956b092012
" Here is a simple introduction to the node operations scylla supports and some of the issues. - Replace operation It is used to replace a dead node. The token ring does not change. It pulls data from only one of the replicas which might not be the latest copy. - Rebuild operation It is used to get all the data this node owns form other nodes. It pulls data from only one of the replicas which might not be the latest copy. - Bootstrap operation It is used to add a new node into the cluster. The token ring changes. Do no suffer from the "not the latest replica” issue. New node pulls data from existing nodes that are losing the token range. Suffer from failed streaming. We split the ranges in 10 groups and we stream one group at a time. Restream the group if failed, causing unnecessary data transmission on wire. Bootstrap is not resumable. Failure after 99.99% of data is streamed. If we restart the node again, we need to stream all the data again even if the node already has 99.99% of the data. - Decommission operation It is used to remove a live node form the cluster. Token ring changes. Do not suffer “not the latest replica” issue. The leaving node pushes data to existing nodes. It suffers from resumable issue like bootstrap operation. - Removenode operation It is used to remove a dead node out of the cluster. Existing nodes pulls data from other existing nodes for the new ranges it own. It pulls from one of the replicas which might not be the latest copy. To solve all the issues above. We could use repair based node operation. The idea behind repair based node operations is simple: use repair to sync data between replicas instead of streaming. The benefits: - Latest copy is guaranteed - Resumable in nature - No extra data is streamed on wire E.g., rebuild twice, will not stream the same data twice - Unified code path for all the node operations - Free repair operation during bootstrap, replace operation and so on. Fixes: #3003 Fixes: #4208 Tests: update_cluster_layout_tests.py + replace_address_test.py + manual test " * 'repair_for_node_ops' of https://github.com/asias/scylla: docs: Add doc for repair_based_node_ops storage_service: Enable node repair based ops for bootstrap storage_service: Enable node repair based ops for decommission storage_service: Enable node repair based ops for replace storage_service: Enable node repair based ops for removenode storage_service: Enable node repair based ops for rebuild storage_service: Use the same tokens as previous bootstrap storage_service: Add is_repair_based_node_ops_enabled helper config: Add enable_repair_based_node_ops repair: Add replace_with_repair repair: Add rebuild_with_repair repair: Add do_rebuild_replace_with_repair repair: Add removenode_with_repair repair: Add decommission_with_repair repair: Add do_decommission_removenode_with_repair repair: Add bootstrap_with_repair repair: Introduce sync_data_using_repair repair: Propagate exception in tracker::run
Scylla developer documentation
This folder (and its subfolders) contain developer-oriented documentation concerning the Scylla codebase. We also have a wiki, which contains additional developer-oriented documentation. There is currently no clear definition of what goes where, so when looking for something be sure to check both.
Seastar documentation can be found here.
User documentation can be found on docs.scylladb.com
For information on how to build Scylla and how to contribute visit HACKING.md and CONTRIBUTING.md.