Load Base Split
Load Base Split is a new feature introduced in TiDB 4.0. It aims to solve the hotspot issue caused by unbalanced access between Regions, such as full table scans for small tables.
Scenarios
In TiDB, it is easy to generate hotspots when the load is concentrated on certain nodes. PD tries to schedule the hot Regions so that they are distributed as evenly as possible across all nodes for better performance.
However, the minimum unit for PD scheduling is Region. If the number of hotspots in a cluster is smaller than the number of nodes, or if a few hotspots have far more load than other Regions, PD can only move the hotspot from one node to another, but not make the entire cluster share the load.
This scenario is especially common with workloads that are mostly read requests, such as full table scans and index lookups for small tables, or frequent access to some fields.
Previously, the solution to this problem was to manually execute a command to split one or more hotspot Regions, but this approach has two problems:
- Evenly splitting a Region is not always the best choice, because requests might be concentrated on a few keys. In such cases, hotspots might still be on one of the Regions after evenly splitting, and it might take multiple even splits to realize the goal.
- Human intervention is not timely or simple.
Implementation principles
Load Base Split automatically splits the Region based on statistics. It identifies the Regions whose read load or CPU usage consistently exceeds the threshold for 10 seconds, and splits these Regions at a proper position. When choosing the split position, Load Base Split tries to balance the access load of both Regions after the split and avoid access across Regions.
The Region split by Load Base Split will not be merged quickly. On the one hand, PD's MergeChecker
skips the hot Regions; on the other hand, PD also determines whether to merge two Regions according to QPS
in the heartbeat information, to avoid the merging of two Regions with high QPS
.
Usage
The Load Base Split feature is currently controlled by the following parameters:
split.qps-threshold
: The QPS threshold at which a Region is identified as a hotspot. The default value is3000
per second.split.byte-threshold
: The traffic threshold at which a Region is identified as a hotspot. The unit is byte and the default value is 30 MiB per second. (Introduced in v5.0)split.region-cpu-overload-threshold-ratio
: The CPU usage threshold (the percentage of CPU time of the read thread pool) at which a Region is identified as a hotspot. The default value is0.25
. (Introduced in v6.2.0)
If a Region meets one of the following conditions for 10 consecutive seconds, TiKV tries to split the Region:
- the sum of its read requests exceeds
split.qps-threshold
. - its traffic exceeds
split.byte-threshold
. - its CPU usage in the Unified Read Pool exceeds
split.region-cpu-overload-threshold-ratio
.
Load Base Split is enabled by default, but the parameter is set to a rather high value. If you want to disable this feature, set split.qps-threshold
and split.byte-threshold
high enough and set split.region-cpu-overload-threshold-ratio
to 0
at the same time.
To modify the parameter, take either of the following two methods:
Use a SQL statement:
# Set the QPS threshold to 1500 SET config tikv split.qps-threshold=1500; # Set the byte threshold to 15 MiB (15 * 1024 * 1024) SET config tikv split.byte-threshold=15728640; # Set the CPU usage threshold to 50% SET config tikv split.region-cpu-overload-threshold-ratio=0.5;Use TiKV:
curl -X POST "http://ip:status_port/config" -H "accept: application/json" -d '{"split.qps-threshold":"1500"}' curl -X POST "http://ip:status_port/config" -H "accept: application/json" -d '{"split.byte-threshold":"15728640"}' curl -X POST "http://ip:status_port/config" -H "accept: application/json" -d '{"split.region-cpu-overload-threshold-ratio":"0.5"}'
Accordingly, you can view the configuration by either of the following two methods:
Use a SQL statement:
show config where type='tikv' and name like '%split.qps-threshold%';Use TiKV:
curl "http://ip:status_port/config"