Assess and correct clustered columnstore index health in a dedicated SQL pool
Applies to: Azure Synapse Analytics
This article introduces a slightly different approach to assessing clustered columnstore index (CCI) health. Follow the steps in the following sections or execute the steps in the notebook via Azure Data Studio.
Note
Before attempting to open this notebook, make sure that Azure Data Studio is installed on your local machine. To install it, go to Learn how to install Azure Data Studio.
In general, two major factors affect the quality of a CCI:
Compact rowgroups and metadata - The actual rowgroup count is close to the ideal count for the number of rows in the rowgroup.
Compressed rowgroups - Rowgroups are using columnstore compression.
Other conditions, such as small tables, over-partitioned tables, or under-partitioned tables, are arguably of poor quality or health. However, these conditions are better classified as design improvement opportunities that can be assessed in Step 4.
Step 1: Analyze a summary of your CCI health
Use the following query to get a single row of metrics.
WITH cci_detail AS (
SELECT t.object_id,
rg.partition_number,
COUNT(*) AS total_rowgroup_count,
SUM(CASE WHEN rg.state = 1 THEN 1 END) AS open_rowgroup_count,
CEILING((SUM(rg.[total_rows]) - SUM(rg.deleted_rows))/COUNT(DISTINCT rg.distribution_id)/1048576.) * COUNT(DISTINCT rg.distribution_id) AS [ideal_rowgroup_count],
SUM(rg.size_in_bytes/1024/1024.) AS size_in_mb,
SUM(CASE WHEN rg.state = 1 THEN rg.size_in_bytes END /1024/1024.) AS open_size_in_mb
FROM sys.pdw_nodes_column_store_row_groups rg
JOIN sys.pdw_nodes_tables nt ON rg.object_id = nt.object_id
AND rg.pdw_node_id = nt.pdw_node_id
AND rg.distribution_id = nt.distribution_id
JOIN sys.pdw_table_mappings mp ON nt.name = mp.physical_name
JOIN sys.tables t ON mp.object_id = t.object_id
GROUP BY t.object_id,
rg.partition_number
)
SELECT COUNT(DISTINCT object_id) AS tables_assessed_count,
COUNT(*) AS partitions_assessed_count,
SUM(total_rowgroup_count) AS actual_rowgroup_count,
SUM(ideal_rowgroup_count) AS ideal_rowgroup_count,
SUM(open_rowgroup_count) AS uncompressed_rowgroup_count,
CAST(SUM(size_in_mb) AS DECIMAL(19, 4)) AS actual_size_in_mb,
CAST(SUM(open_size_in_mb) AS DECIMAL(19, 4)) AS uncompressed_size_in_mb,
CAST(((SUM(total_rowgroup_count) - SUM(ideal_rowgroup_count)) / SUM(total_rowgroup_count)) * 100. AS DECIMAL(9, 4)) AS excess_pct,
CAST(((SUM(total_rowgroup_count) - SUM(ideal_rowgroup_count)) / SUM(total_rowgroup_count)) * 1. AS DECIMAL(9, 4)) * SUM(size_in_mb) AS excess_size_in_mb
FROM cci_detail
From the result, you can get an overview of the CCI health for your dedicated SQL pool. This information isn't directly actionable but helps you understand the importance of maintenance routines for achieving an ideal state.
| Column name | Description |
|---|---|
tables_assessed_count |
Count of CCI tables |
partitions_assessed_count |
Count of partitions Note: Non-partitioned tables will be counted as 1. |
actual_rowgroup_count |
Physical count of rowgroups |
ideal_rowgroup_count |
Calculated number of rowgroups that would be ideal for the number of rows |
uncompressed_rowgroup_count |
Number of rowgroups that contain uncompressed data. (Also known as: OPEN rows) |
actual_size_in_mb |
Physical size of CCI data in MB |
uncompressed_size_in_mb |
Physical size of uncompressed data in MB |
excess_pct |
Percent of rowgroups that could be further optimized |
excess_size_in_mb |
Estimated MB from unoptimized rowgroups |
Step 2: Analyze detailed CCI information
The following query provides a detailed report of which table partitions are candidates for rebuilding. CCI details are provided in three metrics that help identify and prioritize tables/partitions that would benefit most from maintenance. Set the appropriate threshold values for these metrics in the WHERE clause, and then in the ORDER BY clause, use the metrics that are of most interest to you. The detailed information can also be useful to determine whether your dedicated SQL pool is being impacted by a large number of small, fragmented tables, which can lead to delays in compilation.
Note
The commented fnMs_GenerateIndexMaintenanceScript function is a table-valued function (TVF) that can generate common scripts for maintaining indexes. If you want to get the maintenance scripts in the result, uncomment lines 37 and 39. And before you run the query, use the script in the section Generate index maintenance scripts to create the function. When running the maintenance script you get from the result, be sure to use an appropriately-sized resource class, such as largerc or xlargerc.
| Column name | Quality characteristic | Description |
|---|---|---|
excess_pct |
Compactness | Percentage of rowgroups that could be further compacted |
excess_size_in_mb |
Compactness | Estimated MB from unoptimized rowgroups |
OPEN_rowgroup_size_in_mb |
Compression | Actual MB of uncompressed data in the index |
WITH cci_info AS(
SELECT t.object_id AS [object_id],
MAX(schema_name(t.schema_id)) AS [schema_name],
MAX(t.name) AS [table_name],
rg.partition_number AS [partition_number],
COUNT(DISTINCT rg.distribution_id) AS [distribution_count],
SUM(rg.size_in_bytes/1024/1024) AS [size_in_mb],
SUM(rg.[total_rows]) AS [row_count_total],
COUNT(*) AS [total_rowgroup_count],
CEILING((SUM(rg.[total_rows]) - SUM(rg.[deleted_rows]))/COUNT(DISTINCT rg.distribution_id)/1048576.) * COUNT(DISTINCT rg.distribution_id) AS [ideal_rowgroup_count],
SUM(CASE WHEN rg.[State] = 1 THEN 1 ELSE 0 END) AS [OPEN_rowgroup_count],
SUM(CASE WHEN rg.[State] = 1 THEN rg.[total_rows] ELSE 0 END) AS [OPEN_rowgroup_rows],
CAST(SUM(CASE WHEN rg.[State] = 1 THEN rg.[size_in_bytes]/1024./1024. ELSE 0 END) AS DECIMAL(19, 4)) AS [OPEN_rowgroup_size_in_mb],
SUM(CASE WHEN rg.[State] = 2 THEN 1 ELSE 0 END) AS [CLOSED_rowgroup_count],
SUM(CASE WHEN rg.[State] = 2 THEN rg.[total_rows] ELSE 0 END) AS [CLOSED_rowgroup_rows],
CAST(SUM(CASE WHEN rg.[State] = 2 THEN rg.[size_in_bytes]/1024./1024. ELSE 0 END) AS DECIMAL(19, 4)) AS [CLOSED_size_in_mb],
SUM(CASE WHEN rg.[State] = 3 THEN 1 ELSE 0 END) AS [COMPRESSED_rowgroup_count],
SUM(CASE WHEN rg.[State] = 3 THEN rg.[total_rows] ELSE 0 END) AS [COMPRESSED_rowgroup_rows],
CAST(SUM(CASE WHEN rg.[State] = 3 THEN rg.[size_in_bytes]/1024./1024. ELSE 0 END) AS DECIMAL(19, 4)) AS [COMPRESSED_size_in_mb],
SUM(CASE WHEN rg.[State] = 3 THEN rg.[deleted_rows] ELSE 0 END) AS [COMPRESSED_rowgroup_rows_DELETED]
FROM sys.[pdw_nodes_column_store_row_groups] rg
JOIN sys.[pdw_nodes_tables] nt ON rg.[object_id] = nt.[object_id]
AND rg.[pdw_node_id] = nt.[pdw_node_id]
AND rg.[distribution_id] = nt.[distribution_id]
JOIN sys.[pdw_table_mappings] mp ON nt.[name] = mp.[physical_name]
JOIN sys.[tables] t ON mp.[object_id] = t.[object_id]
GROUP BY t.object_id,
rg.partition_number
)
, calc_excess AS(
SELECT *,
CAST(((total_rowgroup_count - ideal_rowgroup_count) / total_rowgroup_count) * 100. AS DECIMAL(9, 4)) AS [excess_pct],
CAST(((total_rowgroup_count - ideal_rowgroup_count) / total_rowgroup_count) * 1. AS DECIMAL(9, 4)) * size_in_mb AS [excess_size_in_mb]
FROM cci_info
)
SELECT calc_excess.*
-- , script.*
FROM calc_excess
-- CROSS APPLY dbo.fnMs_GenerateIndexMaintenanceScript(object_id, partition_number) AS script
WHERE -- set your own threshold(s) for the following; 0 is the ideal, but usually not practical
calc_excess.[excess_size_in_mb] > 300
OR calc_excess.excess_pct > 0.1
OR calc_excess.OPEN_rowgroup_size_in_mb > 100
ORDER BY calc_excess.[excess_size_in_mb] DESC;
Step 3: What to do when maintenance doesn't make the CCI health better
Performing maintenance on a table/partition may result in one of the following scenarios:
excess_pctorexcess_size_in_mbis larger than it was before maintenance.- The maintenance statement fails with insufficient memory.
Typical causes
- Insufficient resources.
- Insufficient service level (DWU).
- Table is large and not partitioned.
Recommended mitigations
- Increase resources for the maintenance statements by changing the executing user's resource class or workload group.
- Temporarily increase the DWU level to perform the maintenance.
- Implement a partitioning strategy for the problematic table and then perform maintenance on the partitions.
Step 4: Check for design improvement opportunities
Though not comprehensive, the following query can help you identify potential opportunities commonly found to cause performance or maintenance issues concerning CCIs.
| Opportunity title | Description | Recommendations |
|---|---|---|
| Small table | Table contains fewer than 15M rows | Consider changing the index from CCI to:
|
| Partitioning opportunity or under-partitioned table | Calculated ideal rowgroup count is greater than 180M (or ~188M rows) | Implement a partitioning strategy or change the existing partitioning strategy to reduce the number of rows per partition to less than 188M (approximately three row groups per partition per distribution) |
| Over-partitioned table | Table contains fewer than 15M rows for the largest partition | Consider:
|
WITH cci_info AS (
SELECT t.object_id AS [object_id],
MAX(SCHEMA_NAME(t.schema_id)) AS [schema_name],
MAX(t.name) AS [table_name],
rg.partition_number AS [partition_number],
SUM(rg.[total_rows]) AS [row_count_total],
CEILING((SUM(rg.[total_rows]) - SUM(rg.[deleted_rows]))/COUNT(DISTINCT rg.distribution_id)/1048576.) * COUNT(DISTINCT rg.distribution_id) AS [ideal_rowgroup_count]
FROM sys.[pdw_nodes_column_store_row_groups] rg
JOIN sys.[pdw_nodes_tables] nt ON rg.[object_id] = nt.[object_id]
AND rg.[pdw_node_id] = nt.[pdw_node_id]
AND rg.[distribution_id] = nt.[distribution_id]
JOIN sys.[pdw_table_mappings] mp ON nt.[name] = mp.[physical_name]
JOIN sys.[tables] t ON mp.[object_id] = t.[object_id]
GROUP BY t.object_id,
rg.partition_number
)
SELECT object_id,
MAX(SCHEMA_NAME),
MAX(TABLE_NAME),
COUNT(*) AS number_of_partitions,
MAX(row_count_total) AS max_partition_row_count,
MAX(ideal_rowgroup_count) partition_ideal_row_count,
CASE
-- non-partitioned tables
WHEN COUNT(*) = 1 AND MAX(row_count_total) < 15000000 THEN 'Small table'
WHEN COUNT(*) = 1 AND MAX(ideal_rowgroup_count) > 180 THEN 'Partitioning opportunity'
-- partitioned tables
WHEN COUNT(*) > 1 AND MAX(row_count_total) < 15000000 THEN 'Over-partitioned table'
WHEN COUNT(*) > 1 AND MAX(ideal_rowgroup_count) > 180 THEN 'Under-partitioned table'
END AS warning_category
FROM cci_info
GROUP BY object_id
Generate index maintenance scripts
Run the following query to create dbo.fnMs_GenerateIndexMaintenanceScript function on your dedicated SQL pool. This function generates scripts to optimize your CCI in three ways. You can use this function to maintain not only CCIs but also clustered (rowstore) indexes.
Parameters
| Parameter name | Required | Description |
|---|---|---|
@object_id |
Y | object_id of the table to target |
@partition_number |
Y | partition_number from sys.partitions to target. If the table isn't partitioned, specify 1. |
Output table
| Column name | Description |
|---|---|
rebuild_script |
Generated ALTER INDEX ALL ... REBUILD statement for the given table/partition. Non-partitioned heaps will return NULL. |
reorganize_script |
Generated ALTER INDEX ALL ... REORGANIZE statement for the given table/partition. Non-partitioned heaps will return NULL. |
partition_switch_script |
Applies only to partitioned tables; will be NULL if the table isn't partitioned or if an invalid partition number is specified. If the CCI was created with an ORDER clause, it will be rendered. |
CREATE FUNCTION dbo.fnMs_GenerateIndexMaintenanceScript (@object_id INT, @partition_number INT = 1)
RETURNS TABLE
AS
RETURN(
WITH base_info AS (
SELECT
t.object_id
, SCHEMA_NAME(t.schema_id) AS [schema_name]
, t.name AS table_name
, i.index_type
, i.index_cols
, i.index_type_desc
, tdp.distribution_policy_desc
, c.name hash_distribution_column_name
FROM sys.tables t
JOIN (
SELECT
i.object_id
, i.index_id
, MAX(i.type) AS index_type
, MAX(CASE WHEN i.type = 5 AND ic.column_store_order_ordinal != 0 THEN ' ORDER ' ELSE '' END)
+ '(' + STRING_AGG(
CASE
WHEN i.type IN (1, 5)
AND (ic.key_ordinal != 0 OR ic.column_store_order_ordinal != 0)
THEN c.name + CASE WHEN ic.is_descending_key = 1 THEN ' DESC' ELSE '' END
END
, ',') WITHIN GROUP(ORDER BY ic.column_store_order_ordinal, ic.key_ordinal) + ')' AS index_cols
, MAX(i.type_desc)
+ CASE
WHEN MAX(i.type) IN (1, 5) THEN ' INDEX'
ELSE ''
END COLLATE SQL_Latin1_General_CP1_CI_AS AS index_type_desc
FROM sys.indexes i
JOIN sys.index_columns ic ON i.object_id = ic.object_id AND i.index_id = ic.index_id
JOIN sys.columns c ON ic.object_id = c.object_id AND ic.column_id = c.column_id
WHERE i.index_id <= 1
GROUP BY i.object_id, i.index_id
) AS i
ON t.object_id = i.object_id
JOIN sys.pdw_table_distribution_properties tdp ON t.object_id = tdp.object_id
LEFT JOIN sys.pdw_column_distribution_properties cdp ON t.object_id = cdp.object_id AND cdp.distribution_ordinal = 1
LEFT JOIN sys.columns c ON cdp.object_id = c.object_id AND cdp.column_id = c.column_id
WHERE t.object_id = @object_id
)
, param_data_type AS (
SELECT
pp.function_id
, typ.name AS data_type_name
, CAST(CASE
WHEN typ.collation_name IS NOT NULL THEN 1
WHEN typ.name LIKE '%date%' THEN 1
WHEN typ.name = 'uniqueidentifier' THEN 1
ELSE 0
END AS BIT) AS use_quotes_on_values_flag
FROM sys.partition_parameters pp
JOIN sys.types typ ON pp.user_type_id = typ.user_type_id
)
, boundary AS (
SELECT
t.object_id
, c.name AS partition_column_name
, pf.boundary_value_on_right
, prv.boundary_id
, prv.boundary_id + CASE WHEN pf.boundary_value_on_right = 1 THEN 1 ELSE 0 END AS [partition_number]
, CASE
WHEN pdt.use_quotes_on_values_flag = 1 THEN '''' + CAST(
CASE pdt.data_type_name
WHEN 'date' THEN CONVERT(char(10), prv.value, 120)
WHEN 'smalldatetime' THEN CONVERT(VARCHAR, prv.value, 120)
WHEN 'datetime' THEN CONVERT(VARCHAR, prv.value, 121)
WHEN 'datetime2' THEN CONVERT(VARCHAR, prv.value, 121)
ELSE prv.value
END
AS VARCHAR(32)) + ''''
ELSE CAST(prv.value AS VARCHAR(32))
END AS boundary_value
FROM sys.tables t
JOIN sys.indexes i ON t.object_id = i.object_id AND i.index_id <= 1
JOIN sys.index_columns ic ON i.object_id = ic.object_id AND i.index_id = ic.index_id AND ic.partition_ordinal = 1
JOIN sys.columns c ON ic.object_id = c.object_id AND ic.column_id = c.column_id
JOIN sys.partition_schemes ps ON i.data_space_id = ps.data_space_id
JOIN sys.partition_functions pf ON ps.function_id = pf.function_id
JOIN param_data_type pdt ON pf.function_id = pdt.function_id
JOIN sys.partition_range_values prv ON pf.function_id = prv.function_id
WHERE t.object_id = @object_id
)
, partition_clause AS (
SELECT
object_id
, COUNT(*) - 1 -- should always be the 2nd to last partition in stage table
+ CASE WHEN MAX([partition_number]) = @partition_number THEN 1 ELSE 0 END -- except when last partition
AS [source_partition_number]
, 'WHERE ' + MAX(partition_column_name)
+ CASE WHEN MAX(CAST(boundary_value_on_right AS TINYINT)) = 1 THEN
' >= ' + MIN(CASE WHEN [partition_number] = @partition_number THEN boundary_value END)
ELSE
' <= ' + MAX(CASE WHEN [partition_number] = @partition_number THEN boundary_value END)
END
+ ' AND ' + MAX(partition_column_name)
+ CASE WHEN MAX(CAST(boundary_value_on_right AS TINYINT)) = 1 THEN
' < ' + MAX(boundary_value)
ELSE
' > ' + MIN(boundary_value)
END AS filter_clause
, ', PARTITION (' + MAX(partition_column_name) + ' RANGE '
+ CASE WHEN MAX(CAST(boundary_value_on_right AS TINYINT)) = 1 THEN 'RIGHT' ELSE 'LEFT' END
+ ' FOR VALUES(' + STRING_AGG(boundary_value, ',') + '))' AS [partition_clause]
FROM boundary
WHERE [partition_number] BETWEEN @partition_number - 1 AND @partition_number + 1
GROUP BY object_id
)
SELECT
CASE WHEN index_type IN (1, 5) THEN 'ALTER INDEX ALL ON [' + [schema_name] + '].[' + [table_name] + '] REBUILD'
+ CASE WHEN partition_clause.[object_id] IS NOT NULL THEN ' PARTITION = ' + CAST(@partition_number AS VARCHAR(16)) ELSE '' END + ';' END AS [rebuild_script]
, CASE WHEN index_type IN (1, 5) THEN 'ALTER INDEX ALL ON [' + [schema_name] + '].[' + [table_name] + '] REORGANIZE'
+ CASE WHEN partition_clause.[object_id] IS NOT NULL THEN ' PARTITION = ' + CAST(@partition_number AS VARCHAR(16)) ELSE '' END
+ CASE WHEN index_type = 5 THEN ' WITH (COMPRESS_ALL_ROW_GROUPS = ON)' ELSE '' END + ';' END AS [reorganize_script]
, 'CREATE TABLE [' + schema_name + '].[' + table_name + '_p' + CAST(@partition_number AS VARCHAR(16)) + '_tmp] WITH(' + index_type_desc + ISNULL(index_cols, '')
+ ', DISTRIBUTION = ' + distribution_policy_desc + CASE WHEN distribution_policy_desc = 'HASH' THEN '(' + hash_distribution_column_name + ')' ELSE '' END
+ partition_clause.partition_clause + ') AS SELECT * FROM [' + [schema_name] + '].[' + [table_name] + '] ' + filter_clause + CASE WHEN index_type = 5 AND index_cols IS NOT NULL THEN ' OPTION(MAXDOP 1)' ELSE '' END + ';'
+ ' ALTER TABLE [' + schema_name + '].[' + table_name + '_p' + CAST(@partition_number AS VARCHAR(16)) + '_tmp] SWITCH PARTITION ' + CAST(source_partition_number AS VARCHAR(16))
+ ' TO [' + [schema_name] + '].[' + [table_name] + '] PARTITION ' + CAST(@partition_number AS VARCHAR(16))
+ ' WITH (TRUNCATE_TARGET = ON);'
+ ' DROP TABLE [' + schema_name + '].[' + table_name + '_p' + CAST(@partition_number AS VARCHAR(16)) + '_tmp];' AS [partition_switch_script]
FROM base_info
LEFT JOIN partition_clause
ON base_info.object_id = partition_clause.object_id
);
GO
More information
To gain a more in-depth understanding and acquire extra assessment tools for CCI on the dedicated SQL pool, see: