Hello,
And just to iterate on this, just look at the code duplication (double the maintenance for an async and sync with eventlet code path).
We’d need to duplicate oslo.service, oslo.messaging and all other common libraries being called in a asyncio code path, and not only
that we’d need to consolidate everyone on a Pecan (REST API library) replacement that would support asyncio (a new oslo library?) and
just imagine that migration and consolidation effort for multiple projects.
Even then we’d also need to port every single project to using it without regressions in performance.
It’s a crazy massive effort, but I fully agree that eventlet should have been gone long ago.
Best regards
Tobias
On 21 Jul 2023, at 16:16, smooney@redhat.com wrote:
On
Fri, 2023-07-21 at 15:40 +0200, Thomas Goirand wrote:
On 7/21/23 13:32, smooney@redhat.com wrote:
On Fri, 2023-07-21 at 11:59 +0200, Radosław Piliszek wrote:
On Fri, 21 Jul 2023 at 09:03, Thomas Goirand <zigo@debian.org> wrote:
On 7/15/23 16:27, Tobias Urdin wrote:
The effort is mostly now to get/find a supported third-party Python library that does not
use asyncio for usage in a oslo.messaging driver.
Sorry to ask, but what's the problem with asyncio? Do we want something
*not* async? If yes, why?
Large swaths of OpenStack code are not ready for async.
We want to take baby steps and get to some better place with tangible benefits.
that not entirly correct.
much of openstack is already async but we made a decision ot use eventlet as our async model many many
years ago after we ported form twisted.
The problem is you cannot really mix eventlet code and asyncio safely in the python process.
That means you cannot have the async nats lib used in nova,swift,cinder,neutron, glance and progrably other
projects today.
i proposed one option a long time ago which is to take a privsep like approch where we spawn the oslo messaging
driver
into a sperate process and comunicate with it over a unix socket.
IRRC, back in 2014, at the Hongkong summit, there was a large consensus
that Eventlet was a wrong choice. I can still remember Jay Pipes
shouting "die Eventlet, die...". Fast forward nearly 10 years later,
nothing has moved in that direction, and we're still stuck with that
horrible hack. Why don't we just get rid of Eventlet and switch to
asyncio instead? That'd be a *HUGE* improvement in terms of ... well
many things ... less bugs, more maintainability, less breakage on each
minor Python release, etc. That'd be a lot of work, I'm sure... At the
same time, everyone would be happy to finally get rid of it.
we
have consider that in the past and the reality is it would be a massive rewrite
we
considerd doing this when we were dicussiong moving to python3 and at various points since
the
problem is two fold. 1 we dont have contibutors able to work on this currently. 2 we dont have
a
good solution for how to do this incrementally.
in
nova at least we did breicly consider tyring to add asyncio console-script entries points
so
that we couuld have an eventlet version and non eventlet version of each service to act as a bridge
its
somethign we could discuss again but makign such a change without introducing bugs or performance regressions
would
require a lot of careful engenieing work.
i
would love to regarin the ablity ot run nova under a debugger and drop the dep on eventlet but i dont want
to
do that at the expense of stablity or perfromace for our customer.
some
services liek the api and schduler would be relitivly simple to port. the compute agent and conductor on the other
hand
rely on the implict acync behavior to fucntion at all. nova-comptue with libvirt as at lest 3 while true loops
to
compunticate with rabbit, libvirt and privsep to name just one of the the issues tha towuld need to be resolved.
Cheers,
Thomas Goirand (zigo)