To help improve your cluster's performance, CockroachDB can automatically merge small ranges of data together to form fewer, larger ranges. This can both improve query latency, as well as your cluster's survivability.
Settings
Enable/disable range merges
Use SET CLUSTER SETTING to set kv.range_merge.queue_enabled to:
trueto enable range merges (default)falseto disable range merges
> SET CLUSTER SETTING kv.range_merge.queue_enabled = true;
When to enable range merges
Range merges can provide performance improvements if you're working with a CockroachDB engineer and running CockroachDB 2.1.3 or newer.
When to disable range merges
Disable (or do not use) range merges if you want to manually control range splits using SPLIT AT.
This limitation has been lifted in v19.2. If you disabled automatic range merging in order to use manual splits, and you are upgrading to v19.2, consider setting kv.range_merge.queue_enabled to true to improve performance.
How range merges work
Overview
CockroachDB splits your cluster's data into many ranges (64MiB by default), which are defined by the range of keys they contain. For example, your cluster might have a range for customers whose IDs are between [1000, 2000). If that range grows beyond 64MiB of data, the range is split into two 32MiB ranges.
However, as you delete data from your cluster, a range might contain far less data. Over the lifetime of a cluster, this could lead to a number of small ranges.
To reduce the number of small ranges, your cluster can have any range below a certain threshold (16MiB by default) try to merge with its "right-hand neighbor", i.e., the range that starts where this range ends. Using our example above, this might be the range for customers whose IDs are between [2000, 3000).
If the combined size of the small range and its neighbor is less than the maximum range size, the ranges merge into a single range. In our example, this would create a new range of keys [1000, 3000).
Ranges only attempt to merge with their right-hand neighbor. Ranges do not currently attempt to merge with their left-hand neighbor (i.e., the range that ends where this range begins).
Why range merges improve performance
Query latency
Queries in CockroachDB must contact a replica of each range involved in the query. This creates the following issues for clusters with many small ranges:
- Queries incur a fixed overhead in terms of processing time for each range they must coordinate with.
- Having many small ranges can increase the number of machines your query must coordinate with. This exposes your query to a greater likelihood of running into issues like network latency or overloaded nodes.
By merging small ranges, CockroachDB can greatly reduce the number of ranges involved in queries and reduce their latency.
Survivability
CockroachDB automatically rebalances the distribution of ranges in your cluster whenever nodes come online or go offline.
During rebalancing, it's preferable to replicate a few larger ranges across nodes. It requires less coordination and often completes more quickly.
By merging ranges together, your cluster needs to rebalance fewer ranges, which ultimately improves your cluster's performance, especially in the face of availability events like node outages.
Was this helpful?
