Move CAS deprecation to TransportShardBulkAction #42641
Conversation
dnhatn
added
>non-issue
:Distributed Indexing/Distributed
|
Pinging @elastic/es-distributed |
|
@ywelsch This is ready. Please have a look. Thank you! |
|
This change is not water tight. An update request can still fail if the primary shard which resolves the update request gets relocated before it executes the request while replicas are still on old nodes. |
only in case of a single update request (TransportUpdateAction), not an update that's done as part of the bulk APIs (TransportShardBulkAction), right? |
|
How about moving the deprecation logging that's currently done by |
Yes, it would be a good option. |
I made changes for this. However, there is still a case where we issue an unwanted deprecation warning for a single update request. If the cluster has an old node when TransportUpdateAction processes the request but that old node left when TransportShardBulkAction executes it. How about moving this deprecation to the rest layer? |
I think that's fine, this should be sufficient for our REST tests. It does not need to be perfect, but has to pass our tests.
This means that requests from the transport client won't be logged, and also internal requests which mistakenly might use versioning would not be logged either. I prefer the previously proposed solution. |
dnhatn
changed the title
|
@ywelsch This is ready again. Can you please take a look? Thank you! |
|
@elasticmachine test this please |
|
@ywelsch Thanks for reviewing! |
Today CrudIT expects CAS deprecation warnings from DocWriteRequest#validate which is executed locally on the client side before sending the request. This suite is currently failing after we moved the CAS deprecation check to TransportShardBulkAction (in #42641). We need to adjust this suite to verify the CAS deprecation warnings from the responses. Closes #42881
After we moved CAS deprecation check to TransportShardBulkAction (#42641), a mixed cluster consisting of 6.7 and 6.8.1 nodes can issue two deprecation warnings for versioning write requests. The way we deduplicate response headers does not work for warning headers from different versions. This commit skips CAS deprecation tests until all nodes on 6.8.1+.
If the primary on 6.6+ but replicas on older versions, we can not use CAS using ifSeqNo since it requires all nodes on 6.6+. In this case, we have to fall back to use CAS with _version and should not issue a deprecation warning log during validation. This change moves the CAS deprecation to TransportShardBulkAction and only issues deprecation when the cluster is ready for ifSeqNo.
Closes #42796
Relates #42596