SingleStore caches memory (allocated by the Buffer Manager) that is not in use, instead of returning it to the system. The buffer_manager_cached_memory variable keeps a track of this cached memory. Let’s say a query consumes XXX GBs of memory for execution. Once the query execution completes, instead of being returned to the system (Linux), this memory is stored as the Buffer Manager cached memory. Thereafter, if some other query needs memory, it will get that memory from the cache. This process avoids system call/contention over access to memory, which can be substantial. The size of memory reserved for each node in the cache can be viewed in the output of the SHOW STATUS EXTENDED command for each node. For example,

SHOW STATUS EXTENDED LIKE 'buffer_manager_cached%';

+------------------------------+---------+
| Variable_name                |  Value  |
+------------------------------+---------+
| Buffer_manager_cached_memory | 80.0 MB |
+------------------------------+---------+

buffer_manager_cached_memory has following limitations:

• buffer_manager_cached_memory is capped at 25% of the maximum_memory.

• Buffers will only be cached if the total_server_memory is less than 90% of the maximum_memory

For more information on buffer_manager_cached_memory and total_server_memory, see the Summary variables topic.