Add IME doc

TOC.md

@@ -319,6 +319,7 @@
     - [Tune TiKV Threads](/tune-tikv-thread-performance.md)
     - [Tune TiKV Memory](/tune-tikv-memory-performance.md)
     - [TiKV Follower Read](/follower-read.md)
+    - [TiKV MVCC In-Memory Engine](/tikv-in-memory-engine.md)
     - [Tune Region Performance](/tune-region-performance.md)
     - [Tune TiFlash Performance](/tiflash/tune-tiflash-performance.md)
     - [Coprocessor Cache](/coprocessor-cache.md)

analyze-slow-queries.md

@@ -10,7 +10,7 @@
 1. Among many queries, identify which type of queries are slow.
 2. Analyze why this type of queries are slow.
 
-You can easily perform step 1 using the [slow query log](/dashboard/dashboard-slow-query.md) and the [statement summary table](/statement-summary-tables.md) features. It is recommended to use [TiDB Dashboard](/dashboard/dashboard-intro.md), which integrates the two features and directly displays the slow queries in your browser. 
+You can easily perform step 1 using the [slow query log](/dashboard/dashboard-slow-query.md) and the [statement summary table](/statement-summary-tables.md) features. It is recommended to use [TiDB Dashboard](/dashboard/dashboard-intro.md), which integrates the two features and directly displays the slow queries in your browser.
 
 This document focuses on how to perform step 2 - analyze why this type of queries are slow.
 
@@ -98,9 +98,9 @@
 
 The log above shows that a `cop-task` sent to the `10.6.131.78` instance waits `110ms` before being executed. It indicates that this instance is busy. You can check the CPU monitoring of that time to confirm the cause.
 
-#### Too many outdated keys
+#### Expired MVCC versions and excessive keys
 
-A TiKV instance has much outdated data, which needs to be cleaned up for data scan. This impacts the processing speed.
+Too many expired MVCC versions on TiKV or long GC time can result in an accumulation of excessive MVCC versions, which will affect the scan speed due to the need to process these redundant MVCC versions.
 
 Check `Total_keys` and `Processed_keys`. If they are greatly different, the TiKV instance has too many keys of the older versions.
 
@@ -110,6 +110,8 @@
 ...
 ```
 
+TiDB v8.5.0 introduces the TiKV MVCC In-Memory Engine feature, which can accelerate this type of slow query. For more information, see [TiKV MVCC In-Memory Engine](/tikv-in-memory-engine.md).
+
 ### Other key stages are slow
 
 #### Slow in getting timestamps

tikv-configuration-file.md

@@ -2501,3 +2501,36 @@ Configuration items related to [Load Base Split](/configure-load-base-split.md).
 
 + Specifies the amount of data sampled by Heap Profiling each time, rounding up to the nearest power of 2.
 + Default value: `512KiB`
+
+## in-memory-engine <span class="version-mark">New in v8.5.0</span>
+
+TiKV MVCC in-memory engine configuration items related to the storage layer.
+
+### `enable` <span class="version-mark">New in v8.5.0</span>
+
+> **Note:**
+>
+> This configuration item cannot be queried via SQL statements but can be configured in the configuration file.
+
++ Whether to enable the in-memory engine to accelerate multi-version queries. For more information about the in-memory engine, see [TiKV MVCC In-Memory Engine](/tikv-in-memory-engine.md)
++ Default value: false (the in-memory engine is disabled)
+
+### `capacity` <span class="version-mark">New in v8.5.0</span>
+
+> **Note:**
+>
+> + When the in-memory engine is enabled, `block-cache.capacity` will automatically decrease by 10%.
+> + When you manually configure `capacity`, `block-cache.capacity` will not automatically decrease. In this case, you need to manually adjust the value to avoid OOM.
+
++ Configure the maximum memory size that the in-memory engine can use. The maximum value is 5 GiB. You can manually configure it to use more memory.
++ Default value: 10% of the system memory.
+
+### `gc-run-interval` <span class="version-mark">New in v8.5.0</span>
+
++ Control the time interval for the In-memory Engine GC cached MVCC versions. Reducing this parameter can speed up the GC frequency, reduce MVCC records, but will increase GC CPU consumption and increase the probability of in-memory engine cache miss.
++ Default value: 3m
+
+### `mvcc-amplification-threshold` <span class="version-mark">New in v8.5.0</span>
+
++ Control the threshold for the in-memory engine to select loading Region when MVCC read amplification occurs. The default value is `10`, indicating that when the number of MVCC versions processed for reading a row record in a certain Region exceeds 10, it might be loaded into the in-memory engine.
++ Default value: 10

tikv-in-memory-engine.md

@@ -0,0 +1,131 @@
+---
+title: TiKV MVCC In-Memory Engine
+summary: Learn the applicable scenarios and working principles of the in-memory engine, and how to use the in-memory engine to accelerate queries for MVCC versions.
+---
+
+# TiKV MVCC In-Memory Engine
+
+TiKV MVCC In-Memory Engine (IME) is primarily used to accelerate queries that need to scan a large number of MVCC historical versions, that is, [the total number of versions scanned (total_keys) is much greater than the number of versions processed (processed_keys)](/analyze-slow-queries.md#expired-mvcc-versions-and-excessive-keys).
+
+TiKV MVCC in-memory engine is suitable for the following scenarios:
+
+- An application requires frequent queries on frequently updated or deleted records.
+- An application requires adjusting the [`tidb_gc_life_time`](/garbage-collection-configuration.md#garbage-collection-configuration) to make TiDB retain historical versions for a longer period (for example, 24 hours).
+
+## Working principles
+
+TiKV MVCC in-memory engine caches the latest written MVCC versions in memory and implements an MVCC GC mechanism independent of TiDB, allowing it to quickly perform GC on MVCC versions in memory, reducing the number of versions scanned during queries, and achieving the effect of reducing request latency and CPU overhead.
+
+The following diagram illustrates how TiKV organizes MVCC versions.
+
+<div style="text-align: center;"><img src="./media/tikv-ime-data-organization.png" alt="IME caches recent versions to reduce CPU overhead" width="400" /></div>
+
+The diagram shows two rows of records, each with 9 MVCC versions. The behavior is compared between the cases with and without IME enabled:
+
+- On the left, without IME enabled, the table records are stored in RocksDB in ascending order by primary key, with the same row's MVCC versions adjacent to each other.
+- On the right, with IME enabled, the data in RocksDB is consistent with the left side, and IME caches the latest 2 MVCC versions of the 2 rows of records.
+- When TiKV processes a scan request with a range of `[k1, k2]` and a start timestamp of `8`, the left side without IME enabled needs to process 11 MVCC versions, while the right side with IME enabled only needs to process 4 MVCC versions, reducing request latency and CPU consumption.
+- When TiKV processes a scan request with a range of `[k1, k2]` and a start timestamp of `7`, because the right side lacks the historical versions that need to be read, the IME cache becomes invalid, and it falls back to reading data from RocksDB.
+
+## Usage
+
+Enabling IME requires adjusting the TiKV configuration and restarting. The following example explains the configuration items:
+
+```toml
+[in-memory-engine]
+# This parameter is the switch for the in-memory engine feature, which is disabled by default. You can set it to true to enable it.
+enable = false
+#
+# This parameter controls the memory capacity that in-memory engine can use. The default value is 10% of the system memory, and the maximum value is 5 GiB.
+# You can manually configure it to use more memory.
+# Note: When in-memory engine is enabled, block-cache.capacity will be reduced by 10%.
+#capacity = "5GiB"
+#
+# This parameter controls the time interval for in-memory engine to GC the cached MVCC versions.
+# The default value is 3 minutes, representing that GC is performed every 3 minutes on the cached MVCC versions.
+# Decreasing the value of this parameter can speed up the GC frequency, reduce MVCC versions, but will increase GC CPU consumption and increase the probability of cache miss.
+#gc-run-interval = "3m"
+#
+# This parameter controls the threshold for in-memory engine to select and load Regions based on MVCC read amplification.
+# The default value is 10, indicating that when the number of MVCC versions processed for a row of records in a Region exceeds 10, it might be loaded into the in-memory engine.
+#mvcc-amplification-threshold = 10
+```
+
+> **Note:**
+>
+> + The in-memory engine is disabled by default. After you enable it, you need to restart TiKV.
+> + Except for `enable`, all the other configuration items can be dynamically adjusted.
+
+### Automatic loading
+
+After you enable the in-memory engine, Regions will be automatically loaded based on their read traffic and MVCC amplification. The process is as follows:
+
+1. Regions are sorted by the number of next (RocksDB Iterator next API) and prev (RocksDB Iterator next API) operations in the recent time period.
+2. Regions are filtered using `mvcc-amplification-threshold` (`10` by default. MVCC amplification measures read amplification, calculated as (next + prev) / processed_keys).
+3. The top N Regions with severe MVCC amplification are loaded, where N is based on memory estimation.
+
+IME also periodically performs Region eviction. The process is as follows:
+
+1. IME evicts Regions with low read traffic or low MVCC amplification.
+2. If memory usage reaches 90% of `capacity` and new Regions need to be loaded, IME will filter Regions based on read traffic and evict Regions.
+
+## Compatibility
+
++ [BR](/br/br-use-overview.md): IME can be used with BR, but BR restore will evict IME Regions involved in the restore. After BR restore is complete, if the corresponding Region is still a hotspot, it will be automatically loaded by IME.
++ [TiDB Lightning](/tidb-lightning/tidb-lightning-overview.md): IME can be used with TiDB Lightning, but TiDB Lightning's physical import mode will evict IME Regions involved in the import. After TiDB Lightning completes the import, if the corresponding Region is still a hotspot, it will be automatically loaded by IME.
++ [Follower Read](/develop/dev-guide-use-follower-read.md) and [Stale Read](/develop/dev-guide-use-stale-read.md): IME can be enabled with these two features, but IME can only accelerate Leader coprocessor requests and cannot accelerate Follower Read and Stale Read.
++ [`FLASHBACK CLUSTER`](/sql-statements/sql-statement-flashback-cluster.md): IME can be used with Flashback, but Flashback will cause IME cache invalidation. After Flashback is complete, IME will automatically load hotspot Regions.
+
+## FAQ
+
+### Can in-memory engine reduce write latency and increase write throughput?
+
+No, in-memory engine can only accelerate read requests that scan a large number of MVCC versions.
+
+### How to determine if in-memory engine can improve my scenario?
+
+You can execute the following SQL statement to check if there are slow queries with `Total_keys` much greater than `Process_keys`.
+
+```sql
+SELECT
+    Time,
+    DB,
+    Index_names,
+    Process_keys,
+    Total_keys,
+    CONCAT(
+        LEFT(REGEXP_REPLACE(Query, '\\s+', ' '), 20),
+        '...',
+        RIGHT(REGEXP_REPLACE(Query, '\\s+', ' '), 10)
+    ) as Query,
+    Query_time,
+    Cop_time,
+    Process_time
+FROM
+    INFORMATION_SCHEMA.SLOW_QUERY
+WHERE
+    Is_internal = 0
+    AND Cop_time > 1
+    AND Process_keys > 0
+    AND Total_keys / Process_keys >= 10
+    AND Time >= NOW() - INTERVAL 10 MINUTE
+ORDER BY Total_keys DESC
+LIMIT 5;
+```
+
+Example:
+
+The following result shows that there are queries with severe MVCC amplification on the `db1.tbl1` table. TiKV processes 1358517 MVCC versions and only returns 2 versions.
+
+```
++----------------------------+-----+-------------------+--------------+------------+-----------------------------------+--------------------+--------------------+--------------------+
+| Time                       | DB  | Index_names       | Process_keys | Total_keys | Query                             | Query_time         | Cop_time           | Process_time       |
++----------------------------+-----+-------------------+--------------+------------+-----------------------------------+--------------------+--------------------+--------------------+
+| 2024-11-18 11:56:10.303228 | db1 | [tbl1:some_index] |            2 |    1358517 |  SELECT * FROM tbl1 ... LIMIT 1 ; | 1.2581352350000001 |         1.25651062 |        1.251837479 |
+| 2024-11-18 11:56:11.556257 | db1 | [tbl1:some_index] |            2 |    1358231 |  SELECT * FROM tbl1 ... LIMIT 1 ; |        1.252694002 |        1.251129038 |        1.240532546 |
+| 2024-11-18 12:00:10.553331 | db1 | [tbl1:some_index] |            2 |    1342914 |  SELECT * FROM tbl1 ... LIMIT 1 ; |        1.473941872 | 1.4720495900000001 | 1.3666103170000001 |
+| 2024-11-18 12:01:52.122548 | db1 | [tbl1:some_index] |            2 |    1128064 |  SELECT * FROM tbl1 ... LIMIT 1 ; |        1.058942591 |        1.056853228 |        1.023483875 |
+| 2024-11-18 12:01:52.107951 | db1 | [tbl1:some_index] |            2 |    1128064 |  SELECT * FROM tbl1 ... LIMIT 1 ; |        1.044847031 |        1.042546122 |        0.934768555 |
++----------------------------+-----+-------------------+--------------+------------+-----------------------------------+--------------------+--------------------+--------------------+
+5 rows in set (1.26 sec)
+```

troubleshoot-hot-spot-issues.md

@@ -184,3 +184,7 @@
 ## Scatter read hotspots
 
 In a read hotspot scenario, the hotspot TiKV node cannot process read requests in time, resulting in the read requests queuing. However, not all TiKV resources are exhausted at this time. To reduce latency, TiDB v7.1.0 introduces the load-based replica read feature, which allows TiDB to read data from other TiKV nodes without queuing on the hotspot TiKV node. You can control the queue length of read requests using the [`tidb_load_based_replica_read_threshold`](/system-variables.md#tidb_load_based_replica_read_threshold-new-in-v700) system variable. When the estimated queue time of the leader node exceeds this threshold, TiDB prioritizes reading data from follower nodes. This feature can improve read throughput by 70% to 200% in a read hotspot scenario compared to not scattering read hotspots.
+
+## Use TiKV MVCC in-memory engine to mitigate read hotspots caused by high MVCC read amplification
+
+During long GC times or frequent updates and deletions, read hotspots might occur due to scanning a large number of MVCC versions. To alleviate this type of hotspot, you can enable the in-memory engine feature. For more information, see [TiKV MVCC In-Memory Engine](/tikv-in-memory-engine.md).