Troubleshoot TiDB OOM Issues

This document describes how to troubleshoot TiDB OOM (Out of Memory) issues, including phenomena, causes, solutions, and diagnostic information.

Typical OOM phenomena

The following are some typical OOM phenomena:

• The client side reports the following error: SQL error, errno = 2013, state = 'HY000': Lost connection to MySQL server during query.

• The Grafana dashboard shows:

  • TiDB > Server > Memory Usage shows that the process/heapInUse metric keeps rising, and suddenly drops to zero after reaching the threshold.
  • TiDB > Server > Uptime suddenly drops to zero.
  • TiDB-Runtime > Memory Usage shows that the estimate-inuse metric keeps rising.

• Check tidb.log, and you can find the following log entries:

  • An alarm about OOM: [WARN] [memory_usage_alarm.go:139] ["tidb-server has the risk of OOM. Running SQLs and heap profile will be recorded in record path"]. For more information, see memory-usage-alarm-ratio.
  • A log entry about restart: [INFO] [printer.go:33] ["Welcome to TiDB."].

Overall troubleshooting process

When you troubleshoot OOM issues, follow this process:

1. Confirm whether it is an OOM issue.

   Execute the following command to check the operating system logs. If there is an oom-killer log near the time when the problem occurs, you can confirm that it is an OOM issue.

   dmesg -T | grep tidb-server
   

   The following is an example of the log that contains oom-killer:

   ......
   Mar 14 16:55:03 localhost kernel: tidb-server invoked oom-killer: gfp_mask=0x201da, order=0, oom_score_adj=0
   Mar 14 16:55:03 localhost kernel: tidb-server cpuset=/ mems_allowed=0
   Mar 14 16:55:03 localhost kernel: CPU: 14 PID: 21966 Comm: tidb-server Kdump: loaded Not tainted 3.10.0-1160.el7.x86_64 #1
   Mar 14 16:55:03 localhost kernel: Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS rel-1.14.0-0-g155821a1990b-prebuilt.qemu.org 04/01/2014
   ......
   Mar 14 16:55:03 localhost kernel: Out of memory: Kill process 21945 (tidb-server) score 956 or sacrifice child
   Mar 14 16:55:03 localhost kernel: Killed process 21945 (tidb-server), UID 1000, total-vm:33027492kB, anon-rss:31303276kB, file-rss:0kB, shmem-rss:0kB
   Mar 14 16:55:07 localhost systemd: tidb-4000.service: main process exited, code=killed, status=9/KILL
   ......
   

2. After confirming that it is an OOM issue, you can further investigate whether the OOM is caused by deployment or the database.

   • If the OOM is caused by a deployment issue, you need to investigate the resource configuration and impact of hybrid deployment.
   • If the OOM is caused by a database issue, the following are some possible causes:
     • TiDB handles large data traffic, such as large queries, large writes, and data import.
     • TiDB is in a high concurrency scenario, where multiple SQL statements consume resources concurrently or operator concurrency is high.
     • TiDB has a memory leak and resources are not released.

   Refer to the following sections for specific troubleshooting methods.

Typical causes and solutions

OOM issues are usually caused by the following:

• Deployment issues
• Database issues
• Client side issues

Deployment issues

The following are some causes of OOM due to improper deployment:

• The memory capacity of the operating system is too small.
• The TiUP configuration resource_control is not appropriate.
• In the case of hybrid deployments (meaning that TiDB and other applications are deployed on the same server), TiDB is killed accidentally by oom-killer due to lack of resources.

Database issues

This section describes the causes and solutions for OOM caused by database issues.

Executing SQL statements consumes too much memory

You can take the following measures to reduce the memory usage of SQL statements, depending on the different causes of OOM issues.

• If the execution plan of SQL is not optimal, for example, due to lack of proper indexes, outdated statistics, or optimizer bugs, a wrong execution plan of SQL might be selected. A huge intermediate result set will then be accumulated in the memory. In this case, consider the following measures:

  • Add appropriate indexes.
  • Use the disk spill feature for execution operators.
  • Adjust the JOIN order between tables.
  • Use hints to optimize SQL statements.

• Some operators and functions are not supported to be pushed down to the storage level, resulting in a huge accumulation of intermediate result sets. In this case, you need to refine the SQL statements or use hints to optimize, and use the functions or operators that support pushing down.

• The execution plan contains the operator HashAgg. HashAgg is executed concurrently by multiple threads, which is faster but consumes more memory. Instead, you can use STREAM_AGG().

• Reduce the number of Regions to be read simultaneously or reduce the concurrency of operators to avoid memory problems caused by high concurrency. The corresponding system variables include:

  • tidb_distsql_scan_concurrency
  • tidb_index_serial_scan_concurrency
  • tidb_executor_concurrency

• The concurrency of sessions is too high near the time point when the problem occurs. In this case, consider scaling out the TiDB cluster by adding more TiDB nodes.

Large transactions or large writes consume too much memory

You need to plan for memory capacity. When a transaction is executed, the memory usage of the TiDB process is scaled up comparing with the transaction size, up to two to three times or more of the transaction size.

You can split a single large transaction to multiple smaller transactions.

The process of collecting and loading statistical information consumes too much memory

A TiDB node needs to load statistics into memory after it starts. TiDB consumes memory when collecting statistical information. You can control memory usage in the following ways:

• Specify a sampling rate, only collect statistics for specific columns, and reduce ANALYZE concurrency.
• Since TiDB v6.1.0, you can use the system variable tidb_stats_cache_mem_quota to control the memory usage for statistical information.
• Since TiDB v6.1.0, you can use the system variable tidb_mem_quota_analyze to control the maximum memory usage when TiDB updates statistics.

For more information, see Introduction to Statistics.

Prepared statements are overused

The client side keeps creating prepared statements but does not execute deallocate prepare stmt, which causes memory consumption to continue to rise and eventually triggers TiDB OOM. The reason is that the memory occupied by a prepared statement is not released until the session is closed. This is especially important for long-time connection sessions.

To solve the problem, consider the following measures:

• Adjust the session lifecycle.
• Adjust the wait_timeout and max_execution_time of the connection pool.
• Use the system variable max_prepared_stmt_count to control the maximum number of prepared statements in a session.

tidb_enable_rate_limit_action is not configured properly

The system variable tidb_enable_rate_limit_action controls memory usage effectively when an SQL statement only reads data. When this variable is enabled and computing operations (such as join or aggregation operations) are required, memory usage might not be under the control of tidb_mem_quota_query, which increases the risk of OOM.

It is recommended that you disable this system variable. Since TiDB v6.3.0, this system variable is disabled by default.

Client side issues

If OOM occurs on the client side, investigate the following:

• Check the trend and speed on Grafana TiDB Details > Server > Client Data Traffic to see if there is a network blockage.
• Check whether there is an application OOM caused by wrong JDBC configuration parameters. For example, if the defaultFetchSize parameter for streaming read is incorrectly configured, it can cause data to be heavily accumulated on the client side.

Diagnostic information to be collected to troubleshoot OOM issues

To locate the root cause of an OOM issue, you need to collect the following information:

• Collect the memory-related configurations of the operating system:

  • TiUP configuration: resource_control.memory_limit
  • Operating system configurations:
    • Memory information: cat /proc/meminfo
    • Kernel parameters: vm.overcommit_memory
  • NUMA information:
    • numactl --hardware
    • numactl --show

• Collect the version information and the memory-related configurations of the database:

  • TiDB version
  • tidb_mem_quota_query
  • memory-usage-alarm-ratio
  • mem-quota-query
  • oom-action
  • tidb_enable_rate_limit_action
  • tidb_server_memory_limit
  • oom-use-tmp-storage
  • tmp-storage-path
  • tmp-storage-quota
  • tidb_analyze_version

• Check the daily usage of TiDB memory on the Grafana dashboard: TiDB > Server > Memory Usage.

• Check the SQL statements that consume more memory.

  • View SQL statement analysis, slow queries, and memory usage on the TiDB Dashboard.
  • Check SLOW_QUERY and CLUSTER_SLOW_QUERY in INFORMATION_SCHEMA.
  • Check tidb_slow_query.log on each TiDB node.
  • Run grep "expensive_query" tidb.log to check the corresponding log entries.
  • Run EXPLAIN ANALYZE to check the memory usage of operators.
  • Run SELECT * FROM information_schema.processlist; to check the value of the MEM column.

• Run the following command to collect the TiDB Profile information when memory usage is high:

  curl -G http://{TiDBIP}:10080/debug/zip?seconds=10" > profile.zip
  

• Run grep "tidb-server has the risk of OOM" tidb.log to check the path of the alert file collected by TiDB Server. The following is an example output:

  ["tidb-server has the risk of OOM. Running SQLs and heap profile will be recorded in record path"] ["is tidb_server_memory_limit set"=false] ["system memory total"=14388137984] ["system memory usage"=11897434112] ["tidb-server memory usage"=11223572312] [memory-usage-alarm-ratio=0.8] ["record path"="/tmp/0_tidb/MC4wLjAuMDo0MDAwLzAuMC4wLjA6MTAwODA=/tmp-storage/record"]
  

See also

• TiDB Memory Control
• Tune TiKV Memory Parameter Performance