investigate if we get server side sharded list and watch out of the box with k8s 1.36 #4459

Activity

1. neolit123

   self-assigned this

   on Oct 12, 2026
2. neolit123

   added

   kind/choreCategorizes issue or PR as related to maintenance and other usually non-code changes.Categorizes issue or PR as related to maintenance and other usually non-code changes.

   on Oct 12, 2026
3. kcp-ci-bot

   added this to kcpon Oct 12, 2026
4. neolit123 commented on Oct 13, 2026

   

   neolit123

   on Oct 13, 2026

   Last edited by neolit123

   ContributorAuthor

   More actions

   ran a test script to verify various cases:

   ./kcp-verify-list-watch-sharding.sh
   
   building kcp...
   starting kcp with ShardedListAndWatch=true...
   creating test data...
   
   test case 1: list configmaps in one workspace; the two halves must add up to the full list and return shardInfo
     full=21 lower=10 upper=11 watch(lower)=10 shardInfo=shardRange(object.metadata.uid, '0x0000000000000000', '0x8000000000000000')
     -> PASS
   
   test case 2: watch configmaps in one workspace; only newly created objects in the lower half must arrive
     received=[cm-101 cm-102 cm-106 cm-113 cm-119]
     expected=[cm-101 cm-102 cm-106 cm-113 cm-119]
     -> PASS
   
   test case 3: list and watch configmaps across all workspaces (wildcard /clusters/*)
     full=50 lower=20 upper=30 watch(lower)=20 shardInfo=shardRange(object.metadata.uid, '0x0000000000000000', '0x8000000000000000')
     -> PASS
   
   test case 4: list and watch a CRD-backed resource in a child workspace
     full=20 lower=9 upper=11 watch(lower)=9 shardInfo=<none>
     -> FAIL
   
   test case 5: list and watch through the APIExport virtual workspace (tenancy.kcp.io workspacetypes)
     full=5 lower=4 upper=1 watch(lower)=4 shardInfo=<none>
     -> FAIL
   
   test case 6: shard by namespace across all workspaces; works, but every workspace's 'default' namespace lands in the same half
     full=50 lower=42 upper=8 watch(lower)=42 shardInfo=shardRange(object.metadata.namespace, '0x0000000000000000', '0x8000000000000000')
     -> PASS
   
   restarting kcp with ShardedListAndWatch=false...
   
   test case 7: with the feature gate off, the selector must be ignored and no shardInfo returned
     full=41 lower=41 shardInfo=<none>
     -> PASS
   

   the 4 and 5 cases are with CRDs and they fail. upstream needs this fix, so i submitted a PR; don't know if it can land in .38:

   • metav1: expose ShardInfo getter and setter on UnstructuredList kubernetes/kubernetes#142975
5. neolit123 commented on Oct 14, 2026

   

   neolit123

   on Oct 14, 2026

   ContributorAuthor

   More actions

   update

   • metav1: expose ShardInfo getter and setter on UnstructuredList kubernetes/kubernetes#142975

   ^ was approved and will merge for 1.38.
   the feature is maybe graduating to beta in 1.38 too.

   summary: all the above cases will work in 1.38, but 4 and 5 (which is effectively CRD support) don't work for < 1.38.
   it's a bug, but backporting fixes for alpha features is generally not accepted unless it's very blocking and core types do support the sharding, so it kind of is not very blocking. so this is borderline backportable to older k8s but i would prefer to not propose it, for now.
6. neolit123

   closed this as completedon Oct 14, 2026
7. github-project-automation

   moved this from New to Done in kcpon Oct 14, 2026