Troubleshoot a TiFlash Cluster

This section describes some commonly encountered issues when using TiFlash, the reasons, and the solutions.

TiFlash fails to start

The issue might occur due to different reasons. It is recommended that you troubleshoot it following the steps below:

  1. Check whether your system is CentOS8.

    CentOS8 does not have the libnsl.so system library. You can manually install it via the following command:

    dnf install libnsl
  2. Check your system's ulimit parameter setting.

    ulimit -n 1000000
  3. Use the PD Control tool to check whether there is any TiFlash instance that failed to go offline on the node (same IP and Port) and force the instance(s) to go offline. For detailed steps, refer to Scale in a TiFlash cluster.

If the above methods cannot resolve your issue, save the TiFlash log files and email to info@pingcap.com for more information.

TiFlash replica is always unavailable

This is because TiFlash is in an abnormal state caused by configuration errors or environment issues. Take the following steps to identify the faulty component:

  1. Check whether PD enables the Placement Rules feature:

    echo 'config show replication' | /path/to/pd-ctl -u http://${pd-ip}:${pd-port}
  2. Check whether the TiFlash process is working correctly by viewing UpTime on the TiFlash-Summary monitoring panel.

  3. Check whether the TiFlash proxy status is normal through pd-ctl.

    echo "store" | /path/to/pd-ctl -u http://${pd-ip}:${pd-port}

    The TiFlash proxy's store.labels includes information such as {"key": "engine", "value": "tiflash"}. You can check this information to confirm a TiFlash proxy.

  4. Check whether pd buddy can correctly print the logs (the log path is the value of log in the [flash.flash_cluster] configuration item; the default log path is under the tmp directory configured in the TiFlash configuration file).

  5. Check whether the number of configured replicas is less than or equal to the number of TiKV nodes in the cluster. If not, PD cannot replicate data to TiFlash:

    echo 'config placement-rules show' | /path/to/pd-ctl -u http://${pd-ip}:${pd-port}

    Reconfirm the value of default: count.

  6. Check whether the remaining disk space of the machine (where store of the TiFlash node is) is sufficient. By default, when the remaining disk space is less than 20% of the store capacity (which is controlled by the low-space-ratio parameter), PD cannot schedule data to this TiFlash node.

TiFlash query time is unstable, and the error log prints many Lock Exception messages

This is because large amounts of data are written to the cluster, which causes that the TiFlash query encounters a lock and requires query retry.

You can set the query timestamp to one second earlier in TiDB. For example, if the current time is '2020-04-08 20:15:01', you can execute set @@tidb_snapshot='2020-04-08 20:15:00'; before you execute the query. This makes less TiFlash queries encounter a lock and mitigates the risk of unstable query time.

Some queries return the Region Unavailable error

If the load pressure on TiFlash is too heavy and it causes that TiFlash data replication falls behind, some queries might return the Region Unavailable error.

In this case, you can balance the load pressure by adding more TiFlash nodes.

Data file corruption

Take the following steps to handle the data file corruption:

  1. Refer to Take a TiFlash node down to take the corresponding TiFlash node down.
  2. Delete the related data of the TiFlash node.
  3. Redeploy the TiFlash node in the cluster.

TiFlash analysis is slow

If a statement contains operators or functions not supported in the MPP mode, TiDB does not select the MPP mode. Therefore, the analysis of the statement is slow. In this case, you can execute the EXPLAIN statement to check for operators or functions not supported in the MPP mode.

create table t(a datetime); alter table t set tiflash replica 1; insert into t values('2022-01-13'); set @@session.tidb_enforce_mpp=1; explain select count(*) from t where subtime(a, '12:00:00') > '2022-01-01' group by a; show warnings;

In this example, the warning message shows that TiDB does not select the MPP mode because TiDB 5.4 and earlier versions do not support the subtime function.

+---------+------+-----------------------------------------------------------------------------+ > | Level | Code | Message | +---------+------+-----------------------------------------------------------------------------+ | Warning | 1105 | Scalar function 'subtime'(signature: SubDatetimeAndString, return type: datetime) is not supported to push down to tiflash now. | +---------+------+-----------------------------------------------------------------------------+

Data is not replicated to TiFlash

After deploying a TiFlash node and starting replication (by performing the ALTER operation), no data is replicated to it. In this case, you can identify and address the problem by following the steps below:

  1. Check whether the replication is successful by running the ALTER table <tbl_name> set tiflash replica <num> command and check the output.

    • If there is output, go to the next step.
    • If there is no output, run the SELECT * FROM information_schema.tiflash_replica command to check whether TiFlash replicas have been created. If not, run the ALTER table ${tbl_name} set tiflash replica ${num} command again, check whether other statements (for example, add index) have been executed, or check whether DDL executions are successful.
  2. Check whether the TiFlash process runs correctly.

    Check whether there is any change in progress, the flash_region_count parameter in the tiflash_cluster_manager.log file, and the Grafana monitoring item Uptime:

    • If yes, the TiFlash process runs correctly.
    • If no, the TiFlash process is abnormal. Check the tiflash log for further information.
  3. Check whether the Placement Rules function has been enabled by using pd-ctl:

    echo 'config show replication' | /path/to/pd-ctl -u http://<pd-ip>:<pd-port>
  4. Check whether the max-replicas configuration is correct:

    • If the value of max-replicas does not exceed the number of TiKV nodes in the cluster, go to the next step.
    • If the value of max-replicas is greater than the number of TiKV nodes in the cluster, the PD does not replicate data to the TiFlash node. To address this issue, change max-replicas to an integer fewer than or equal to the number of TiKV nodes in the cluster.
    curl -X POST -d '{ "group_id": "pd", "id": "default", "start_key": "", "end_key": "", "role": "voter", "count": 3, "location_labels": [ "host" ] }' <http://172.16.x.xxx:2379/pd/api/v1/config/rule>
  5. Check whether the connection between TiDB or PD and TiFlash is normal.

    Search the flash_cluster_manager.log file for the ERROR keyword.

    • If no ERROR is found, the connection is normal. Go to the next step.

    • If ERROR is found, the connection is abnormal. Perform the following check.

      • Check whether the log records PD keywords.

        If PD keywords are found, check whether raft.pd_addr in the TiFlash configuration file is valid. Specifically, run the curl '{pd-addr}/pd/api/v1/config/rules' command and check whether there is any output in 5s.

      • Check whether the log records TiDB-related keywords.

        If TiDB keywords are found, check whether flash.tidb_status_addr in the TiFlash configuration file is valid. Specifically, run the curl '{tidb-status-addr}/tiflash/replica' command and check whether there is any output in 5s.

      • Check whether the nodes can ping through each other.

  6. Check whether placement-rule is created for tables.

    Search the flash_cluster_manager.log file for the Set placement rule … table-<table_id>-r keyword.

    • If the keyword is found, go to the next step.
    • If not, collect logs of the corresponding component for troubleshooting.
  7. Check whether the PD schedules properly.

    Search the pd.log file for the table-<table_id>-r keyword and scheduling behaviors like add operator.

    • If the keyword is found, the PD schedules properly.
    • If not, the PD does not schedule properly. Contact PingCAP technical support for help.

Data replication gets stuck

If data replication on TiFlash starts normally but then all or some data fails to be replicated after a period of time, you can confirm or resolve the issue by performing the following steps:

  1. Check the disk space.

    Check whether the disk space ratio is higher than the value of low-space-ratio (defaulted to 0.8. When the space usage of a node exceeds 80%, the PD stops migrating data to this node to avoid exhaustion of disk space).

    • If the disk usage ratio is greater than or equal to the value of low-space-ratio, the disk space is insufficient. To relieve the disk space, remove unnecessary files, such as space_placeholder_file (if necessary, set reserve-space to 0MB after removing the file) under the ${data}/flash/ folder.
    • If the disk usage ratio is less than the value of low-space-ratio, the disk space is sufficient. Go to the next step.
  2. Check the network connectivity between TiKV, TiFlash, and PD.

    In flash_cluster_manager.log, check whether there are any new updates to flash_region_count corresponding to the table that gets stuck.

    • If no, go to the next step.

    • If yes, search for down peer (replication gets stuck if there is a peer that is down).

      • Run pd-ctl region check-down-peer to search for down peer.
      • If down peer is found, run pd-ctl operator add remove-peer\<region-id> \<tiflash-store-id> to remove it.
  3. Check CPU usage.

    On Grafana, choose TiFlash-Proxy-Details > Thread CPU > Region task worker pre-handle/generate snapshot CPU. Check the CPU usage of <instance-ip>:<instance-port>-region-worker.

    If the curve is a straight line, the TiFlash node is stuck. Terminate the TiFlash process and restart it, or contact PingCAP technical support for help.

Data replication is slow

The causes may vary. You can address the problem by performing the following steps.

  1. Adjust the value of the scheduling parameters.

  2. Adjust the load on TiFlsh.

    Excessively high load on TiFlash can also result in slow replication. You can check the load of TiFlash indicators on the TiFlash-Summary panel on Grafana:

    • Applying snapshots Count: TiFlash-summary > raft > Applying snapshots Count
    • Snapshot Predecode Duration: TiFlash-summary > raft > Snapshot Predecode Duration
    • Snapshot Flush Duration: TiFlash-summary > raft > Snapshot Flush Duration
    • Write Stall Duration: TiFlash-summary > Storage Write Stall > Write Stall Duration
    • generate snapshot CPU: TiFlash-Proxy-Details > Thread CPU > Region task worker pre-handle/generate snapshot CPU

    Based on your service priorities, adjust the load accordingly to achieve optimal performance.