation through android studio (although sdkmanager would likely work as well)

I can't get it to work on

  • aarch64 macOS

This likely won't work because ANDROID_SDK_ROOT isn't pointing to anything, I would need to test this before setting it on both platforms

x86 nixOS (emulated via qemu via docker on aarch64 Darwin host):

error: unable to load seccomp BPF program: Invalid argument

This looks related to https://github.com/NixOS/nix/issues/5258, which is a docker-related issue, but the current workaround is to modify nix's behavior.

Harbormaster failed remote builds in B10694: Diff 14726!

This was intermittent network issue while fetching dependencies, restarting fixed it.

Was chatting with @atul and neither of us is really sure what the product impact here is. The team (outside of @jon) doesn't use NixOS, so the only benefits I can see here are:

  1. To improve @jon's personal custom workflow. Here, I wonder if we should be spending cycles supporting a custom niche use case, as opposed to having @jon use macOS primarily like the rest of the team
  2. To provide an alternative to the CI build that we currently do with Debian. Here, it's not really clear if the change will be that impactful

At a high-level, I think we should be prioritizing Nix work that has an immediate impact on dev velocity and dev experience across all devs on the team. Until the whole team is actively using Nix for the primary development workflow, I'd prefer if we could deprioritize any other Nix work.

To help make the prioritization super clear, here are some clarificiations:

  1. NixOS support should be deprioritized in favor of macOS Nix support
  2. CI Nix support should be deprioritized in favor of dev environment Nix support
  3. Nix as a replacement for Docker should be deprioritized in favor of Nix as a replacement for Homebrew and the dev environment instructions
nix/dev-shell.nix
93 ↗(On Diff #14726)

We should avoid single-bracket conditionals in bash. We spent way too much time trying to fix these to break it again here... see ENG-157

That said this is a small diff so not opposed to landing it. But unfortunately because it seems to only pertain to a custom workflow of @jon's, @atul's unable to test it...

Spun up a x86 nixOS machine on EC2 to give it a try, and got the following on nix develop:

[root@ip-172-31-26-187:~/comm]# nix develop --extra-experimental-features "flakes nix-command"
error: builder for '/nix/store/hc01f564s3wx40f6fnxi85p0jikjzqj7-arcanist-20220517.drv' failed with exit code 2;
       last 10 log lines:
       > g++ -c -fPIC -Wall -O3 -minline-all-stringops -o parser.yacc.o parser.yacc.cpp
       > parser.y:8:10: fatal error: node_names.hpp: No such file or directory
       >     8 | #include "node_names.hpp"
       >       |          ^~~~~~~~~~~~~~~~
       > compilation terminated.
       > make: *** [Makefile:66: parser.yacc.o] Error 1
       > make: *** Waiting for unfinished jobs....
       > Wrote C++ definition.
       > Wrote PHP definition.
       > make: Leaving directory '/build/source/support/xhpast'
       For full logs, run 'nix log /nix/store/hc01f564s3wx40f6fnxi85p0jikjzqj7-arcanist-20220517.drv'.
error: 1 dependencies of derivation '/nix/store/4r6habzv850108kc5p9g043mb2cmk0s8-nix-shell-env.drv' failed to build

Could definitely be messing something up here, anything else I should try?

atul requested changes to this revision.Jul 20 2022, 3:18 PM

That said this is a small diff so not opposed to landing it. But unfortunately because it seems to only pertain to a custom workflow of @jon's, @atul's unable to test it...

Quickly set up a nixOS x86 EC2 instance so I'm able to test it now.

The build didn't succeed for me, but as soon it does--whether that requires changes to the diff or something on my end--I'm happy to accept it and have it landed.

This revision now requires changes to proceed.Jul 20 2022, 3:18 PM

Could definitely be messing something up here, anything else I should try?

Interesting... I have the same exact drv, but able to build it fine:

[16:32:47] jon@jon-desktop /home/jon/comm/comm (jonringer/tunnelbroker|REBASE)
$ nix eval .#legacyPackages.x86_64-linux.arcanist.drvPath
"/nix/store/hc01f564s3wx40f6fnxi85p0jikjzqj7-arcanist-20220517.drv"
[16:32:49] jon@jon-desktop /home/jon/comm/comm (jonringer/tunnelbroker|REBASE)
$ nix-build /nix/store/hc01f564s3wx40f6fnxi85p0jikjzqj7-arcanist-20220517.drv
/nix/store/4hqv72jgh77vb51ax9h19mi08dsp33j5-arcanist-20220517

This is likely an issue of the arcanist build itself, in that the build isn't thread safe. I see that they use bison, which means that it's possible that they don't have a build rule for the outputs generated, so a later build step is executing before those files exist. In this case "node_names.hpp".

To attempt this, you can try nix-build /nix/store/hc01f564s3wx40f6fnxi85p0jikjzqj7-arcanist-20220517.drv --cores 1, to make the build single threaded

I did use these changes to do D4448.

I think it's still a move in the right directions, even though the main benefactor is NixOS (which I know is not a primary use case).

As for prioritization, this diff took very little time, but unblocked me for D4448

I did use these changes to do D4448.

I think it's still a move in the right directions, even though the main benefactor is NixOS (which I know is not a primary use case).

As for prioritization, this diff took very little time, but unblocked me for D4448

That's really confusing, since D4448 is landed but this one isn't! Does that mean that D4448 doesn't actually need this diff? Or does that mean that things are in a broken state (for macOS Android builds) until this diff is landed?

  1. If D4448 works fine for macOS Android builds without this diff, and this diff is just adding NixOS support, then I don't think it should've been prioritized. (Happy to still land it now that it's up.)
  2. On the other hand, if D4448 breaks macOS Android builds without this diff, it shouldn't've been landed before this diff was.

This is likely an issue of the arcanist build itself, in that the build isn't thread safe. I see that they use bison, which means that it's possible that they don't have a build rule for the outputs generated, so a later build step is executing before those files exist. In this case "node_names.hpp".

To attempt this, you can try nix-build /nix/store/hc01f564s3wx40f6fnxi85p0jikjzqj7-arcanist-20220517.drv --cores 1, to make the build single threaded

Is there a way to configure nix-build so that the user doesn't have to manually pass --cores 1 to it every time they build? One thing that worries me about Nix is that it often feels fragile when working with code that doesn't follow its "best practices". The reality is that lots of tools we use will probably not follow Nix "best practices", so it's critical that Nix has ways to configure things to handle those tools well.

Separately, it's a bit confusing to me why we need to be "building" Arcanist at all... in my case it's just a simple interpreted PHP script. Why can't we just use the PHP interpreter here? It feels like scope creep that we're trying to build Arcanist at all.

Does that mean that D4448 doesn't actually need this diff?

For FHS linux distros no. The existing docker CI gate was already handling the Android SDK stuff, so the diff was more or less to align with the CommonCpp being the include path for header files.

Or does that mean that things are in a broken state (for macOS Android builds) until this diff is landed?

They shouldn't be broken, the c++ related changes are reflected in both the cmake and header files. If there was successful workflow before, it will be successful after the diff. I was commenting on that I hadn't tested on mac.

Is there a way to configure nix-build so that the user doesn't have to manually pass --cores 1 to it every time they build?

In /etc/nix/nix.conf, you can state cores = 1. But builds will take a LONG time on average. I can just upstream the pinning in nixpkgs, and later update to the official cache. It's unfortunate that arcanist is in a largely unmaintained status though. https://secure.phabricator.com/D21853

Upstream PR: https://github.com/NixOS/nixpkgs/pull/182364

One thing that worries me about Nix is that it often feels fragile when working with code that doesn't follow its "best practices".

One of the reason's why I've had to make patches to protobuf, fbjni, and soon aws sdk. These usually aren't long term issues, but it's kind of sad how much software "barely works".

Separately, it's a bit confusing to me why we need to be "building" Arcanist at all... in my case it's just a simple interpreted PHP script. Why can't we just use the PHP interpreter here? It feels like scope creep that we're trying to build Arcanist at all.

That's how nix deals with interpreted scripts, it bundles the script and "nixifies" them. This enables the, "just add it to the dev shell, and it just works" workflow and avoids the "also install php 7.0, and place this script somewhere in your PATH" workflow.

Arcanist + phabricator is not the best example either, it's been officially unmaintained since July 2021, only receiving some community activity.

Separately, it's a bit confusing to me why we need to be "building" Arcanist at all... in my case it's just a simple interpreted PHP script.

Is that just an entrypoint? similar to how java usually exports a wrapper script? the arcanist code base is quite extensive: https://github.com/phacility/arcanist

It has about ~120k LoC (C++ and PHP)

Limit changes to being "good defaults" for all Linux distros

Is this ready for me to check out on the nixOS instance and test? Or should we mark it "Changes Planned" for now?

The Node build step takes a long time, so just let me know when it's ready

Rebase to avoid CI breaking changes

Accidentally squished tunnelbroker

@atul this is ready to test on any x86_64-linux machine + Nix

In D4483#131720, @jon wrote:

@atul this is ready to test on any x86_64-linux machine + Nix

Nice! Will check out the commit and give it a shot

Is there no way to use prebuilt binaries for Node? Would that be possible if we weren't limited to Node 16.13?

atul requested changes to this revision.EditedJul 21 2022, 3:45 PM

From a x86_64-linux machine

Would've been helpful to specify a distro/version since you mentioned that it needed to be FHS-compliant and whatnot.

(Ended up going with Debian which I assume should work)


Follow existing directions to download android sdk and ndk

Would it be possible to link those? I can't seem to find them anywhere

It's kind of tedious to set this up manually from the command-line, so I'm guessing it's something that would be done once by a developer who primarily uses a Linux machine? I thought we could make this happen automatically with Nix? (https://phab.comm.dev/D3640)


nix develop

Would've been good to specify:

nix develop --experimental-features "flakes nix-command"

I've looked at a few of these so picked up what I actually needed to run, but would generally be good to give these to reviewers up front.


cd native/android && ./gradlew bundleRelease --no-daemon

This command fails if yarn cleaninstall hasn't been run in the root of the repository first:

FAILURE: Build failed with an exception.

* Where:
Settings file '/home/admin/comm/native/android/settings.gradle' line: 3

* What went wrong:
A problem occurred evaluating settings 'Comm'.
> String index out of range: 0

* Try:
Run with --stacktrace option to get the stack trace. Run with --info or --debug option to get more log output. Run with --scan to get full insights.

* Get more help at https://help.gradle.org

BUILD FAILED in 4s

If someone is using Linux as their primary dev machine, it seems like they'd be best off using the existing: https://hub.docker.com/r/reactnativecommunity/react-native-android/?

This revision now requires changes to proceed.Jul 21 2022, 3:45 PM

Would've been good to specify:

This is listed in the nix directions to enable flakes: https://github.com/CommE2E/comm/blob/master/docs/nix_dev_env.md#nix-flakes

This command fails if yarn cleaninstall hasn't been run in the root of the repository first:

That's probably fair, but this is also true in the existing workflow. I'll update the test commands

ashoat requested changes to this revision.Jul 22 2022, 9:27 AM

That's how nix deals with interpreted scripts, it bundles the script and "nixifies" them. This enables the, "just add it to the dev shell, and it just works" workflow and avoids the "also install php 7.0, and place this script somewhere in your PATH" workflow.

I'd prefer to take the more supported workflow. Can we please move to installing Arcanist the same way we do in the dev environment instructions? We should avoid the build step, and in general we should avoid picking Nix best practices over a tried and true process. If the Arcanist install instructions say to install PHP 7 and interpret the PHP, then we should be configuring Nix to install PHP 7 and intrepret the PHP.

In D4483#131647, @jon wrote:

Is there a way to configure nix-build so that the user doesn't have to manually pass --cores 1 to it every time they build?

In /etc/nix/nix.conf, you can state cores = 1. But builds will take a LONG time on average. I can just upstream the pinning in nixpkgs, and later update to the official cache. It's unfortunate that arcanist is in a largely unmaintained status though. https://secure.phabricator.com/D21853

Upstream PR: https://github.com/NixOS/nixpkgs/pull/182364

Your upstream PR doesn't seem related to my request, which was:

Is there a way to configure nix-build so that the user doesn't have to manually pass --cores 1 to it every time they build?

Asking the user to edit /etc/nix/nix.conf to make our dev environment work goes against the whole point of this work, which is to make the dev environment work with a single, simple command.

In D4483#131646, @jon wrote:

Does that mean that D4448 doesn't actually need this diff?

For FHS linux distros no. The existing docker CI gate was already handling the Android SDK stuff, so the diff was more or less to align with the CommonCpp being the include path for header files.

Or does that mean that things are in a broken state (for macOS Android builds) until this diff is landed?

They shouldn't be broken, the c++ related changes are reflected in both the cmake and header files. If there was successful workflow before, it will be successful after the diff. I was commenting on that I hadn't tested on mac.

Unfortunately I'm still confused. Can you try answering these questions again? I still don't understand the relationship between D4448 and this diff, and how it's possible that D4448 required the work here and yet was landed before this diff. I think it would be helpful if you try to write out more... it seems like you might need several paragraphs to answer these questions.

I'd prefer to take the more supported workflow. Can we please move to installing Arcanist the same way we do in the dev environment instructions? We should avoid the build step, and in general we should avoid picking Nix best practices over a tried and true process.

This would be remedied largely by https://linear.app/comm/issue/ENG-1393/add-new-nix-binary-cache, no one would need to build packages if this was fulfilled.

If the Arcanist install instructions say to install PHP 7 and interpret the PHP, then we should be configuring Nix to install PHP 7 and intrepret the PHP.

Nixpkgs has removed php74 because it's end-of-life soonish (probably sooner than it should though). https://github.com/NixOS/nixpkgs/commit/e91811bbe107b7dad31075254344a3ec57cfa4de

What do you want me to do from here? I can:

  • Remove arcanist from dev-shell
  • package php74 locally, but that means that everyone will need to build once (until the cache is running). The error from atul was from trying to build a package.
  • Wait: I have a PR for nixpkgs to pin it to php8.0, https://github.com/NixOS/nixpkgs/pull/182364

Asking the user to edit /etc/nix/nix.conf to make our dev environment work goes against the whole point of this work, which is to make the dev environment work with a single, simple command.

It's already part of the installation instructions to avoid having to pass --experimental-features "nix-command flakes" all the time. However with cores, I would suggest not doing that, the comm-specific nix cache would make all of the local building largely irrelevant, they would just pull from the comm cache.

Unfortunately I'm still confused. Can you try answering these questions again?

I'm trying to slowly make the dev-shell more inclusive of all the workflows. This is just a small step in that direction. I've been able to build it locally on my NixOS with this diff and the NixOS support diff.

That's really confusing, since D4448 is landed but this one isn't! Does that mean that D4448 doesn't actually need this diff? Or does that mean that things are in a broken state (for macOS Android builds) until this diff is landed?

They are ordered this way in the stack because I needed them to be ordered this way for my nix-based workflow. The diff itself is fine in the existing workflow (e.g. download the sdk + ndk, install jdk manually). As part of the rebasing, I've just been prioritizing "what can go in, thus I'm less likely to cause errors when handling 20+ rebasing commits all the time".

If D4448 works fine for macOS Android builds without this diff, and this diff is just adding NixOS support, then I don't think it should've been prioritized. (Happy to still land it now that it's up.)

Maybe, but it's a major QoL for me. And the code changes took me very little time, and the workflow (with exception of the actual 3 lines of nixos-specific logic) can be extended to any linux distro.

On the other hand, if D4448 breaks macOS Android builds without this diff, it shouldn't've been landed before this diff was.

Agreed, but if there was a passing android build before, it should pass after the diff.

Is there a way to configure nix-build so that the user doesn't have to manually pass --cores 1 to it every time they build?

I mentioned --cores 1 because it lowers the non-determinism of the arcanist build significantly. Re-trying nix develop probably would have also worked. Yes, this is not a great experience, but the appropriate action would be to fix it on the arcanist side, as this is an issue which exists outside of nix and in arcanist.

One thing that worries me about Nix is that it often feels fragile when working with code that doesn't follow its "best practices".

Usually this isn't much of an issue if you just want to use nix for development tools as impure workflows are fine. But it's more of an issue if you want to package something using nix (e.g. nix build .#comm.tunnelbroker). Since the current work is just relating to development QoL this usually isn't much of an issue. Also for the software that is failing at best practices, i've been upstreaming fixes for them. A contributing factor to the current "use a docker dev environment" is likely stemming from the fact that some of these upstreams were broken to begin with, so you couldn't make use of the cmake installation workflow, but instead needed to directly import them as if they were a co-located code base.

The reality is that lots of tools we use will probably not follow Nix "best practices", so it's critical that Nix has ways to configure things to handle those tools well.

For development environments, this shouldn't be much of an issue.

Separately, it's a bit confusing to me why we need to be "building" Arcanist at all...

Since I changed the default php version from what's available in nixpkgs, nix will re-build the package locally. This can be avoided if we setup that nix cache.

in my case it's just a simple interpreted PHP script.

Arcanist is quite a large code base: https://github.com/phacility/arcanist

Why can't we just use the PHP interpreter here?

Because arcanist is more than just a php script

It feels like scope creep that we're trying to build Arcanist at all.

Said another way, nix "derivations" are a merkel tree of all inputs. I changed one of the inputs (e.g. php8.1 -> php8.0), so it needs to build the new derivation. Above I mentioned the three options moving forward.

You should request review again so this ends up in @ashoat's queue

@atul for the darwin issues, I created https://github.com/NixOS/nixos-homepage/issues/876 so nix can at least gracefully handle apple update nukes.

I ran into some issues around updating to 12.5, I was able to get it fixed by doing the following:

sudo rm /etc/{bashrc,zshrc}.backup-before-nix
sh <(curl -L https://nixos.org/nix/install)
ashoat requested changes to this revision.Jul 25 2022, 6:09 AM

High-level feedback – this diff is taking a lot of my time to engage on. We should move forward here ASAP. Before responding with a comment that might require a long reply from me, it might be best for @jon to sync with @varun or @atul to get their interpretation before responding.

In D4483#132027, @jon wrote:

I'd prefer to take the more supported workflow. Can we please move to installing Arcanist the same way we do in the dev environment instructions? We should avoid the build step, and in general we should avoid picking Nix best practices over a tried and true process.

This would be remedied largely by https://linear.app/comm/issue/ENG-1393/add-new-nix-binary-cache, no one would need to build packages if this was fulfilled.

The fact that there are ways to make your more complicated workflow work does not mean we should go with the more complicated workflow. We should go with the more simple, supported workflow, which (as far as I understand) is to interpret the PHP.

If the Arcanist install instructions say to install PHP 7 and interpret the PHP, then we should be configuring Nix to install PHP 7 and intrepret the PHP.

Nixpkgs has removed php74 because it's end-of-life soonish (probably sooner than it should though). https://github.com/NixOS/nixpkgs/commit/e91811bbe107b7dad31075254344a3ec57cfa4de

What do you want me to do from here? I can:

  • Remove arcanist from dev-shell
  • package php74 locally, but that means that everyone will need to build once (until the cache is running). The error from atul was from trying to build a package.
  • Wait: I have a PR for nixpkgs to pin it to php8.0, https://github.com/NixOS/nixpkgs/pull/182364

I had an "if" in my statement... is it indeed the case that Arcanist only works with PHP 7.4? If it works with PHP 8.0 then we can use PHP 8.0. I don't have some special interest in PHP 7.4.

Anyways, it seems like there is an option 4 here... just install PHP 8.0 in the Nix config, and use that to interpret Arcanist?

Asking the user to edit /etc/nix/nix.conf to make our dev environment work goes against the whole point of this work, which is to make the dev environment work with a single, simple command.

It's already part of the installation instructions to avoid having to pass --experimental-features "nix-command flakes" all the time. However with cores, I would suggest not doing that, the comm-specific nix cache would make all of the local building largely irrelevant, they would just pull from the comm cache.

As stated before, we should avoid the build step. It is a huge problem that it is does not work "out of the box" in Nix without --cores 1, and that there is no way to configure the build to work on a single core outside of a global config.

The real issue here is that we're trying to build Arcanist at all. As mentioned before, it's my understanding that it should be interpreted. We should follow "stock" simple installation instructions.

Unfortunately I'm still confused. Can you try answering these questions again?

I'm trying to slowly make the dev-shell more inclusive of all the workflows. This is just a small step in that direction. I've been able to build it locally on my NixOS with this diff and the NixOS support diff.

I really see building on NixOS as a non-goal and think we would be able to move faster on this work if you use the two macOS machines we've provided you for your development workflow while working on Comm. If you insist on using NixOS I won't object, but please keep in mind that I see it as slowing you down and don't see the NixOS diffs as moving us forward...

That's really confusing, since D4448 is landed but this one isn't! Does that mean that D4448 doesn't actually need this diff? Or does that mean that things are in a broken state (for macOS Android builds) until this diff is landed?

They are ordered this way in the stack because I needed them to be ordered this way for my nix-based workflow. The diff itself is fine in the existing workflow (e.g. download the sdk + ndk, install jdk manually). As part of the rebasing, I've just been prioritizing "what can go in, thus I'm less likely to cause errors when handling 20+ rebasing commits all the time".

Can you actually explain WHY you are ordering the diffs this way, rather than just telling me that they need to be ordered that way?

I mentioned --cores 1 because it lowers the non-determinism of the arcanist build significantly. Re-trying nix develop probably would have also worked. Yes, this is not a great experience, but the appropriate action would be to fix it on the arcanist side, as this is an issue which exists outside of nix and in arcanist.

What's "appropriate" really doesn't matter here... we should be moving forward with a simple, working approach, which is just to interpret Arcanist.

Usually this isn't much of an issue if you just want to use nix for development tools as impure workflows are fine. But it's more of an issue if you want to package something using nix (e.g. nix build .#comm.tunnelbroker). Since the current work is just relating to development QoL this usually isn't much of an issue. Also for the software that is failing at best practices, i've been upstreaming fixes for them. A contributing factor to the current "use a docker dev environment" is likely stemming from the fact that some of these upstreams were broken to begin with, so you couldn't make use of the cmake installation workflow, but instead needed to directly import them as if they were a co-located code base.

Looks like there is a high cost to packing something using Nix. We should probably continue deferring this work, and instead prioritize using Nix as a development tool. Regarding impure workflows, we should embrace those and focus and getting the dev environment working with nix develop rather trying to use Nix for everything and making everything perfectly pure and Nix-compatible.

Separately, it's a bit confusing to me why we need to be "building" Arcanist at all...

Since I changed the default php version from what's available in nixpkgs, nix will re-build the package locally. This can be avoided if we setup that nix cache.

Huh? The Nix cache is not relevant here... it appears to me that you made a very opinionated (and expensive) decision to build Arcanist here (a non-standard workflow), and I don't understand the justification at all.

in my case it's just a simple interpreted PHP script.

Arcanist is quite a large code base: https://github.com/phacility/arcanist

Your opinion of the size of the codebase seems hardly relevant. It works as an interpreted script on every dev's computer right now and nobody has had an issue. Improving the performance of Arcanist was never a goal of your work and is a non-issue on the team.

Why can't we just use the PHP interpreter here?

Because arcanist is more than just a php script

I don't understand. Are you denying that the Arcanist install instructions tell you to interpret Arcanist? It really doesn't matter if you think of it as a "php script" or not... the reality is that it appears to be executed as a PHP script according to the default installation instructions.

It feels like scope creep that we're trying to build Arcanist at all.

Said another way, nix "derivations" are a merkel tree of all inputs. I changed one of the inputs (e.g. php8.1 -> php8.0), so it needs to build the new derivation. Above I mentioned the three options moving forward.

It feels like you're avoiding addressing the critique here, and responding with misdirection about Nix builds. The details of what complicated Nix internals are leading to this build being complicated are entirely irrelevant, as far as I understand. We simply should not be doing the Nix build!

This revision now requires changes to proceed.Jul 25 2022, 6:09 AM

Anyways, it seems like there is an option 4 here... just install PHP 8.0 in the Nix config, and use that to interpret Arcanist?

That is the current state. It just happened to fail during a nix develop which triggered a build.

It is a huge problem that it is does not work "out of the box" in Nix without --cores 1

It can work. It worked with me and I have --cores 24 as my default. It's just arcanist doesn't have a strict build DAG which enforces that generated parser files are created before some steps which need them.

Huh? The Nix cache is not relevant here... it appears to me that you made a very opinionated (and expensive) decision to build Arcanist here (a non-standard workflow), and I don't understand the justification at all.

https://github.com/NixOS/nixpkgs/pull/182364 was merged, we should just be able to update the flake

Looks like there is a high cost to packing something using Nix. We should probably continue deferring this work, and instead prioritize using Nix as a development tool. Regarding impure workflows, we should embrace those and focus and getting the dev environment working with nix develop rather trying to use Nix for everything and making everything perfectly pure and Nix-compatible.

Agreed.

@ashoat I created https://phab.comm.dev/D4646 to address arcanist issue. It no longer needs to be built by us. It's available with php80 from nixpkgs.

ashoat requested changes to this revision.Jul 27 2022, 6:20 AM

Let's follow up on the Arcanist issue in D4646. As for this diff, I'll repeat my feedback from earlier:

Was chatting with @atul and neither of us is really sure what the product impact here is. The team (outside of @jon) doesn't use NixOS, so the only benefits I can see here are:

  1. To improve @jon's personal custom workflow. Here, I wonder if we should be spending cycles supporting a custom niche use case, as opposed to having @jon use macOS primarily like the rest of the team
  2. To provide an alternative to the CI build that we currently do with Debian. Here, it's not really clear if the change will be that impactful

At a high-level, I think we should be prioritizing Nix work that has an immediate impact on dev velocity and dev experience across all devs on the team. Until the whole team is actively using Nix for the primary development workflow, I'd prefer if we could deprioritize any other Nix work.

To help make the prioritization super clear, here are some clarificiations:

  1. NixOS support should be deprioritized in favor of macOS Nix support
  2. CI Nix support should be deprioritized in favor of dev environment Nix support
  3. Nix as a replacement for Docker should be deprioritized in favor of Nix as a replacement for Homebrew and the dev environment instructions
This revision now requires changes to proceed.Jul 27 2022, 6:20 AM

Talked to @jon offline – we're going to unblock this as soon as he provides detailed step-by-step instructions in the Test Plan for how to test this using macOS

@ashoat this is very similar to D4605, but changes ANDROID_SDK_HOME to the default for linux. Since we don't count linux as a tier 1 development platform, do you mind if we consider this done and move on?

I think considering linux "there but unsupported" will help with getting a lot of the existing diffs in without my having to do a fair amount of git commit refactoring

@atul you're still a blocking reviewer. Since it worked on darwin, are you okay with rubber stamping this for now?

This revision is now accepted and ready to land.Jul 28 2022, 8:42 PM