[debian-mysql] MariaDB upgrades for Debian/Ubuntu pending: request for review and contribtions

[Robie Basak]

On Mon, Nov 22, 2021 at 09:31:11AM -0800, Otto Kekäläinen wrote:
> There is an upstream issue where Marko explains more about what is
> going on: https://jira.mariadb.org/browse/MDEV-26569

I don't agree with your statement there "There is nothing to be fixed in
MariaDB for this". Based on comments there, MariaDB can already avoid
the runtime requirement simply with a configuration flag. So why is the
default for that configuration flag tied to the kernel availability of
the feature *at build time*? In most deployments, the build environment
is not the runtime environment. It is therefore wrong for upstream to
make this assumption. MariaDB should detect if the kernel supports
io_uring and adjust itself accordingly.

I'm told that liburing is already designed to be able to cope seamlessly
with this kind of situation. There's no reason for it to be designed
this way only for MariaDB to then trip itself up on it.

You could insist that there's some minimal kernel featureset that must
be supported for builds to work at all as benchmarked by the build
environment, and this wouldn't necessarily be wrong. However it would be
unnecessarily unhelpful for users who bump into this situation given
that runtime support to function without the feature already exists.

> I don't think this issue is grave. Let's just wait for Launchpad
> kernels to catch up and then it should work.

I understand that there's a liburing2 transition underway in Ubuntu at
the moment which may get blocked on this, and especially for an LTS
release it's probably good to test testing for MariaDB 10.6 underway
soon, rather than waiting for an *estimate* of next year. In particular
bumping the kernel version used for builds can (and, I'm told, has)
cause issues elsewhere, so we might find ourselves blocked on this
longer than expected. So maybe it's better to get Ubuntu building
sooner, with a delta if necessary.

Setting -DCMAKE_DISABLE_FIND_PACKAGE_URING=1 would work for now, and
wouldn't be a feature regression as MariaDB has never had aio_uring
support in Ubuntu previously. Or the test runner could be adjusted to
set innodb_use_native_aio=OFF, so that users could still benefit from
the feature. We could do this in an Ubuntu delta, or use dpkg-vendor in
Debian to keep a unified source.

Alternatively, the tests could (presumably temporarily) be moved
exclusively to autopkgtest. From a quick, unconfirmed glance, I think
the autopkgtest environment in Ubuntu has a recent enough kernel. Then
Ubuntu wouldn't regress in aio_uring test coverage, either.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 819 bytes
Desc: not available
URL: <http://alioth-lists.debian.net/pipermail/pkg-mysql-maint/attachments/20211122/dde7e92c/attachment.sig>