Please provide a binary for aarch64 + deb10

[tarleb]

It seems that the list of official binaries does currently not include a build for aarch64 on deb10. Adding this build would allow us to publish a new Docker image with GHC 9.6 and cabal-install 3.10.

See also: haskell/docker-haskell#102

[Mikolaj]

Hi! The cabal team is not able to produce these binaries, but @hasufell over at ghcup has built them (if these are what you need): haskell/ghcup-metadata@b703c55

[Mikolaj]

BTW, is a GHC binary for exactly this platform produced by gitlab CI? If so, could you point me to a gitlab pipeline doing exactly that?

[hasufell]

The cabal team is not able to produce these binaries

Why?

[tarleb]

Official Docker images cannot contain unofficial binaries afaik, so we are somewhat dependent on the cabal project to provide those. How can we help to produce those, and where would we start?

[hasufell]

It appears this was a commit by @chreekat removing aarch64 linux: 0ea1001

The rationale was

GHCUp fails to run on aarch64-linux-deb10 due to a glibc mismatch.

aarch64-linux-deb11 would work, but there is no GHC release for that
version yet.

At the time of the commit, this was already fixed in GHCup. See the changelog: https://github.com/haskell/ghcup-hs/blob/master/CHANGELOG.md#01191----2023-2-19

And the ticket: haskell/ghcup-hs#762

Further, the second part of the rationale doesn't make sense to me. You should be able to install the deb10 GHC on deb11.

It seems I wasn't consulted on the PR either.

[Kleidukos]

Looks like the path to resolution is clearer now.

[chreekat]

It seems I wasn't consulted on the PR either.

My mistake, that was an oversight. I was just trying to get CI to green and forgot to double back on some of the issues I ran into.

I'm surprised to hear it was already fixed in GHCUp, though. The CI script just basically runs curl https://get-ghcup.haskell.org | sh — does it need to do something different?

[Kleidukos]

@hasufell @chreekat Any chance we could sync on this btw? CI is highly critical for cabal and I need to make sure that institutional knowledge gets written and understood by other maintainers.

[angerman]

Can I bring up again, that we should have a mailinglist for these kinds of issues?

Currently you have to sift through various GitHub notifications, potentially some IRC/Matrix/Telegram chats... we should really have a mailing list with the interested CI related parties involved to keep everyone up to date what's happening across the whole GitHub:haskell ecosystem.

[Kleidukos]

Yes there is considerable overhead.

[hasufell]

I'm surprised to hear it was already fixed in GHCUp, though. The CI script just basically runs curl https://get-ghcup.haskell.org | sh — does it need to do something different?

I'm surprised too.

The bootstrap script was updated in time: haskell/ghcup-hs@ea6c8d3

The ghcup release CI shows that deb10 was used to build: https://github.com/haskell/ghcup-hs/actions/runs/4254450877/jobs/7400717833

GHCup knows only Tier 1 platforms. So everything like that is a bug that I'll fix next weekend.

[chreekat]

Maybe I should run an apt-get update or something?

I re-enabled aarch64-deb10 and aarch64-deb11 jobs, and sure enough the same thing happens:

https://gitlab.haskell.org/haskell/cabal/-/jobs/1404272

[chreekat]

deb11 seems to be faring better — it's at least compiling Cabal now [edit: it has completed successfully]. https://gitlab.haskell.org/haskell/cabal/-/jobs/1404273.

I hadn't considered that "there is no GHC release for aarch64-deb11" wouldn't be a blocker.

[chreekat]

Doing an update && upgrade with apt-get doesn't help.

Installing the system requirements doesn't help, either (although it does install missing packages...)

Unrelated: in Debian 11, installing the listed system requirements fails with E: Unable to locate package libffi6

[chreekat]

Can I bring up again, that we should have a mailinglist for these kinds of issues?

Almost missed this message. I will see about getting a mailing list created. (There is an existing ghc-devops group, but its charter was explicitly "only ghc", which is not what we want.)