[3.4] mvcc: reduce count-only range overhead

[fuweid]

Backports

• mvcc: reduce count-only range overhead #11771 (Only including the code change and test)
• etcdctl: support query count only of specified prefix #11743

The #11771 's test case needs #11743 change. So I backport two prs in one pr.

Please read https://github.com/etcd-io/etcd/blob/main/CONTRIBUTING.md#contribution-flow.

---

Related to #15072

[serathius]

We don't backport features.

[fuweid]

@serathius Thanks for the comment.

Before I file #15072 pr, I checked the contributor document about the backport policy.

The commits should be restricted to bug fixes and security patches.

Except that it introduces new feature --count-only into the etcdctl, the CountRevisions is new internal interface to prevent unnecessary memory allocation. It won't change the existing interface to end-user. So from the point of view of interface, could the enhancement be considered to be bugfix? :) Thanks

And I'm also trying to backport #12933 and wondering how to handle this case.

We are running etcd-3.4 (upgraded from 3.3) and it provides concurrent read transaction feature.
But with the high read-tx ratio, the etcd-3.4 needs more memory than etcd-3.3. It is easy to run into memory spike.
It seems that it is kind of memory regression. If the following release has a patch to fix the memory issue and the patch is very suitable to older release, is it possible to consider the patch as bugfix?

Look forward to the release managers' comments. Thanks!

[ahrtr]

In principle, we do not backport features. But we should always make judgement based on benefit and risk.

The etcdctl CLI flag --count-only is new to 3.4, but clientv3.WithCountOnly() is already in release-3.4 there for a long time. It's an issue to me that previous implementation of this feature (CountOnly) in 3.4 isn't complete. So I am not against to make it complete as long as it's a simple & minor change.

The simple improvement in (ti *treeIndex) CountRevisions can definitely reduce the memory usage when users perform only Count-Only operation, because we don't need to cache all K/V pairs any more. Again, it's a simple & minor change. So I am not against adding it in release-3.4. But @fuweid please backport the related unit test case (e.g. TestIndexRevision) or add a new unit test as well.

[fuweid]

Thanks @ahrtr for the comment! Updated the TestIndexRevision ut and it is ready to review. PTAL.

[ahrtr]

LGTM

Thank you @fuweid

Please also add a changelog item for 3.4.24.

[fuweid]

Please also add a changelog item for 3.4.24.

Will do. Thanks!

[fuweid]

ping @serathius @ptabor ~ is it acceptable?