On 11/27/24 01:06, Daniel P. Berrangé wrote:
On Tue, Nov 26, 2024 at 04:54:18PM -0600, Richard Henderson wrote:
On 11/26/24 11:52, Thomas Huth wrote:
I think we want to continue to maek failing downloads as test failures,
otherwise we'll never notice when an asset is not available from the
internet anymore (since SKIPs just get ignored).
I disagree. Download failures are not rare.
Failures of the test to download assets will be rare *if* we have the
CI runner cache fixed. We only need to successfully download each
asset once, and it should be cached forever with no expiry timeout.
So we have an initially bootstrapping problem once caching is fixed,
where download failures could impact us. Once the cache is primed,
we'll only be at risk of download failures when introducing new
asset URLs, so I think it is fair to say failures should be rare
*if* we get the caching fixed.
With regards,
Daniel
Beyond the QEMU CI, we should think about users trying to run tests, and
having the same kind of problems, but without having access to the magic
cache.
Regarding the assets download, why don't we mirror them somewhere
reliable instead of relying on third party storage?
It could be another github repo (and we push files in a "release", where
size is unlimited). In my experience, it's very reliable, and uptime is
99.999%.
The change is not complicated, and there is no risk to introduce
regressions, as the hashes will be exactly the same.
We can even do it on a best-effort basis, when one dep start failing.
Pierrick
