Garbage Collection Configuration
You can configure garbage collection (GC) using the following system variables:
tidb_gc_enable: controls whether to enable garbage collection for TiKV.tidb_gc_run_interval: specifies the GC interval.tidb_gc_life_time: specifies the time limit during which data is retained for each GC.tidb_gc_concurrency: specifies the number of threads in the Resolve Locks step of GC.tidb_gc_scan_lock_mode: specifies the way of scanning locks in the Resolve Locks step of GC.tidb_gc_max_wait_time: specifies the maximum time that active transactions block the GC safe point.
For more information about how to modify the value of a system variable, see System variables.
GC I/O limit
TiKV supports the GC I/O limit. You can configure gc.max-write-bytes-per-sec to limit writes of a GC worker per second, and thus to reduce the impact on normal requests.
0 indicates disabling this feature.
You can dynamically modify this configuration using tikv-ctl:
tikv-ctl --host=ip:port modify-tikv-config -n gc.max-write-bytes-per-sec -v 10MB
Changes in TiDB 5.0
In previous releases of TiDB, garbage collection was configured via the mysql.tidb system table. While changes to this table continue to be supported, it is recommended to use the system variables provided. This helps ensure that any changes to configuration can be validated, and prevent unexpected behavior (#20655).
The CENTRAL garbage collection mode is no longer supported. The DISTRIBUTED GC mode (which has been the default since TiDB 3.0) will automatically be used in its place. This mode is more efficient, since TiDB no longer needs to send requests to each TiKV region to trigger garbage collection.
For information on changes in previous releases, refer to earlier versions of this document using the TIDB version selector in the left hand menu.
Changes in TiDB 6.1.0
Before TiDB v6.1.0, the transaction in TiDB does not affect the GC safe point. Starting from v6.1.0, TiDB considers the startTS of the transaction when calculating the GC safe point, to resolve the problem that the data to be accessed has been cleared. If the transaction is too long, the safe point will be blocked for a long time, which affects the application performance.
In TiDB v6.1.0, the system variable tidb_gc_max_wait_time is introduced to control the maximum time that active transactions block the GC safe point. After the value is exceeded, the GC safe point is forwarded forcefully.
GC in Compaction Filter
Based on the DISTRIBUTED GC mode, the mechanism of GC in Compaction Filter uses the compaction process of RocksDB, instead of a separate GC worker thread, to run GC. This new GC mechanism helps to avoid extra disk read caused by GC. Also, after clearing the obsolete data, it avoids a large number of left tombstone marks which degrade the sequential scan performance.
The following example shows how to enable the mechanism in the TiKV configuration file:
[gc]
enable-compaction-filter = true
You can also enable this GC mechanism by modifying the configuration dynamically. See the following example:
show config where type = 'tikv' and name like '%enable-compaction-filter%';
+------+-------------------+-----------------------------+-------+
| Type | Instance | Name | Value |
+------+-------------------+-----------------------------+-------+
| tikv | 172.16.5.37:20163 | gc.enable-compaction-filter | false |
| tikv | 172.16.5.36:20163 | gc.enable-compaction-filter | false |
| tikv | 172.16.5.35:20163 | gc.enable-compaction-filter | false |
+------+-------------------+-----------------------------+-------+
set config tikv gc.enable-compaction-filter = true;
show config where type = 'tikv' and name like '%enable-compaction-filter%';
+------+-------------------+-----------------------------+-------+
| Type | Instance | Name | Value |
+------+-------------------+-----------------------------+-------+
| tikv | 172.16.5.37:20163 | gc.enable-compaction-filter | true |
| tikv | 172.16.5.36:20163 | gc.enable-compaction-filter | true |
| tikv | 172.16.5.35:20163 | gc.enable-compaction-filter | true |
+------+-------------------+-----------------------------+-------+