Repair¶

Features¶

• Glob patterns to select keyspaces or tables to repair

• Parallel repairs

• Control over repair intensity and parallelism even for ongoing repairs

• Repair order improving performance and stability

• Resilience to schema changes

• Retries

• Pause and resume

Parallel repairs¶

Each node can take part in at most one Scylla repair job at any given moment, but Scylla Manager can repair distinct replica sets in a token ring in parallel. This is beneficial for big clusters. The following diagram presents a benchmark results comparing different parallel flag values. In a benchmark we ran 9 Scylla 2020.1 nodes on AWS i3.2xlarge machines under 50% load, for details check this blog post

By default Scylla Manager runs repairs with full parallelism, you can change that using sctool repair –parallel flag.

Maximal effective parallelism¶

Max parallelism is determined by:

• the constraint that each node can only take part in one ScyllaDB repair job at any given moment.

• ScyllaDB repair job targeting the full replica set of the repaired token range.

For example, let’s assume a cluster with 2 datacenters, 5 nodes each. When you repair the keyspace my_keyspace with replication = {'class': 'NetworkTopologyStrategy', 'dc1': 2, 'dc2': 3}, max parallelism is equal to 1, because each ScyllaDB repair job targets a full replica set of the repaired token range. Every replica set consists of 2 nodes from dc1 and 3 nodes from dc2, so it’s impossible to schedule 2 repair jobs to run simultaneously (dc2 lacks one more node for it to be possible).

Repair is performed table by table and keyspace by keyspace, so max effective parallelism might change depending on which keyspace is being repaired.

Repair intensity¶

Intensity specifies how many token ranges can be repaired in a Scylla node at every given time. The default intensity is one, you can change that using sctool repair –intensity flag.

Scylla Manager 2.2 adds support for intensity value zero. In that case, the number of token ranges is calculated based on node memory and adjusted to ScyllaDB’s maximum number of ranges that can be repaired in parallel. If you want to repair faster, try using intensity zero.

Note that the less the cluster is loaded the more it makes sense to increase intensity. If you increase intensity on a loaded cluster it may not give speed benefits since cluster have no resources to process more repairs. In our experiments in a 50% loaded cluster increasing intensity from 1 to 2 gives about 10-20% boost and increasing it further will have little impact.

Maximal effective intensity¶

Max intensity is calculated based on the max_repair_ranges_in_parallel value (present in ScyllaDB logs). This value might be different for each node in the cluster.

As each ScyllaDB repair job targets some subset of all nodes and ScyllaDB Manager avoids repairing more than max_repair_ranges_in_parallel on any node, the max effective intensity for a given repair job is equal to the minimum max_repair_ranges_in_parallel value of nodes taking part in the job.

Changing repair speed¶

Repair speed is controlled by two parameters: --parallel and --intensity Those parameters can be set when you:

• Schedule a repair with sctool repair

• Update a repair specification with sctool repair update

• Update a running repair task with sctool repair control

More on the topic of repair speed can be found in Repair faster and Repair slower articles.

Repair order¶

Scylla Manager repairs keyspace by keyspace and table by table in order to achieve greater repair stability and performance.

Keyspaces and tables are ordered according to the following rules:

• repair internal (with system prefix) tables before user tables

• repair base tables before Materialized Views and Secondary Indexes

• repair smaller keyspaces and tables first