Performance tuning with ordered clustered columnstore index
When users query a columnstore table in Synapse SQL pool, the optimizer checks the minimum and maximum values stored in each segment. Segments that are outside the bounds of the query predicate aren't read from disk to memory. A query can get faster performance if the number of segments to read and their total size are small.
Ordered vs. non-ordered clustered columnstore index
By default, for each table created without an index option, an internal component (index builder) creates a non-ordered clustered columnstore index (CCI) on it. Data in each column is compressed into a separate CCI rowgroup segment. There's metadata on each segment's value range, so segments that are outside the bounds of the query predicate aren't read from disk during query execution. CCI offers the highest level of data compression and reduces the size of segments to read so queries can run faster. However, because the index builder doesn't sort data before compressing them into segments, segments with overlapping value ranges could occur, causing queries to read more segments from disk and take longer to finish.
When creating an ordered CCI, the Synapse SQL engine sorts the existing data in memory by the order key(s) before the index builder compresses them into index segments. With sorted data, segment overlapping is reduced allowing queries to have a more efficient segment elimination and thus faster performance because the number of segments to read from disk is smaller. If all data can be sorted in memory at once, then segment overlapping can be avoided. Due to large tables in data warehouses, this scenario doesn't happen often.
To check the segment ranges for a column, run the following command with your table name and column name:
SELECT o.name, pnp.index_id, cls.row_count, pnp.data_compression_desc, pnp.pdw_node_id, pnp.distribution_id, cls.segment_id, cls.column_id, cls.min_data_id, cls.max_data_id, cls.max_data_id-cls.min_data_id as difference FROM sys.pdw_nodes_partitions AS pnp JOIN sys.pdw_nodes_tables AS Ntables ON pnp.object_id = NTables.object_id AND pnp.pdw_node_id = NTables.pdw_node_id JOIN sys.pdw_table_mappings AS Tmap ON NTables.name = TMap.physical_name AND substring(TMap.physical_name,40, 10) = pnp.distribution_id JOIN sys.objects AS o ON TMap.object_id = o.object_id JOIN sys.pdw_nodes_column_store_segments AS cls ON pnp.partition_id = cls.partition_id AND pnp.distribution_id = cls.distribution_id JOIN sys.columns as cols ON o.object_id = cols.object_id AND cls.column_id = cols.column_id WHERE o.name = '<Table Name>' and cols.name = '<Column Name>' and TMap.physical_name not like '%HdTable%' ORDER BY o.name, pnp.distribution_id, cls.min_data_id
In an ordered CCI table, the new data resulting from the same batch of DML or data loading operations are sorted within that batch, there is no global sorting across all data in the table. Users can REBUILD the ordered CCI to sort all data in the table. In Synapse SQL, the columnstore index REBUILD is an offline operation. For a partitioned table, the REBUILD is done one partition at a time. Data in the partition that is being rebuilt is "offline" and unavailable until the REBUILD is complete for that partition.
A query's performance gain from an ordered CCI depends on the query patterns, the size of data, how well the data is sorted, the physical structure of segments, and the DWU and resource class chosen for the query execution. Users should review all these factors before choosing the ordering columns when designing an ordered CCI table.
Queries with all these patterns typically run faster with ordered CCI.
- The queries have equality, inequality, or range predicates
- The predicate columns and the ordered CCI columns are the same.
- The predicate columns are used in the same order as the column ordinal of ordered CCI columns.
In this example, table T1 has a clustered columnstore index ordered in the sequence of Col_C, Col_B, and Col_A.
CREATE CLUSTERED COLUMNSTORE INDEX MyOrderedCCI ON T1 ORDER (Col_C, Col_B, Col_A)
The performance of query 1 can benefit more from ordered CCI than the other three queries.
-- Query #1: SELECT * FROM T1 WHERE Col_C = 'c' AND Col_B = 'b' AND Col_A = 'a'; -- Query #2 SELECT * FROM T1 WHERE Col_B = 'b' AND Col_C = 'c' AND Col_A = 'a'; -- Query #3 SELECT * FROM T1 WHERE Col_B = 'b' AND Col_A = 'a'; -- Query #4 SELECT * FROM T1 WHERE Col_A = 'a' AND Col_C = 'c';
Data loading performance
The performance of data loading into an ordered CCI table is similar to a partitioned table. Loading data into an ordered CCI table can take longer than a non-ordered CCI table because of the data sorting operation, however queries can run faster afterwards with ordered CCI.
Here is an example performance comparison of loading data into tables with different schemas.
Here is an example query performance comparison between CCI and ordered CCI.
Reduce segment overlapping
The number of overlapping segments depends on the size of data to sort, the available memory, and the maximum degree of parallelism (MAXDOP) setting during ordered CCI creation. Below are options to reduce segment overlapping when creating ordered CCI.
Use xlargerc resource class on a higher DWU to allow more memory for data sorting before the index builder compresses the data into segments. Once in an index segment, the physical location of the data cannot be changed. There's no data sorting within a segment or across segments.
Create ordered CCI with MAXDOP = 1. Each thread used for ordered CCI creation works on a subset of data and sorts it locally. There's no global sorting across data sorted by different threads. Using parallel threads can reduce the time to create an ordered CCI but will generate more overlapping segments than using a single thread. Currently, the MAXDOP option is only supported in creating an ordered CCI table using CREATE TABLE AS SELECT command. Creating an ordered CCI via CREATE INDEX or CREATE TABLE commands does not support the MAXDOP option. For example,
CREATE TABLE Table1 WITH (DISTRIBUTION = HASH(c1), CLUSTERED COLUMNSTORE INDEX ORDER(c1) ) AS SELECT * FROM ExampleTable OPTION (MAXDOP 1);
- Pre-sort the data by the sort key(s) before loading them into tables.
Here is an example of an ordered CCI table distribution that has zero segment overlapping following above recommendations. The ordered CCI table is created in a DWU1000c database via CTAS from a 20-GB heap table using MAXDOP 1 and xlargerc. The CCI is ordered on a BIGINT column with no duplicates.
Create ordered CCI on large tables
Creating an ordered CCI is an offline operation. For tables with no partitions, the data won't be accessible to users until the ordered CCI creation process completes. For partitioned tables, since the engine creates the ordered CCI partition by partition, users can still access the data in partitions where ordered CCI creation isn't in process. You can use this option to minimize the downtime during ordered CCI creation on large tables:
- Create partitions on the target large table (called Table_A).
- Create an empty ordered CCI table (called Table_B) with the same table and partition schema as Table A.
- Switch one partition from Table A to Table B.
- Run ALTER INDEX <Ordered_CCI_Index> ON <Table_B> REBUILD PARTITION = <Partition_ID> on Table B to rebuild the switched-in partition.
- Repeat step 3 and 4 for each partition in Table_A.
- Once all partitions are switched from Table_A to Table_B and have been rebuilt, drop Table_A, and rename Table_B to Table_A.
A. To check for ordered columns and order ordinal:
SELECT object_name(c.object_id) table_name, c.name column_name, i.column_store_order_ordinal FROM sys.index_columns i JOIN sys.columns c ON i.object_id = c.object_id AND c.column_id = i.column_id WHERE column_store_order_ordinal <>0
B. To change column ordinal, add or remove columns from the order list, or to change from CCI to ordered CCI:
CREATE CLUSTERED COLUMNSTORE INDEX InternetSales ON InternetSales ORDER (ProductKey, SalesAmount) WITH (DROP_EXISTING = ON)
For more development tips, see development overview.