Columnstore replica

Columnstore replica (C replica) is a read-only replica type in OceanBase Database, primarily used to handle analytical (AP) read-only queries. Data on a C replica is stored in columnar format and is used in conjunction with full-featured replicas (F replicas) and read-only replicas (R replicas). Transactional writes and regular strong reads still primarily occur on F or R replicas; AP scans can benefit from the more suitable storage format on C replicas. On a C replica, all user tables (including replicated tables, but excluding index tables, internal tables, and system tables) are stored in columnar format. OLAP workloads access C replicas through a dedicated proxy endpoint and perform decision analysis tasks using weak-consistency reads.

Columnstore replicas have the following main characteristics:

• Includes complete logs, MemTables, and SSTables.

• Major SSTable: User tables on the same log stream are organized in columnar format on a C replica, facilitating large-scale scanning and AP computation.

• Incremental and fusion: Incremental data such as MemTables and minor compactions is still organized in row format. The C replica asynchronously catches up with the source replica via clogs and replays them locally.

• Cannot form a Paxos member group. It does not participate in log voting as a Paxos member but acts as an observer, continuously catching up with the logs of Paxos members and replaying them locally, thus avoiding increased transaction commit latency due to more voting members.

• Can provide read-only services when the business does not require high consistency for data reads.

• On the same log stream, user tables are stored in columnar format at the major SSTable level.

• Cannot be converted to other replica types.

More features and their descriptions are shown in the following table.

Scenarios where the storage format on a C replica is temporarily row-based

In OceanBase Database, if a table is created as a row-based table, the system creates a corresponding pure columnstore table on the C replica. If a table is created as a columnstore table, its storage format on the C replica remains consistent with that on the F replica. Therefore, for a C replica, it only converts the user partitions of row-based user tables from the F replica into columnar storage. This description refers only to the "final" state being columnar. On a C replica, partitions are not always in columnar format. In the following scenarios, user table partitions on the C replica are temporarily row-based, requiring the system to automatically schedule row-to-column conversion tasks to convert the latest baseline data into columnar format.

In the above scenarios, during the conversion from row-based to columnar storage, although the optimizer generates a columnar query plan, the actual execution is still a query against the row-based baseline. Users can query the available information and the progress of the row-to-column conversion on a C replica using the views CDB_OB_CS_REPLICA_STATS (system tenant) and DBA_OB_CS_REPLICA_STATS (user tenant). After the row-to-column conversion task is completely finished, you can then perform queries on the C replica.

In the sys tenant, the following example shows how to query the progress of converting tablets in the log streams of all tenants' C replicas:

obclient[oceanbase]> SELECT * FROM oceanbase.CDB_OB_CS_REPLICA_STATS;

The query result is as follows:

+-----------+----------------+----------+-------+------------------+----------------------+-----------------------+---------------------------+-----------+
| TENANT_ID | SVR_IP         | SVR_PORT | LS_ID | TOTAL_TABLET_CNT | AVAILABLE_TABLET_CNT | TOTAL_MACRO_BLOCK_CNT | AVAILABLE_MACRO_BLOCK_CNT | AVAILABLE |
+-----------+----------------+----------+-------+------------------+----------------------+-----------------------+---------------------------+-----------+
|      1004 | xx.xxx.xxx.212 |    63000 |  1001 |             1019 |                 1019 |                 10706 |                     10706 | TRUE      |
|      1006 | xx.xxx.xxx.212 |    63000 |  1001 |              133 |                  133 |                   875 |                       875 | TRUE      |
+-----------+----------------+----------+-------+------------------+----------------------+-----------------------+---------------------------+-----------+
2 rows in set

From the first line of the query result, we can see that for the tenant with ID 1004, the server hosting its C replica is xx.xxx.xxx.212, with port number 63000. The ID of the C replica log stream is 1001. Currently, there are 1019 partitions that need to be converted to columnar format, and 1019 partitions are available. The total number of baseline macroblocks is 10706, and 10706 are available. You can roughly estimate the progress of the row-to-column conversion using AVAILABLE_TABLET_CNT / TOTAL_TABLET_CNT or AVAILABLE_MACRO_BLOCK_CNT / TOTAL_MACRO_BLOCK_CNT. A log stream is fully available only when all tablets in it are available.