[
https://issues.apache.org/jira/browse/CASSANALYTICS-181?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=18110597#comment-18110597
]
Bianca Stanciu commented on CASSANALYTICS-181:
----------------------------------------------
[https://github.com/apache/cassandra-analytics/pull/223] is ready for another
look.
Time-skew used to fan out to the per-replica FQDNs from the token map, which
executors can’t reach when Sidecar is behind a load balancer. Now it just calls
timeSkew() against the contact points instead. That’s fine because the
time-skew endpoint doesn’t care
which instance answers (TimeSkewHandler doesn’t look at the Host header), so
any Sidecar behind the LB can respond.
Like you suggested, I made it generic instead of coordinated-only: renamed it
to {*}LoadBalancedCassandraClusterInfo{*}, moved it out of the coordinated
package, and the flag now works for both single-cluster and coordinated writes
(picked in CassandraClusterInfo.create).
Default behavior is unchanged and there are tests for it.
cc [~skoppu]
> Support Sidecar behind a load balancer for time-skew validation in
> coordinated writes
> -------------------------------------------------------------------------------------
>
> Key: CASSANALYTICS-181
> URL: https://issues.apache.org/jira/browse/CASSANALYTICS-181
> Project: Apache Cassandra Analytics
> Issue Type: Improvement
> Components: Writer
> Reporter: Bianca Stanciu
> Assignee: Bianca Stanciu
> Priority: High
> Time Spent: 1h 20m
> Remaining Estimate: 0h
>
> ## Problem
>
>
> Time-skew validation before a coordinated bulk write fans out HTTP calls
> directly to per-replica FQDNs resolved from the Cassandra token map. In
> deployments where Sidecar is fronted by a *load balancer,* those FQDNs are
> *not* routable from Spark executors, only the load balancer address is. This
> causes time-skew validation to fail at the start of every coordinated write
> job, blocking bulk writes entirely in load-balanced Sidecar deployments.
>
> ## Solution
>
>
> Add a _*SIDECAR_BEHIND_LOAD_BALANCER*_ writer option (default: false). When
> true, time-skew validation skips the per-replica fan-out and routes through
> the preconfigured load balancer contact points instead:
>
> {code:java}
> sidecarClient.timeSkew() // routes through LB, no per-replica list{code}
>
>
> Implemented via a *new CoordinatedCassandraClusterInfo* subclass that
> overrides only the time-skew fetch step, leaving all other behavior unchanged.
>
>
> ## Acceptance Criteria
>
>
> - [ ] Flag has no effect on single-cluster writes (warning logged if
> misconfigured)
> - [ ] Default behavior (flag=false) is unchanged
> - [ ] Unit test verifies timeSkew(List) is never called when flag is
> enabled
> - [ ] CHANGES.txt entry added under 0.5.0
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
