[
https://issues.apache.org/jira/browse/CASSANDRA-19948?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17885495#comment-17885495
]
Josh McKenzie commented on CASSANDRA-19948:
-------------------------------------------
>As long as the CQL client sees a write is successful, the CDC consumer will
>definitely see that write too.
Will it? I don't think this is true:
* Given nodes A, B, C
* CDC enabled on B and C (QUORUM)
* Write hits node A and B, fails on C (CL=QUORUM success)
* Ack to client that write is successful
* B dies
Now you have data on node A that is not available to CDC with a write at QUORUM
that matched your CL request. Anti-entropy efforts _should_ resolve this, but
there's no guarantees here.
Or put more formally: if you non-equal sets of overlapping quorums, loss of the
overlapping members leaves you with inconsistent views of the data between the
quorums.
The original requirements I was working with on CDC were "at least once
delivery, period, full-stop". Which necessitates both a: using the same file
for CDC log as WAL log (split f-sync == nightmare fuel for CDC reconciliation),
and b: having your # CDC logs == your # replicas so you never have a replica
write that doesn't make it to a CDC log.
> Changing cdc table property can cause schema disagreement
> ---------------------------------------------------------
>
> Key: CASSANDRA-19948
> URL: https://issues.apache.org/jira/browse/CASSANDRA-19948
> Project: Cassandra
> Issue Type: Bug
> Components: Cluster/Schema
> Reporter: Bowen Song
> Priority: Normal
> Fix For: 4.1.x, 5.0.x, 5.x
>
> Attachments: 4.1.1.txt, 4.1.6.txt, 5.0.0-corrected.txt,
> cdc_schema_disagreement.sh
>
> Time Spent: 10m
> Remaining Estimate: 0h
>
> In the cassandra.yaml file, there is a parameter named "cdc_enabled" which
> allows CDC to be enabled or disabled on each individual nodes.
> It has been found that it can cause schema disagreement or discrepancy when
> an "ALTER TABLE ... WITH cdc=..." statement is ran against a node which has
> "cdc_enabled" set to "false" in a cluster in which nodes have mixed
> "true"/"false" values for the "cdc_enabled" settings.
> The exact behaviour of the above is version-dependant.
> On Cassandra 4.1.1, the cluster will end up in the schema disagreement state.
> A rolling restart will bring the schema back in sync, but the changes made to
> the `cdc` table property will be lost.
> On Cassandra 4.1.6, the cluster will not have visible schema disagreement in
> the "nodetool describecluster" command's output, but the "ALTER TABLE"
> statement only has cosmetic effect on the node it is run. The node with
> "cdc_enabled" set to "false" will show the "cdc" table property has changed,
> but this does not affect its behaviour in any way. At the same time, other
> nodes do not see that table property change at all. This is perhaps even
> worse than on 4.1.1, because the alter table statement is silently failing.
> On Casandra 5.0.0, the behaviour is the same as 4.1.6.
> A shell script for reproducing the above described behaviours in Docker, and
> the outputs of it on both 4.1.1 and 4.1.6 and 5.0.0 are attached.
>
> Edit on 25 Sep: added test result on 5.0.0
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org
