Identifying and Reducing Memory Usage

Total_server_memory

Total_server_memory tracks the server’s overall memory use. SingleStore does not let this value grow higher than maximum_memory. When Total_server_memory reaches maximum_memory, memory allocations start failing. Queries then fail with the error 1712 - "Not enough memory available to complete the current request. The request was not processed."

In addition, the trace log shows the following: "Nonfatal buffer manager memory allocation failure. The maximum_memory parameter (XXXXX MB) has been reached."

Alloc_table_memory

Alloc_table_memory tracks the memory stored inside of all rowstore tables (memory for rows, indexes, variable-length columns like VARCHAR or JSON that are stored off row). Once Alloc_table_memory reaches maximum_table_memory, INSERT, UPDATE, and LOAD DATA operations against the tables receive the following error: 1720 - "Memory usage by SingleStore for tables (XXXXX MB) has reached the value of 'maximum_table_memory' global variable (YYYYY MB). This query cannot be executed.".

Buffer_manager_memory

Buffer_manager_memory tracks memory that is allocated by the Buffer Manager for SingleStore’s built-in memory allocators. The Buffer Manager is a component that consumes memory from the Linux OS in 128KB blocks and manages that memory out to memory allocators used by rowstore tables or by query execution. If the application makes heavy use of rowstore tables, it’s normal for Buffer_manager_memory to be a large percentage of Total_server_memory.

Buffer_manager_cached_memory

Buffer_manager_cached_memory tracks cached memory blocks allocated by the Buffer Manager that are currently not in use, but still mapped. Keeping these blocks cached allows later allocations to reuse them without requesting memory from the operating system. Table garbage collection and query execution commonly add freed memory to this cache. Other freed Buffer Manager memory can also be added when it is safe to reuse. Buffer_manager_unrecycled_memory reflects the memory that may still be referenced until the check is complete.

While Total_server_memory is at or below 90% of maximum_memory or at least 10 GB of free memory remains, Buffer_manager_cached_memory is capped at 25% of maximum_memory on each node. When the cache size reaches the capped limit, additional freed memory blocks (or buffers) are returned to the operating system. Buffers that are already cached are still mapped and left in place, therefore crossing the capped limit does not trim the cache. The cache shrinks when a subsequent allocation uses the cached buffer, and it can get unmapped if another allocation may otherwise exceed maximum_memory.

When Total_server_memory is above 90% of maximum_memory and less than 10 GB of free memory remains, newly freed buffers are cached only until the cache size reaches approximately 1% of maximum_memory (vs. the default 25%). Existing cached buffers are not trimmed when this limit applies.

As cached memory stays mapped, Total_server_memory includes Buffer_manager_cached_memory. After a large rowstore DELETE, Alloc_table_memory can fall while Total_server_memory stays high because the freed memory blocks are still held as unrecycled or cached memory.

Refer to buffer_manager_cached_memory for more information.

Alloc_query_execution

Alloc_query_execution tracks memory allocated by currently executing queries for sorts, hash tables, result tables, etc. If no queries are running, this value should be 0.

Alloc_variable

Alloc_variable tracks memory allocated for variable-length columns inside rowstore tables, or for other variable-length memory allocations inside query execution (i.e. temporary allocations inside of string expressions, etc.).

Alloc_vector_index

Alloc_vector_index specifies the total size of vector indexes currently residing in memory. The vector index cache loads indexes when they are used and evicts them using an LRU policy when the cache reaches its capacity.

Alloc_skiplist_towers

Alloc_skiplist_towers tracks memory used by the towers for skiplist indexes. Each skiplist index uses on average 40 bytes of memory per row using this allocator. The exact amount of memory per row is probabilistic. It depends on the randomized tower height of the particular row.

Alloc_table_primary

Alloc_table_primary tracks memory used for on-row data for rowstore tables. SingleStore tables share a single row memory allocation amongst all indexes on a particular table. Variable-length columns are not stored in this allocator (VARCHAR, VARBINARY, BLOB, TEXT, JSON, etc). Instead, they are stored in Alloc_variable that was previously discussed in this topic.

Alloc_deleted_version

Alloc_deleted_version tracks memory used to mark rows as deleted in rowstore tables. DELETE queries in SingleStore don’t free up memory when they commit. They mark rows as deleted and the garbage collector frees this memory up when it is safe to do so (i.e. no query or operation is using the deleted row anymore). If this number is large, it means the garbage collector is behind or some operation is preventing the garbage collector from physically freeing the memory used by deleted rows. Examples of this could be a snapshot or a backup, or a long running query, etc.

Alloc_hash_buckets

Alloc_hash_buckets tracks memory used for HASH index buckets (by default 4 million buckets per index, which would use 32 MB).

Alloc_replication

Alloc_replication tracks the amount of memory used during replication.

Malloc_active_memory

Malloc_active_memory tracks memory allocated directly from the Linux OS and managed by the C runtime allocators (not SingleStore’s built-in memory allocators that use the Buffer Manager). The memory use here should be approximately 1-2 GBs for most workloads. Column store tables, open connections, and memory for metadata about tables, columns, etc. are the biggest consumers of memory.

Alloc_thread_stacks

Alloc_thread_stacks tracks memory used by thread stacks. SingleStore caches threads used to run queries. Each thread has a 1 MB stack by default. This can be controlled by the thread_stack session variable, but SingleStore does not recommend changing this value. SingleStore kills threads it hasn’t used for 24 hours which frees up stack memory (this can be controlled by the idle_thread_lifetime_seconds variable).