Configuring the Columnstore to Work Effectively


Configuration options described in this section may have significant performance impact on the workload. Make sure to test on a staging environment first before deploying the configuration changes.

Configuring Segment Size in Columnstore Tables

In SingleStore, data in a columnstore table are organized into multiple row segments. For certain workloads, the size of row segments significantly affects performance. Considerations include:

  • In general, larger segments compress better than smaller ones. This leads to less disk usage and faster scan across the table.

  • On the other hand, smaller segments could benefit more from segment elimination. As a result, queries with highly selective filters on the columnstore index run faster with smaller segments.

In SingleStore, the default size of the row segments is controlled by the global variable columnstore_segment_rows. By default, columnstore_segment_rows has a value of 1024000, meaning that each segment contains 1024000 rows by default.


The variable columnar_segment_rows from previous SingleStore versions was deprecated in favor of columnstore_segment_rows. columnar_segment_rows exists as an alias to columnstore_segment_rows.

A cluster upgraded from a version before 6.0 will have the previous default value of 102400 for columnstore_segment_rows.

Refer to Upgrade to SingleStore 7.5 for more information.

In addition to the global variable, it is possible to overwrite the global setting and set the segment size for a specific columnstore table. This can be done during the table creation, or by altering an existing table. For example:

CREATE TABLE t (id INT, SORT KEY (id) WITH (columnstore_segment_rows=100000));
| Table | Create Table                                                            |
| t     | CREATE TABLE `t` (                                                      |
|       |   `id` int(11) DEFAULT NULL,                                            |
|       |   SORT KEY `id` (`id`)                                                  |
|       |   /*!90619 */ /*!90621 WITH(COLUMNSTORE_SEGMENT_ROWS=100000) */         |
|       | )                                                                       |

Now, modify the value of columnstore_segment_rows for this columnstore. The MODIFY KEY keyName clause identifies the index whose settings are to be changed (in this case, the columnstore index).

ALTER TABLE t MODIFY KEY id SET (columnstore_segment_rows=20000);
| Table | Create Table                                                                      |
| t     | CREATE TABLE `t` (                                                                |
|       |   `id` int(11) DEFAULT NULL,                                                      |
|       |    SORT KEY `id` (`id`)                                                           |
|       |    /*!90619 */ /*!90621 WITH(COLUMNSTORE_SEGMENT_ROWS=20000) */                   |
|       | )                                                                                 |

Configuring the Rowstore-backed Segment Size in Columnstore Tables

In addition to the on-disk segments, each SingleStore columnstore also has an in-memory rowstore-backed segment.

In some cases, inserts and updates will write to the rowstore-backed segment first. In these cases, the background flusher process periodically compresses those recently inserted rows and creates on-disk segments. In other cases, inserts and updates will bypass the rowstore-backed segment, and the writes are made to the column-oriented format on disk.

The size of the rowstore-backed segment is controlled by the global variable columnstore_flush_bytes. The background flusher process starts to create on-disk segments when the amount of data in the rowstore-backed segment exceeds columnstore_flush_bytes, with a default of 32 MB. Additionally, insert and load operations are considered small-batch if they write to each partition less than columnstore_flush_bytes * columnstore_disk_insert_threshold, which is by default 16 MB at a time. The amount of time (in seconds) the background flusher waits before trying to flush a table is set via internal_columnstore_idle_flush_wait_seconds. If the table row count has not changed in that period of time then the rows are flushed regardless of the row count. This flushes idle tables that have not reached the threshold.

The minimum size of the disk-backed row segment created by insert and load operations is controlled by the engine variable columnstore_disk_insert_threshold . It is a fractional value with a default of 0.5. Note that if OPTIMIZE TABLE FLUSH is manually run, the minimum segment size can be much smaller.

Considerations about tuning the rowstore-backed segment size include:

  • The rowstore-backed segment is stored in-memory. Therefore, the table consumes less memory when the rowstore-backed segment is smaller.

  • The background flusher process can write more rows to disk at once if the rowstore-backed segment is larger, reducing the number of disk write operations during data ingestion.

Similar to columnstore_segment_rows, the columnstore_flush_bytes value can also be configured per-table with the following syntax:

CREATE TABLE t2 (id INT, SORT KEY (id) WITH (columnstore_flush_bytes=4194304, columnstore_segment_rows=100000));
ALTER TABLE t2 MODIFY KEY id SET (columnstore_flush_bytes=8388608);

Last modified: May 28, 2024

Was this article helpful?