Fix for row-level repairs

Upgrade to 2.0.2 for improved repair speed in Scylla 3.1 and up

One of the useful features of the Manager is how it handles repairs. Manager breaks down token ranges into smaller segments in order distribute load over all available shards. Result of this approach is more efficient repair execution.

With the release of Scylla 3.1 new improvement called row-level repair was introduced. This change approaches repair on the more granular level which makes optimizations done by the Manager obsolete. In practice we noticed degradation of repair execution time with repairs done by Manager on Scylla clusters with row-level repair feature enabled.

Manager 2.0.2 introduces fix for handling repairs on clusters with row-level repair feature enabled. Manager will detect support for this feature and will delegate sharding to the Scylla node by avoiding any split to shards it self. If you experience slow repairs please upgrade to Manager 2.0.2 or newer.

Slowdowns are still possible if manager is not configured correctly. If segments_per_repair configuration option (scylla-manager.yaml) is set to low value repair can still take long time to finish. So for clusters with row-level repair it is recommended to set segments_per_repair to at least 16. Row-level repairs are available from the Scylla Open Source 3.1.