You are viewing a plain text version of this content. The canonical link for it is here.
Posted to pr@cassandra.apache.org by ptnapoleon <gi...@git.apache.org> on 2017/11/23 11:52:13 UTC
[GitHub] cassandra-dtest issue #13: Install ccm from github branch
Github user ptnapoleon commented on the issue:
https://github.com/apache/cassandra-dtest/pull/13
So, I am -0.5 on this. We used to do this. The problem is that ccm has other consumers than dtest, and worse, is both an app and a library. As a result, sometimes changes go into ccm which break dtests, despite our best efforts at testing ccm changes. When we switched to versioning ccm, and manually pulling in ccm versions, it made development in both systems much easier.
For the scope of this project, we obviously only care about difficulty in developing on dtest. Anecdotally, people vastly preferred manual changes to the ccm dependency, as that's much more bisectable when debugging failures. The recent upgrade failures on C* trunk are due to a ccm change. When debugging that, I was able to bisect it to precisely the commit that bumped the ccm version, which let me easily find the ccm commit that caused the issue. Then I released ccm 3.1.1 with the fix.
When we used to pull in ccm automatically, bisecting issues caused by ccm didn't work, and one would have to tie in the time at which CI jobs ran to the time at which ccm commits were made. It's much more unpleasant. I do know we pull in the python driver from a branch. However, that branch isn't pushed to with every new commit to the new driver. Rather, commits are made there in a batch when a new, tested version of the driver is released, or very explicitly go in tied with new C* features (think a cqlsh feature going in alongside a driver feature). The people developing the drivers have far more resources to guarantee the correctness of the driver than I have to do the same with ccm.
Now, I know that's a lot of negatives up-front, sorry about that. Possible points in favor that I am aware of are:
1. People don't test dtest when bumping the ccm version anyway, so why bother with the formality
2. We can increase the speed at which ccm fixes are used in dtest CI
3. It is too inconvenient to get committers to push changes to dtest deps.
I think (1) is throwing the baby out with the bathwater, and for (2), is a handful of hours to run CI such a huge deal? It is not like the OSS C* project relies on dtest fixes having a very fast SLA. (3) is definitely true.
On the ccm side, how would you want me, effectively the only maintainer right now, to manage this branch? Either it gets every commit, in which case let's just pull from master, or it gets batched updates which are equivalent to releases. In the latter option, that's just pushing the burden of testing dtest to me, rather than the dtest developers (yes that also includes me, I know). It's also identical to just pulling in releases on the dtest side, except for how those releases are pulled in. If you want a branch that selectively pulls in commits just for dtest, that will diverge from master, my answer as the ccm maintainer is no, I will not maintain a fork.
Again, just -0.5 on this. Frankly, I don't develop on dtest much, so if this is easier for you and the other people who do, I totally okay with supporting the branch on the ccm side if you let me know how you want that branch maintained.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: pr-unsubscribe@cassandra.apache.org
For additional commands, e-mail: pr-help@cassandra.apache.org