[CASSSIDECAR-79] Sidecar should be able to load metadata even if the local instance is unavailable

• Type: Improvement
• Status: Resolved
• Priority: Normal
• Resolution: Fixed
• Fix Version/s: 1.0
• Component/s: Rest API
• Labels: pull-request-available
• Change Category: Operability
• Complexity: Challenging

Source Control Link

https://github.com/apache/cassandra-sidecar/commit/91f31399042a444c2ccfc55d28d64bfe36a72a74

Test and Documentation Plan

New integration tests covering starting/stopping instances.

• Assignee: Doug Rohrer
• Reporter: Doug Rohrer

Authors

Doug Rohrer

Reviewers

Francisco Guerrero, Yifan Cai

• Created: 20/Oct/23 01:10
• Updated: 05/Dec/23 17:18
• Resolved: 05/Dec/23 17:17

Description

Today, the sidecar’s CassandraAdapter/CqlSessionProvider uses a load balancing policy that only allows connections to a single instance configured for one of the managed instances of Cassandra. However, this is both unnecessary and potentially harmful, as requests to that instance’s hostname will fail to retrieve information like keyspace/cluster metadata when that instance is unavailable, and will not receive any updates to metadata either, which can cause stale metadata to be used.

Instead, the CqlSessionProvider should take a list of contact points and use them to connect to the whole cluster (or a subset of nodes) without the single-instance load balancing policy. Where it is necessary to query an individual instance (system tables, attempts to check the health of a particular instance) we can, instead, use the driver’s ability to direct queries to a specific host on a per-statement basis.

Issue Links

• links to: GitHub Pull Request #74