Category: columnstore

This post contains all of the possible causes for delta store creation that I’ve found. I cannot say with certainty that it’s a complete list, but some of them may be new or unexpected to the reader.

Why Care about Delta Stores?

---

Microsoft and many others will be quick to tell you that loading data into CCIs is much faster when you can bypass the delta store. In SQL Server 2016 and beyond, delta stores are uncompressed rowstore mini-tables that serve as a temporary holding data until the data can be compressed into columnar format. They’re good when you have a trickle of data to load into a CCI, but bad in all possible ways for a data warehouse workload.

Reviewing the Documentation

---

I briefly reviewed the documentation written by Microsoft concerning the appearance of delta stores. Here’s a quote:

Rows go to the deltastore when they are:
Inserted with the INSERT INTO VALUES statement.
At the end of a bulk load and they number less than 102,400.
Updated. Each update is implemented as a delete and an insert.

There are also a few mentions of how partitioning can lead to the creation of multiple delta stores from a single insert. It seems as if the document is incomplete or a little misleading, but I admit that I didn’t exhaustively review everything. After all, Microsoft hides columnstore documentation all over the place.

Test Data

---

The source data for the CCI inserts is fairly uninteresting. I put four rowgroups worth of rows into a rowstore table with a BIGINT column and a randomly generated VARCHAR(16) value.

DROP TABLE IF EXISTS dbo.STAGING_TABLE;

CREATE TABLE dbo.STAGING_TABLE (
	ID BIGINT NOT NULL,
	STR1 VARCHAR(16) NOT NULL,
	PRIMARY KEY (ID)
);

INSERT INTO dbo.STAGING_TABLE WITH (TABLOCK)
SELECT TOP (4 * 1048576)
  ROW_NUMBER() OVER (ORDER BY (SELECT NULL))
, LEFT(CAST(NEWID() AS VARCHAR(36)), 16)
FROM master..spt_values t1
CROSS JOIN master..spt_values t2
OPTION (MAXDOP 1);

The columns for the table definition for the CCI were chosen to cover all of the demos except for the partitioning one. Your fact table definitions probably don’t look like this.

DROP TABLE IF EXISTS dbo.DELTA_STORE_DUMPING_GROUND;
CREATE TABLE dbo.DELTA_STORE_DUMPING_GROUND (
	ID BIGINT NULL,
	STR1 VARCHAR(100) NULL,
	STR2 VARCHAR(100) NULL,
	STR3 VARCHAR(100) NULL,
	STR1_MAX VARCHAR(MAX) NULL,
	INDEX CCI CLUSTERED COLUMNSTORE
);

Not Enough Rows For Bulk Load

---

The first reason for delta creation is well known and understood on SQL Server 2016. If you insert fewer than 102400 rows then SQL Server will not attempt to skip the delta store. This behavior is by design. The following query does not do a bulk load:

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND (ID)
SELECT TOP (102399) ID
FROM dbo.STAGING_TABLE
ORDER BY ID
OPTION (MAXDOP 1);

We can see the delta store that was just created with the following query:

SELECT
    row_group_id
  , state_desc
  , total_rows
--, trim_reason_desc
--, deleted_rows
--, partition_number
FROM sys.dm_db_column_store_row_group_physical_stats rg
INNER JOIN sys.tables t ON rg.OBJECT_ID = t.OBJECT_ID
WHERE t.name = 'DELTA_STORE_DUMPING_GROUND';

The results:

The other examples in this post use similar queries to get information about the newly added rowgroups to the table. They will be omitted for brevity. Simply inserting one row results in the delta store getting skipped:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND (ID)
SELECT TOP (102400) ID
FROM dbo.STAGING_TABLE
ORDER BY ID
OPTION (MAXDOP 1);

Now the rowgroup is compressed:

The rules change slightly in SQL Server 2017 with support of VARCHAR(MAX) and other LOB columns in columnstore. The delta store can be skipped with an insert of as few as 251 rows. Whether or not you write to the delta store depends on the amount of data being written. Below is one query that still writes to the delta store:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND (STR1_MAX)
SELECT TOP (251) REPLICATE(STR1, 40)
FROM dbo.STAGING_TABLE
ORDER BY ID
OPTION (MAXDOP 1);

Once again you can see the delta store:

Things are different if we increase the length of the inserted data. The query below writes to a compressed rowgroup and bypasses the delta store:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND (STR1_MAX)
SELECT TOP (251) REPLICATE(STR1, 500)
FROM dbo.STAGING_TABLE
ORDER BY ID
OPTION (MAXDOP 1);

The resulting rowgroup is compressed:

Removing just a single row from the insert brings us back to the delta store.

Inserting to Multiple Partitions

---

If a MAXDOP 1 INSERT query writes to multiple partitions then it could possibly write to multiple delta stores. The number of rows written to each partition is important as opposed to the total number of rows written to the table. Below I define a simple table with 2 partitions:

CREATE PARTITION FUNCTION CLUNKY_SYNTAX_1
(BIGINT)
AS RANGE LEFT
FOR VALUES (
  0
, 2000000
); 

CREATE PARTITION SCHEME CLUNKY_SYNTAX_2
AS PARTITION CLUNKY_SYNTAX_1
ALL TO ( [PRIMARY] );

DROP TABLE IF EXISTS dbo.PARTITIONED_DELTA_STORE;
CREATE TABLE dbo.PARTITIONED_DELTA_STORE (
ID BIGINT NULL,
INDEX CCI CLUSTERED COLUMNSTORE
) ON CLUNKY_SYNTAX_2 (ID);

The insert writes 200k rows to the CCI which you might expect to bypass the delta store, but since the rows are evenly spread over two partitions we end up with two delta stores:

INSERT INTO dbo.PARTITIONED_DELTA_STORE (ID)
SELECT ID
FROM dbo.STAGING_TABLE
WHERE ID BETWEEN 1900001 AND 2100000
OPTION (MAXDOP 1);

With MAXDOP 8 INSERT queries and the maximum number of partitions defined on a table, it is possible to get 120000 delta stores. I don’t recommend doing this.

Bulk Insert Leftovers

---

Often applications will not insert an exact multiple of 1048576 rows. That means that rows can be left over after a few rowgroups worth of inserted rows are compressed. Those leftover rows can go into a delta store. Consider the following insert query that inserts 100000 rows more than 1048576:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND (ID)
SELECT TOP (1048576 + 100000) ID
FROM dbo.STAGING_TABLE
ORDER BY ID
OPTION (MAXDOP 1);

As expected, the final result is one compressed rowgroup of 1048576 rows and one delta store of 100k rows.

If we inserted just a few thousand more rows than we’d end up with two compressed rowgroups.

Updates

---

UPDATE queries always write to the delta store. There are many other reasons to avoid UPDATES to CCIs if the application makes it possible to do so.

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND (ID)
SELECT TOP (1048576) ID
FROM dbo.STAGING_TABLE
ORDER BY ID
OPTION (MAXDOP 1);

At first there’s just a single compressed rowgroup:

Now run the UPDATE query and go make coffee:

UPDATE DELTA_STORE_DUMPING_GROUND
SET ID = ID;

Our table doesn’t look so hot:

In SQL Server 2016 the Tuple Mover will not clean up this table. Another row needs to be inserted into the table before the rowgroup is marked as CLOSED.

Parallel Insert

---

Many parallel queries have an element of randomess around how rows are distributed to parallel threads. Rows are not moved between threads after they flow to the part of the plan that performs the insert into the CCI. It’s possible to end up with a number of new delta stores equal to the number of parallel threads for the query. Let’s start with a parallel insert that moves 4 * 1048576 rows into the CCI:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND
WITH (TABLOCK) (ID)
SELECT ID
FROM dbo.STAGING_TABLE
OPTION (MAXDOP 4);

It’s possible to end up without any delta stores and the results of the query against sys.dm_db_column_store_row_group_physical_stats will vary, but generally you’ll get at least one:

If we have unnaturally high beauty standards for our rowgroups we can rewrite the query to effectively force rows to be evenly distributed on all threads. The query below does this with a join to a derived table:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND
WITH (TABLOCK) (ID)
SELECT stg.ID
FROM
(
	VALUES
	(0 * 1048576 + 1, 1 * 1048576),
	(1 * 1048576 + 1, 2 * 1048576),
	(2 * 1048576 + 1, 3 * 1048576),
	(3 * 1048576 + 1, 4 * 1048576)
)
v (start_id, end_id)
INNER JOIN dbo.STAGING_TABLE stg ON
	stg.ID BETWEEN v.start_id and v.end_id
OPTION (MAXDOP 4);

Perfection:

I know that you were looking forward to another image of a tiny table, but here’s the important part of the query plan for those who like that sort of thing:

Getting perfect rowgroups can also be accomplished by adding the TOP operator to the original query, but that adds a serial zone to the plan:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND
WITH (TABLOCK) (ID)
SELECT TOP (9999999999999999) stg.ID
FROM dbo.STAGING_TABLE stg
OPTION (MAXDOP 4);

The key here is the parallelism operator in the plan uses a round robin method for distributing rows:

Dictionary Pressure

---

In SQL Server 2016 the maximum size for a column dictionary is 16 MB. This limit is raised in SQL Server 2017 for VARCHAR(MAX) and similar columns. I’m not going to get into the details of dictionaries here but it suffices to say that columns with too many unique string columns can experience dictionary pressure. Dictionary pressure leads to compressed rows that are less than the perfect size of 1048576 rows. Let’s insert the STR1 column into the CCI this time:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND (STR1)
SELECT TOP (1048576) stg.STR1
FROM dbo.STAGING_TABLE stg
ORDER BY ID
OPTION (MAXDOP 1);

Due to dictionary pressure there’s now a delta store with about 73000 rows:

We can see that the dictionary size for the column is close to the limit with the query below:

SELECT csd.entry_count, csd.on_disk_size
FROM sys.column_store_dictionaries csd
INNER JOIN sys.partitions p
    ON csd.partition_id = p.partition_id
INNER JOIN sys.tables t
    ON p.OBJECT_ID = t.OBJECT_ID
WHERE t.name = 'DELTA_STORE_DUMPING_GROUND'
AND csd.column_id = 2;

Here are the results:

Rowgroup Memory Pressure

---

The memory grant for CCI compression for an INSERT is calculated based on DOP and column definitions of target columns in the target table. The memory grant can be insufficient to get a full 1048576 rows into a compressed rowgroup depending on the table definition and the characteristics of the data getting loaded into the table. Consider an example in which data is loaded into three columns of the CCI:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND
(STR1, STR2, STR3)
SELECT TOP (1048576)
  LEFT(STR1, 10)
, LEFT(STR1, 5)
, LEFT(STR1, 6)
FROM
dbo.STAGING_TABLE stg
ORDER BY ID
OPTION (MAXDOP 1);

With the above syntax the memory grant is calculated from just the STR1, STR2, and STR3 columns. The memory grant of 171152 KB isn’t enough to avoid a delta store:

Note that you may not see the same results on your machine due to the randomness of the source data. For my table and source data set, adding a single column and inserting NULL into it bumps the memory grant up enough to avoid memory pressure:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

ALTER TABLE dbo.DELTA_STORE_DUMPING_GROUND
ADD MORE_MEMORY_PLZ VARCHAR(1) NULL;

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND
(STR1, STR2, STR3, MORE_MEMORY_PLZ)
SELECT TOP (1048576)
  LEFT(STR1, 10)
, LEFT(STR1, 5)
, LEFT(STR1, 6)
, NULL
FROM
dbo.STAGING_TABLE stg
ORDER BY ID
OPTION (MAXDOP 1);

The compressed rowgroup contains 1048576 rows now that memory pressure has been addressed.

Cardinality Estimate Less Than 251 Rows

---

SQL Server won’t even ask for a memory grant if the cardinality estimate is less than 251 rows. Perhaps this is because the memory grant would be wasted unless at least 102400 rows were inserted into the table. There’s no second chance at a memory grant here, so it’s possible to insert millions of rows to delta stores. A TOP expression with a variable will default to a cardinality estimate of 100 rows, so this works nicely to show the behavior:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

DECLARE @top_rows BIGINT = 1048576;
INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND (ID)
SELECT TOP (@top_rows) ID
FROM dbo.STAGING_TABLE
ORDER BY ID
OPTION (MAXDOP 1);

Despite inserting 1048576 rows we aren’t able to bypass the delta store:

The same behavior can be observed with a cardinality estimate of 250 rows. The OPTIMIZE FOR query hint is used to control the estimate.

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

DECLARE @top_rows BIGINT = 1048576;
INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND (ID)
SELECT TOP (@top_rows) ID
FROM dbo.STAGING_TABLE
ORDER BY ID
OPTION (MAXDOP 1, OPTIMIZE FOR (@top_rows = 250));

However, if I bump up the estimate by one more row a memory grant is given to the query and the delta store is bypassed:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

DECLARE @top_rows BIGINT = 1048576;
INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND (ID)
SELECT TOP (@top_rows) ID
FROM dbo.STAGING_TABLE
ORDER BY ID
OPTION (MAXDOP 1, OPTIMIZE FOR (@top_rows = 251));

Under this scenario we’ve observed deadlocks when multiple sessions insert into delta stores from the same target table.

Extreme Server Memory Pressure

---

Memory grants for queries that insert into CCIs have a hardcoded timeout of 25 seconds. After 25 seconds they execute with required serial memory and always write to the delta store. In the query below I simulate memory pressure with a MAX_GRANT_PERCENT hint of 0:

TRUNCATE TABLE dbo.DELTA_STORE_DUMPING_GROUND;

INSERT INTO dbo.DELTA_STORE_DUMPING_GROUND (ID)
SELECT TOP (1048576) ID
FROM dbo.STAGING_TABLE
ORDER BY ID
OPTION (MAXDOP 1, MAX_GRANT_PERCENT = 0);

The query always writes to the delta store. It cannot compress data without a memory grant.

Under this scenario we’ve observed deadlocks when multiple sessions insert into delta stores from the same target table.

Final Thoughts

---

It took forever to do the formatting for this one, so I hope that someone finds it useful.

Going Further

---

If this is the kind of SQL Server stuff you love learning about, you’ll love my training. I’m offering a 75% discount to my blog readers if you click from here. I’m also available for consulting if you just don’t have time for that and need to solve performance problems quickly.

I loaded one trillion rows into a nonpartitioned table just to see what would happen. Spoiler: bad things happen.

Hardware and Table Plan

---

I did all of my testing on my home desktop which has an i5-4670 CPU (quad core), 5 GB of RAM for SQL Server, and a Samsung SSD 850 EVO 1 TB. Obviously not ideal to do testing like this but it was enough to get the job done. In order to fit a trillion rows into a table I had to use a CCI. Rowstore tables simply don’t offer good enough compression. For example, consider a page compressed heap with nothing but NULL values. One million rows takes up 10888 KB of disk space:

DROP TABLE IF EXISTS dbo.TEST_HEAP;
CREATE TABLE dbo.TEST_HEAP (
ID BIGINT
) WITH (DATA_COMPRESSION = PAGE);

INSERT INTO dbo.TEST_HEAP WITH (TABLOCK)
SELECT TOP (1000000) NULL
FROM master..spt_values t1
CROSS JOIN master..spt_values t2
OPTION (MAXDOP 1);

EXEC sp_spaceused 'TEST_HEAP';

Therefore, such a table with a trillion rows would require around 10 TB of disk space. That won’t fit on a 1 TB HDD, but the better compression of CCIs gives us some options. Ultimately I decided on building completely full rowgroups of 1048576 rows with each rowgroup only storing a single value. Estimating space for the final table is difficult, but we can hopefully get an upper bound using the following code:

DROP TABLE IF EXISTS dbo.TRIAL_BALLOON;
CREATE TABLE dbo.TRIAL_BALLOON (
ID BIGINT,
INDEX CCI_TRIAL_BALLOON CLUSTERED COLUMNSTORE
) 

INSERT INTO dbo.TRIAL_BALLOON WITH (TABLOCK)
SELECT TOP (10 * 1048576) NULL
FROM master..spt_values t1
CROSS JOIN master..spt_values t2
CROSS JOIN master..spt_values t3
OPTION (MAXDOP 1);

EXEC sp_spaceused 'TRIAL_BALLOON';

The table has 776 KB reserved space and 104 KB space for data. In the worst case we might expect the final table to require 75 GB space on disk.

Serial Population Strategy

---

My strategy for populating the table was to run the same code in four different SQL Server sessions. Each session grabs the next ID from a sequence and does a MAXDOP 1 insert of 1048576 rows into the CCI. With only four concurrent sessions I didn’t expect to run into any locking issues such as the mysterious ROWGROUP_FLUSH wait event. The sequence definition is about as simple as it gets:

DROP SEQUENCE IF EXISTS dbo.CCI_Sequence;
CREATE SEQUENCE dbo.CCI_Sequence AS BIGINT
START WITH 1
INCREMENT BY 1
NO CACHE;

Here’s the code that I used to add rowgroups to the table:

ALTER SERVER CONFIGURATION
SET PROCESS AFFINITY CPU=0;

DECLARE @i BIGINT

SET NOCOUNT ON;

IF EXISTS (SELECT 1 FROM ##stop_table)
BEGIN
	SET @i = 9999999999999;
END
ELSE
BEGIN
	SELECT @i = NEXT VALUE FOR dbo.CCI_Sequence;
END; 

WHILE @i <= 953674
BEGIN
	WITH NUM (n) AS (
	SELECT n
	FROM
	(
	VALUES
	 (@i),(@i),(@i),(@i),(@i),(@i),(@i),(@i)
	,(@i),(@i),(@i),(@i),(@i),(@i),(@i),(@i)
	,(@i),(@i),(@i),(@i),(@i),(@i),(@i),(@i)
	,(@i),(@i),(@i),(@i),(@i),(@i),(@i),(@i)
	) v(n)
	)
	INSERT INTO dbo.BIG_DATA
	SELECT n1.n
	FROM NUM n1
	CROSS JOIN NUM n2
	CROSS JOIN NUM n3
	CROSS JOIN NUM n4
	OPTION (MAXDOP 1);

	IF EXISTS (SELECT 1 FROM ##stop_table)
	BEGIN
		SET @i = 9999999999999;
	END
	ELSE
	BEGIN
		SELECT @i = NEXT VALUE FOR dbo.CCI_Sequence;
	END;
END;

The affinity stuff was to get the work spread out evenly over my four schedulers. Each session was assigned to a different CPU from 0-3. It was also important to run the following in a new session after all four sessions started working:

ALTER SERVER CONFIGURATION
SET PROCESS AFFINITY CPU=AUTO;

It wasn’t clear to me why that was needed to get good throughput. Perhaps part of the work of building a compressed rowgroup is offloaded to a system process?

The references to the ##stop_table temp table are just a way to pause the work as needed without skipping numbers in the sequence. I think that the code to generate 1048576 rows is fairly optimized. I did try to optimize it since this code was going to be run over 950000 times, but I still suspect that there was a better way to do it that I missed.

The jobs finished after about 2 days of running on 4 CPUs. That’s a rate of around 86 million rows per core per minute which I was pretty happy with. After the sessions finished I needed to do one final, very nervous, insert into the table:

INSERT INTO dbo.BIG_DATA
SELECT TOP (331776) 953675
FROM master..spt_values t1
CROSS JOIN master..spt_values t2
OPTION (MAXDOP 1);

Running sp_spaceused has never felt so satisfying:

Under a GB for one trillion rows. Not bad.

Parallel Population Strategy

---

Taking advantage of natural parallelism within SQL Server may also be a viable method to populate a trillion row table (especially on an actual server), but I didn’t test it fully. In order to get perfect rowgroups, you need round robin parallelism with a driving table that’s a multiple of the query’s MAXDOP. For example, here’s a suitable plan:

The Constant Scan contains exactly four rows because I’m running MAXDOP 4. The source CCI has exactly 1048576 rows. The parallel insert happens without a repartition streams so we end up with exactly 1048576 rows on each thread and in each compressed rowgroup. Below is one way to generate such a plan:

DROP TABLE IF EXISTS dbo.SOURCE_CCI;
CREATE TABLE dbo.SOURCE_CCI (
ID BIGINT,
INDEX CCI_SOURCE_CCI CLUSTERED COLUMNSTORE
);

INSERT INTO dbo.SOURCE_CCI WITH (TABLOCK)
SELECT TOP (1048576) 0
FROM master..spt_values t1
CROSS JOIN master..spt_values t2
OPTION (MAXDOP 1);

DROP TABLE IF EXISTS dbo.TRIAL_BALLOON;
CREATE TABLE dbo.TRIAL_BALLOON (
ID BIGINT,
INDEX CCI_TRIAL_BALLOON CLUSTERED COLUMNSTORE
);

INSERT INTO dbo.TRIAL_BALLOON WITH (TABLOCK)
SELECT  driver.n
FROM (
SELECT TOP (4) v.n
FROM (
	VALUES
		(1),(2),(3),(4)
	) v(n)
) driver
INNER JOIN dbo.SOURCE_CCI sc ON sc.ID < driver.n
OPTION (MAXDOP 4, NO_PERFORMANCE_SPOOL);

With nothing else running on my machine I estimate that this method would take about 2 days to complete. The problem is that if one core is busy with something else, such as watching terrible Youtube videos, then the entire insert could be slowed down.

Updating Stats

---

To do anything interesting on a table we want statistics. Gathering statistics for extremely compressed data can be challenging in SQL Server. That is because the target sampled rate is based on the total size of the table as opposed to the number of rows in the table. Consider an 8 billion row table built in the same way as the one trillion row table above. SQL Server generates the following query to gather sampled stats against the table:

SELECT StatMan([SC0])
FROM (
SELECT TOP 100 PERCENT [ID] AS [SC0]
FROM [dbo].[BIG_DATA] WITH (READUNCOMMITTED)
ORDER BY [SC0]
) AS _MS_UPDSTATS_TBL
OPTION (MAXDOP 1)

You may notice the lack of TABLESAMPLE as well as the MAXDOP 1 hint. To gather sampled stats SQL Server will get all eight billion rows from the table, sort them, and build the statistics object using the eight billion rows. On my machine, this took almost 3 hours to complete and tempdb grew to 85 GB.

There is a trick to get more reasonable sampled stats. All that’s needed is to increase the table size while keeping the same data. Soft deletion of compressed rowgroups is a good way to accomplish this. First find a data distribution that doesn’t compress well in CCIs. Here’s one example:

SELECT 9999999 + SUM(RN) OVER (ORDER BY RN)
FROM (
	SELECT TOP (1048576) ROW_NUMBER()
		OVER (ORDER BY (SELECT NULL)) RN
	FROM master..spt_values t1
	CROSS JOIN master..spt_values t2
) t
OPTION (MAXDOP 1);

Each rowgroup has a size of 8392 KB, so adding 100 rowgroups will add 839200 KB to the table. Deleting all of the newly added rows can take a little while and will log quite a bit to the transaction log, but the table size won’t change. Gathering sampled stats after the insert and delete took just a few seconds. The sample size was about 1% of the table. After a REORG the fully deleted rowgroups will be marked as TOMBSTONE and cleaned up by a background process.

For the one trillion row table I decided to roll the dice and go for sampled stats without any tricks. I gave tempdb a maximum size of 280 GB in order to not completely fill my hard drive. The stats update took 3 hours and 44 minutes. Surprisingly, the stat update grew tempdb to its maximum size but it didn’t fail. Perhaps I got very lucky. Here is the hard earned stats object:

Expected Query Performance

---

I expected reasonably fast query performance for queries designed to take advantage of rowgroup elimination. After all, the table was built in a way such that every compressed rowgroup only has a single value. I can get a count of rowgroups along with some other metadata in 20 seconds using the DMVs:

SELECT COUNT(*), MIN(css.min_data_id), MAX(css.max_data_id)
FROM sys.objects o
INNER JOIN sys.columns c ON o.object_id = c.object_id
INNER JOIN sys.partitions p ON o.object_id = p.object_id
INNER JOIN sys.column_store_segments css
    ON p.hobt_id = css.hobt_id
    AND css.column_id = c.column_id
WHERE o.name = 'BIG_DATA'
AND c.name = 'ID';

Getting the relevant segment_ids for a particular filter finishes in under a second:

SELECT segment_id
FROM sys.objects o
INNER JOIN sys.columns c ON o.object_id = c.object_id
INNER JOIN sys.partitions p ON o.object_id = p.object_id
INNER JOIN sys.column_store_segments css
    ON p.hobt_id = css.hobt_id
    AND css.column_id = c.column_id
WHERE o.name = 'BIG_DATA'
AND c.name = 'ID'
AND 500000 BETWEEN css.min_data_id AND css.max_data_id;

I’m dealing with DMVs so I would expect SQL Server to be able to do rowgroup elimination in a much more efficient way than the above queries. Therefore, 30 seconds seemed like a reasonable upper bound for the following query:

SELECT COUNT(*)
FROM dbo.BIG_DATA
WHERE ID = 500000;

The query takes over 15 minutes to complete despite reading only a single segment:

Table ‘BIG_DATA’. Scan count 4, logical reads 0, physical reads 0, read-ahead reads 0, lob logical reads 6, lob physical reads 1, lob read-ahead reads 0.
Table ‘BIG_DATA’. Segment reads 1, segment skipped 953674.

SQL Server Execution Times:
CPU time = 905328 ms, elapsed time = 917478 ms.

The Evil Wait Event

---

The query had a max wait time of 915628 ms for QUERY_TASK_ENQUEUE_MUTEX. This is suspiciously close to the elapsed time of 917478 ms. Unfortunately this is a very unpopular wait event in the industry. The wait event library has almost no information about it as well.

I call this the evil wait event because while it’s happening queries cannot be canceled through SSMS and many unrelated queries won’t even run. Most of the time no useful work can be done on the instance. I can’t read Russian so I’m not sure what the wait event is about. After I restarted SQL Server the wait event no longer appeared as consistently, but query performance did not improve as far as I could tell.

Other Queries

---

I ran a few other tests queries as well, although I was limited in what I could do by the evil wait event. The following query is the only one that I found that wasn’t affected:

SELECT TOP 1 ID
FROM dbo.BIG_DATA;

For reasons I don’t understand the query still took a long time:

Table ‘BIG_DATA’. Scan count 1, logical reads 0, physical reads 0, read-ahead reads 0, lob logical reads 6, lob physical reads 1, lob read-ahead reads 0.
Table ‘BIG_DATA’. Segment reads 1, segment skipped 0.

SQL Server Execution Times:
CPU time = 791625 ms, elapsed time = 811202 ms.

Counting the rows in the table took almost half an hour:

SELECT COUNT_BIG(*)
FROM dbo.BIG_DATA;

Statistics output:

Table ‘BIG_DATA’. Scan count 4, logical reads 0, physical reads 0, read-ahead reads 0, lob logical reads 5722050, lob physical reads 819345, lob read-ahead reads 0.

SQL Server Execution Times:
CPU time = 3734515 ms, elapsed time = 1635348 ms.

A query to sum every ID in the table took over 40 minutes:

SELECT SUM(ID)
FROM dbo.BIG_DATA;

Statistics output:

SQL Server Execution Times:
CPU time = 4810422 ms, elapsed time = 2433689 ms.

As expected performance gets worse without aggregate pushdown. Picking a simple query that’s not supported:

SELECT MAX(ID / 1)
FROM dbo.BIG_DATA;

This query took over an hour:

SQL Server Execution Times:
CPU time = 10218343 ms, elapsed time = 3755976 ms.

Queries against some of the CCIs DMVs don’t do very well either. A simple count took almost ten minutes:

SELECT COUNT(*)
FROM sys.dm_db_column_store_row_group_physical_stats
where OBJECT_ID = OBJECT_ID('BIG_DATA');

All of the work appears to be in the COLUMNSTORE_ROW_GROUPS table-valued function but I didn’t dig any more into it.

If you’re interested in how a query performs let me know in the comments. I will try anything that’s somewhat reasonable.

Final Thoughts

---

Now I can add working with trillion row tables to my resume. The compression for the one trillion row table was very impressive but everything else was decidedly less impressive. The very long QUERY_TASK_ENQUEUE_MUTEX wait times for nearly all queries were especially disappointing. I plan to do more testing at a later date with a partitioned table to see if that helps at all.

Thanks for reading!

Going Further

---

If this is the kind of SQL Server stuff you love learning about, you’ll love my training. I’m offering a 75% discount to my blog readers if you click from here. I’m also available for consulting if you just don’t have time for that and need to solve performance problems quickly.

Rowgroup elimination is a performance optimization based on compressed rowgroup metadata that can allow rowgroups to be skipped during query execution. It’s likely that all of the metadata used for the optimization is exposed in the sys.column_store_segments DMV. This blog post explores some of the less well known rules and limitations for rowgroup elimination.

Test Data

---

To keep things very simple we’ll build 100 rowgroups with exactly 1 million rows in each of them. ID and ID2 increase from 1 to 10000000 and ID_NULL is always NULL. Code to create and populate the table:

DROP TABLE IF EXISTS dbo.MILLIONAIRE_CCI;

CREATE TABLE dbo.MILLIONAIRE_CCI (
	ID BIGINT NULL,
	ID2 BIGINT NULL,
	ID_NULL BIGINT NULL,
	INDEX CCI_MILLIONAIRE_CCI CLUSTERED COLUMNSTORE
);

DECLARE @loop INT = 0;
BEGIN
	SET NOCOUNT ON;
	WHILE @loop < 100
	BEGIN
		INSERT INTO dbo.MILLIONAIRE_CCI WITH (TABLOCK)
		SELECT t.RN, t.RN, NULL
		FROM (
			SELECT TOP (1000000)
				(1000000 * @loop)
				+ ROW_NUMBER()
					OVER (ORDER BY (SELECT NULL)) RN
			FROM master..spt_values t1
			CROSS JOIN master..spt_values t2
			ORDER BY RN
		) t
		OPTION (MAXDOP 1);

		SET @loop = @loop + 1;
	END;
END;

We can expect very good rowgroup elimination on the ID and ID2 columns based on how we built them. That can be verified by calculating the REFF or by looking at sys.column_store_segments:

Code to generate the above result set:

SELECT css.min_data_id, css.max_data_id, css.has_nulls
FROM sys.objects o
INNER JOIN sys.columns c ON o.object_id = c.object_id
INNER JOIN sys.partitions p ON o.object_id = p.object_id
INNER JOIN sys.column_store_segments css
    ON p.hobt_id = css.hobt_id
    AND css.column_id = c.column_id
INNER JOIN sys.dm_db_column_store_row_group_physical_stats s
    ON o.object_id = s.object_id
    AND css.segment_id = s.row_group_id
    AND s.partition_number = p.partition_number
WHERE o.name = 'MILLIONAIRE_CCI'
AND c.name = 'ID'
AND s.[state] = 3
ORDER BY css.min_data_id, css.segment_id;

Many of the test queries below select a single aggregate value. This isn’t done for any special reason other than to limit the size of the result set. The easiest way to see how many rowgroups were skipped is to use SET STATISTICS IO ON and that requires that the results be returned to the client.

Single Column Filtering

---

Consider the following query:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID IN (1000000, 2000001);

Based on how we built the segments for the ID column we might expect that only two segments will need to be read: segment 1 with ID values of 1-1000000 and segment 3 with ID values of 2000001-3000000. As usual, SQL Server does not care about our expectations:

Table ‘MILLIONAIRE_CCI’. Segment reads 3, segment skipped 97.

Why did the storage engine scan two segments instead of three? Running another test makes the problem more clear:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID IN (1, 100000000);

For this query we end up scanning the entire table:

Table ‘MILLIONAIRE_CCI’. Segment reads 100, segment skipped 0.

It seems as if the query optimizer reduces the predicate against the filtered column to be a range of IDs. That range of IDs is used for rowgroup elimination. In some cases it’s possible to write a WHERE clause that won’t return any rows but still isn’t eligible for rowgroup elimination. The storage engine is not able to skip any segments while executing the below query:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID < 0 OR ID > 100000000;

There isn’t an issue when the where clause is filtering on a contiguous range. For example, the following query skips 98 segments as expected:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID BETWEEN 1 AND 2000000;

There also isn’t an issue when filtering down to multiple values as long as those values are sufficiently close together, as shown with the first example. I also wasn’t able to find any liminations around the number of values in the IN clause. The query below reads 1 segment and skips 99 as we might hope:

SELECT MAX(l.ID)
FROM dbo.MILLIONAIRE_CCI l
WHERE l.ID IN (
1, 2, 3, 4, 5, 6, 7, 8, 9, 10
, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20
, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30
, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40
, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50
, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60
, 61, 62, 63, 64
);

If we add one more filter value then the query optimizer changes the plan to use a join:

The above query is eligible for rowgroup elimination but it follows slightly different ruless as covered later in this post.

It is possible to disable the transformation to a join by using the undocumented query hint QueryRuleOff SelToLSJ. With 976 entries in the IN clause I still get rowgroup elimination as expected. With 977 entries nothing was pushed to the scan at all, and we get a truly horrible plan:

This doesn’t appear to be a columnstore limitation. The same behavior can be observed with a clusted rowstore index.

Getting back on track, the internal calculation around which rowgroups to skip isn’t always as simple as calculating the minimum and maximum in the range and using those values to do elimination. It’s possible to end up with no rowgroup elimination even when the maximum and minimum ID in the WHERE clause are close to each other. Consider the following query:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID BETWEEN 1 AND 2
OR ID BETWEEN 2 AND 3;

The storage engine only has to read a single segment. We can see in the query plan that the optimizer was able to simplify the expression into something that happens to qualify for rowgroup elimination:

Now consider the following query:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID BETWEEN 1 AND 2
OR ID BETWEEN 3 AND 4;

It would be helpful if the query optimizer changed the predicate to ID BETWEEN 1 AND 4 when doing calculations around which rowgroups can be skipped. This does not happen, and as a result all 100 rowgroups are scanned. Rowgroup elimination won’t be available when the WHERE clause is a sufficiently complicated mix of AND and OR conditions, even when filtering on just one column.

NULLs

---

Information about NULLs is stored internally and can be used for rowgroup elimination. SQL Server knows that none of the compressed segments for the ID column contain NULL, so the storage engine can skip all 100 segments for the following query:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID IS NULL;

Naturally, reversing the filter for this query will require the storage engine to scan the entire table.

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID IS NOT NULL;

We might expect that query to skip all segments if we change the filter column to ID_NULL. All rows in the rowgroups for ID_NULL are NULL and SQL Server ought to be aware of that fact. However, the storage engine still scans the entire table even for the query below:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID_NULL IS NOT NULL;

The DMV for ID_NULL doesn’t look as we might expect:

sys.column_store_dictionaries has a value of 0 for the entry_count column. It seems likely that the fact that the segments only contain NULL can be deduced from information already tracked by SQL Server. Rowgroup elimination for IS NOT NULL may have not been added because it was thought to be too unlikely of a use case.

Filters on Multiple Columns

---

To state it simply, rowgroup elimination can work quite well with AND predicates on different columns. It will not work with OR predicates on different columns unless the query optimizer can simplify the expression to something that’s eligible for rowgroup elimination.

The following queries are all able to skip 99 rowgroups:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID = 1 AND ID2 = 1;

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID BETWEEN 1 AND 2
AND ID2 BETWEEN 3 AND 4;

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID BETWEEN 1 AND 100000000
AND ID2 BETWEEN 1000001 AND 2000000;

This query skips all 100 rowgroups:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID = 1 AND ID2 = 1000001;

The storage engine doesn’t take the union of rowgroups that could be relevant. It instead takes the intersection, so adding AND predicates won’t increase the number of segments scanned, unless perhaps if you do something very unreasonable. The following query scans one rowgroup as expected:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID BETWEEN 1 AND 100000000
AND ID2 BETWEEN 1000001 AND 2000000
AND ID > ID2;

The final part of the WHERE clause is implemented in a filter operator. The rest of the WHERE clause remains eligible for rowgroup elimination.

Now let’s try a simple query with an OR predicate:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID = 1 OR ID2 = 1;

We might hope that the storage engine is able to deduce that only the first segment is relevant. Instead, rowgroup elimination isn’t even attempted. The predicate is implemented as a filter:

The only situation with OR filters that I’ve found to work with rowgroup elimination is when the optimizer can eliminate one of them. For example, the following query scans 5 segments because the optimizer is able to eliminate the condition on the ID2 column:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID IN (1, 5000000) OR ID2 BETWEEN 1 AND 0;

Joins

---

The query optimizer is able to copy predicates when filtering and joining on the same column. The copied predicates are eligible for rowgroup elimination. Consider the query below:

SELECT MAX(l1.ID)
FROM dbo.MILLIONAIRE_CCI l1
INNER JOIN dbo.MILLIONAIRE_CCI l2 ON l1.ID = l2.ID
WHERE l1.ID BETWEEN 1 AND 1000000;

Only two segments are read because the filter on ID can be applied to both sides of the join. The same behavior can be observed when forcing a merge join. Loop join is a bit different. As covered in the post on CCI string aggregation, rowgroup elimination does not occur on the inner side of a loop. Consider the following query:

SELECT MAX(l1.ID)
FROM dbo.MILLIONAIRE_CCI l1
INNER JOIN dbo.MILLIONAIRE_CCI l2 ON l1.ID = l2.ID
WHERE l1.ID BETWEEN 1 AND 1000
OPTION (LOOP JOIN, NO_PERFORMANCE_SPOOL);

The inner side is scanned 1000 times and the outer side is scanned once. The filter on ID allows all segments to be skipped besides one. So we should read 1001 segments and skip 1001 * 100 – 1001 = 99099 segments. This is what happens:

Table ‘MILLIONAIRE_CCI’. Segment reads 1001, segment skipped 99099.

More segments will be read depending on how many rowgroups the filter crosses. Suppose that we include rows with an ID that’s between 999501 and 1000500:

SELECT MAX(l1.ID)
FROM dbo.MILLIONAIRE_CCI l1
INNER JOIN dbo.MILLIONAIRE_CCI l2 ON l1.ID = l2.ID
WHERE l1.ID BETWEEN 999501 AND 1000500
OPTION (LOOP JOIN, NO_PERFORMANCE_SPOOL);

Now each scan on both the inner and outer side will need to read two segments:

Table ‘MILLIONAIRE_CCI’. Segment reads 2002, segment skipped 98098.

It’s possible to get rowgroup elimination even when filtering and joining on different columns. Consider the following query that joins on ID but filters on ID2:

SELECT MAX(l1.ID)
FROM dbo.MILLIONAIRE_CCI l1
INNER JOIN dbo.MILLIONAIRE_CCI l2 ON l1.ID = l2.ID
WHERE l1.ID2 BETWEEN 1 AND 1000000;

We still get rowgroup elimination against both sides of the join:

Table ‘MILLIONAIRE_CCI’. Segment reads 2, segment skipped 198.

The key is the optimized bitmap:

That allows rowgroup elimination to happen on both sides. Bitmap optimization can only occur with hash joins, so queries written in this way that do a merge or loop join won’t be able to take advantage of rowgroup elimination against both tables.

Less Reasonable Queries

---

Below is a set of sometimes unreasonable queries to test some of the limits around rowgroup elimilation. It was surprising how often the queries remained eligible for rowgroup elimination. For example, local variables seem to cause no issues, even without PEO. The following query reads just one segment:

DECLARE @ID_FILTER BIGINT = 1;
SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID = @ID_FILTER;

Data type conversions on the filtered expression don’t make the query ineligible for rowgroup elimination:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID = '1';

Casting on the filtered column is going to prevent rowgroup elimination. As will “optimizer tricks” like adding zero to the column:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID + 0 = 1;

We read all rowgroups:

Table ‘MILLIONAIRE_CCI’. Segment reads 100, segment skipped 0.

The query below is eligible for rowgroup elimination:

SELECT MAX(ID)
FROM dbo.MILLIONAIRE_CCI
WHERE ID <= CEILING(RAND());

Using scalar UDFs in queries is a terrible idea, but let’s create one for testing purposes:

CREATE OR ALTER FUNCTION dbo.CHEAP_UDF() RETURNS BIGINT
AS
BEGIN
	RETURN 1;
END;

As you might expect, the following query runs without parallelism and cannot skip any segments:

SELECT MAX(l.ID)
FROM dbo.MILLIONAIRE_CCI l
WHERE l.ID = dbo.CHEAP_UDF();

However, if we add SCHEMABINDING to the function definition then we get rowgroup elimination:

Table ‘MILLIONAIRE_CCI’. Segment reads 1, segment skipped 99.

The query below gets rowgroup elimination with and without SCHEMABINDING:

SELECT MAX(l.ID)
FROM dbo.MILLIONAIRE_CCI l
WHERE l.ID = (SELECT MAX(dbo.CHEAP_UDF()));

Query Rewrites for Better Rowgroup Elimination

---

In some cases it’s possible to rewrite queries to get better rowgroup elimination. This requires knowing your data and awareness of the rules around rowgroup elimination. Going back to an earlier example, the following query isn’t eligible for rowgroup elimination (without very convenient constraints):

SELECT *
FROM dbo.MILLIONAIRE_CCI
WHERE ID = 1 OR ID2 = 1;

It can be written to use UNION or UNION ALL. Here’s the UNION query:

SELECT *
FROM dbo.MILLIONAIRE_CCI
WHERE ID = 1

UNION 

SELECT *
FROM dbo.MILLIONAIRE_CCI
WHERE ID2 = 1;

Now the storage engine skips 198 segments and only reads 2:

Table ‘MILLIONAIRE_CCI’. Segment reads 2, segment skipped 198.

In some cases it may be advantageous to avoid the sort. The query below has the same rowgroup elimination:

SELECT *
FROM dbo.MILLIONAIRE_CCI
WHERE ID = 1

UNION ALL

SELECT *
FROM dbo.MILLIONAIRE_CCI
WHERE ID2 = 1 AND ID <> 1;

Here’s the query plan:

Consider another query with a wide range of values in the IN clause, but filtered against a single column. The query below won’t be able to skip any rowgroups because we’re including the minimum and maximum value of ID in the query’s results:

SELECT *
FROM dbo.MILLIONAIRE_CCI
WHERE ID IN (
1, 2, 3, 4, 5, 6, 7, 8, 9, 10
, 100000000
);

It may be impractical to write out the query using UNION. Instead, we can essentially force a join by putting the filter values into a derived table. The optimizer is likely to move the IN list to a constant scan and do a hash join to the CCI. We can get rowgroup elimination through the bitmap filter on the hash join. Here’s one way to rewrite the query:

SELECT c.*
FROM
(
	VALUES (1), (2), (3), (4), (5)
	, (6), (7), (8), (9), (10)
	, (100000000)
) v(x)
INNER JOIN dbo.MILLIONAIRE_CCI c ON c.ID = v.x;

Here’s the plan:

As expected, we only need to scan 2 rowgroups:

Table ‘MILLIONAIRE_CCI’. Segment reads 2, segment skipped 98.

SQL Server 2017 Changes

---

I ran all of the test queries against SQL Server 2017 RC2. I was not able to observe any differences. It may be that Microsoft did not choose to make improvements in this area, or any improvements were missed by my test cases.

Final Thoughts

---

Rowgroup elimination seems designed to reduce IO requirements for queries that filter against contiguous ranges against a column, like filtering against a single month of data from a table, or when joining to the CCI through a hash join. It’s possible to write queries for which rowgroup elimination does not occur, even though SQL Server in theory has all of the information that it would need to perform rowgroup elimination. From a practical point of the view, the biggest limitation is probably around OR logic.

Thanks for reading!

Going Further

---

If this is the kind of SQL Server stuff you love learning about, you’ll love my training. I’m offering a 75% discount to my blog readers if you click from here. I’m also available for consulting if you just don’t have time for that and need to solve performance problems quickly.

Aggregate pushdown is an optimization for aggregate queries against columnstore tables that was introduced in SQL Server 2016. Some aggregate computations can be pushed to the scan node instead of calculated in an aggregate node. I found the documentation to be a little light on details so I tried to find as many restrictions around the functionality as I could through testing.

Test Data

---

Most of my testing was done against a simple CCI with a single compressed rowgroup:

DROP TABLE IF EXISTS dbo.AP_1_RG;
CREATE TABLE dbo.AP_1_RG (
	ID1 bigint NULL,
	ID2 bigint NULL,
	AGG_COLUMN BIGINT NOT NULL,
	INDEX CCI CLUSTERED COLUMNSTORE
);

INSERT INTO dbo.AP_1_RG WITH (TABLOCK)
SELECT
t.RN % 8000
, t.RN % 8000
, 0
FROM
(
	SELECT TOP (1048576)
	ROW_NUMBER() OVER
		(ORDER BY (SELECT NULL)) RN
	FROM master..spt_values t1
	CROSS JOIN master..spt_values t2
) t
OPTION (MAXDOP 1);

I found this table structure and data to be convenient for most of the test cases, but all tests can be reproduced with different data or a different table structure.

Restrictions without GROUP BY

---

Aggregate pushdown is supported both with and without a GROUP BY clause. I found it convenient to test those two cases separately. Below is a list of restrictions that I found or verified during testing.

Data Type of Aggregate Column

The documentation says:

Any datatype <= 64 bits is supported. For example, bigint is supported as its size is 8 bytes but decimal (38,6) is not because its size is 17 bytes. Also, no string types are supported.

However, this isn’t quite accurate. Float data types are not supported. numeric(10,0) is supported despite requiring 9 bytes for storage. Here’s a full table of results:

* bit is not supported for aggregates in general
** not all data types were tested

I would summarize support as all date and time data types are supported except datetimeoffset. All exact numeric data types are supported if they are under 10 bytes. Approximate numerics, strings, and other data types are not supported.

Data Type Conversions

Aggregate pushdown does not appear to be supported if there is any data type conversion in any part of the aggregate expression. Both implicit and explicit data type conversions can cause issues, although it ultimately depends on the rules for data type precedence and if the query optimizer determines if a conversion is needed. For example, the following queries are eligible for pushdown:

SELECT MAX(ID1 + 1)
FROM dbo.AP_1_RG;

SELECT MAX(ID1 + CAST(1 AS BIGINT))
FROM dbo.AP_1_RG;

SELECT MAX(ID1 + CAST(1 AS INT))
FROM dbo.AP_1_RG;

SELECT SUM(ID1 + 1)
FROM dbo.AP_1_RG;

SELECT SUM(1 * ID1)
FROM dbo.AP_1_RG;

SELECT MAX(CAST(ID1 AS BIGINT))
FROM dbo.AP_1_RG;

However, the following queries are not:

SELECT MAX(CAST(ID1 AS INT))
FROM dbo.AP_1_RG;

SELECT SUM(1.5 * ID1)
FROM dbo.AP_1_RG;

SELECT SUM(ID1 + CAST(1 AS BIGINT))
FROM dbo.AP_1_RG;

Sometimes the compute scalar appears with the conversion even when we might not expect it, like for the last query:

Unsupported Operators

Division and modulus prevent aggregate pushdown even when they wouldn’t change the result or if there isn’t a data type conversion in the plan. There are likely other unsupported operators as well. The following queries are eligible for pushdown:

SELECT MAX(ID1 + 1)
FROM dbo.AP_1_RG;

SELECT MAX(ID1 - 1)
FROM dbo.AP_1_RG;

SELECT MAX(ID1 * 2)
FROM dbo.AP_1_RG;

The following queries are not eligible for pushdown:

SELECT MAX(ID1 / 1)
FROM dbo.AP_1_RG;

SELECT MAX(ID1 % 9999999999999)
FROM dbo.AP_1_RG;

Aggregate Cannot be Applied to Scan

The aggregate expression must be applied directly to the scan. If there’s a filter between the aggregate and the scan then it won’t work. A compute scalar node between the aggregate and the scan can be okay.

Filter expressions involving OR on different columns tend to be calculated as a filter. This means that the following query isn’t eligible for pushdown:

SELECT MIN(ID1)
FROM dbo.AP_1_RG
WHERE ID1 > 0 OR ID2 > 0;

It’s likely that this restriction will affect many queries with joins to other tables.

Local Variables Without PEO

Aggregate pushdown is not available if there is a local variable in the aggregate expression unless the optimizer is able to embed the literal parameter value in the query, such as with a RECOMPILE hint. For example, the first query is not eligible for pushdown but the second query is:

DECLARE @var BIGINT = 0;
-- no
SELECT MIN(ID1 + @var)
FROM dbo.AP_1_RG;

-- yes
SELECT MIN(ID1 + @var)
FROM dbo.AP_1_RG
OPTION (RECOMPILE);

Trivial Plans

Simple queries that use SUM, AVG, COUNT, or COUNT_BIG against very small tables may get a trivial plan. In SQL Server 2016 that trivial plan will not be eligible for batch mode so aggregate pushdown will not occur. Consider the following CCI with 10000 rows in a compressed rowgroup:

DROP TABLE IF EXISTS AGG_PUSHDOWN_FEW_ROWS;

CREATE TABLE dbo.AGG_PUSHDOWN_FEW_ROWS (
ID BIGINT NOT NULL,
INDEX CCI2 CLUSTERED COLUMNSTORE
);

INSERT INTO dbo.AGG_PUSHDOWN_FEW_ROWS WITH (TABLOCK)
SELECT t.RN
FROM
(
	SELECT TOP (10000) ROW_NUMBER()
		OVER (ORDER BY (SELECT NULL)) RN
	FROM master..spt_values t1
	CROSS JOIN master..spt_values t2
	CROSS JOIN master..spt_values t3
) t
OPTION (MAXDOP 1);

ALTER INDEX CCI2 ON AGG_PUSHDOWN_FEW_ROWS
REORGANIZE WITH (COMPRESS_ALL_ROW_GROUPS = ON);

With the default cost threshold for parallelism value of 5 I get a trivial plan:

If I decrease the CTFP value to 0 I get batch mode along with aggregate pushdown. As far as I can tell, queries with MAX or MIN do not have this issue.

Lack of Hash Match

For some queries the query optimizer may cost a stream aggregate as a cheaper alternative to a hash match aggregate. Aggregate pushdown is not available with a stream aggregate. If I truncate the AGG_PUSHDOWN_FEW_ROWS table and load 3002 rows into it I get a stream aggregate and no pushdown. With 3003 rows I get a hash match and pushdown. The tipping point depends on the aggregate function, MAXDOP, the data loaded into the table, and so on. The undocumented query hint QUERYRULEOFF GbAggToStrm can be used to encourage a hash match aggregate. If the aggregated column is a string this gets more complicated.

Non-trivial CASE statements

Trivial CASE statements inside the aggregate are still eligible for pushdown:

SELECT SUM(CASE WHEN 1 = 1 THEN ID1 ELSE ID2 END)
FROM dbo.AP_1_RG;

However, many other CASE statements are not. Here is one example:

SELECT SUM(CASE WHEN ID1 < ID2 THEN ID1 ELSE ID2 END)
FROM dbo.AP_1_RG;

Basic Restrictions

For completeness I’ll list a few more of the more obvious restrictions on pushdown. Many of these are documented by Microsoft. The CCI must contain at least one compressed rowgroup. Delta stores are not eligible for pushdown. There are only six aggregate functions supported: COUNT, COUNT_BIG, MIN, MAX, SUM, and AVG. COUNT (DISTINCT col_name) is not supported.

If a query does not get batch mode due to TF 9453 or for other reasons it will not get aggregate pushdown. Undocumented trace flag 9354 directly disables aggregate pushdown.

Restrictions with GROUP BY

---

Queries with a GROUP BY have many, if not, all of the same restrictions on the aggregate expressions. As far as I can tell there are no restrictions on the data type of the GROUP BY columns. More than one GROUP BY column is supported as well. However, there are a few restrictions which only apply to queries with a GROUP BY clause.

Non-direct Column References

The columns in the GROUP BY need to be columns. Adding scalars and other nonsense appears to make the query ineligible. The following queries are not eligible:

SELECT ID1 + 0, SUM(AGG_COLUMN)
FROM dbo.AP_1_RG
GROUP BY ID1 + 0;

SELECT ID1 + ID2, SUM(AGG_COLUMN)
FROM dbo.AP_1_RG
GROUP BY ID1 + ID2;

This one is eligible:

SELECT ID1, ID2, SUM(AGG_COLUMN)
FROM dbo.AP_1_RG
GROUP BY ID1, ID2;

Segment Not Compressed Enough?

A rowgroup appears to be ineligible for aggregate pushdown if the GROUP BY column has a segment size which is too large. This can be observed with the following test data:

DROP TABLE IF EXISTS AP_3_RG;
CREATE TABLE dbo.AP_3_RG (
ID1 bigint NULL,
AGG_COLUMN BIGINT NOT NULL,
INDEX CCI CLUSTERED COLUMNSTORE
);

INSERT INTO dbo.AP_3_RG WITH (TABLOCK)
SELECT
t.RN % 16000
, 0
FROM
(
	SELECT TOP (1 * 1048576) ROW_NUMBER()
		OVER (ORDER BY (SELECT NULL)) RN
	FROM master..spt_values t1
	CROSS JOIN master..spt_values t2
	CROSS JOIN master..spt_values t3
) t
OPTION (MAXDOP 1);

INSERT INTO dbo.AP_3_RG WITH (TABLOCK)
SELECT
t.RN % 17000
, 0
FROM
(
	SELECT TOP (1 * 1048576) ROW_NUMBER()
		OVER (ORDER BY (SELECT NULL)) RN
	FROM master..spt_values t1
	CROSS JOIN master..spt_values t2
	CROSS JOIN master..spt_values t3
) t
OPTION (MAXDOP 1);

INSERT INTO dbo.AP_3_RG WITH (TABLOCK)
SELECT
t.RN % 16000
, 0
FROM
(
	SELECT TOP (1 * 1048576) ROW_NUMBER()
		OVER (ORDER BY (SELECT NULL)) RN
	FROM master..spt_values t1
	CROSS JOIN master..spt_values t2
	CROSS JOIN master..spt_values t3
) t
OPTION (MAXDOP 1);

Rows from the second rowgroup are not locally aggregated for the following query:

SELECT ID1, SUM(AGG_COLUMN)
FROM AP_3_RG
GROUP BY ID1
OPTION (MAXDOP 1);

The following query has 2097152 locally aggregated rows which correspond to the first and third rowgroups. From the actual plan:

Poking around in the sys.column_store_segments and sys.column_store_dictionaries DMVs doesn’t reveal any interesting differences other than the rowgroup with more distinct values has a much larger size (2097736 bytes versus 128576 bytes ). We can go deeper with the undocumented DBCC CSINDEX:

DBCC TRACEON (3604);

DBCC CSINDEX (
7, -- DB_ID
72057613148028928, -- hobt_id
2, -- 1 + column_id
0, -- segment_id
1, -- 1 for segment
0 -- print option
);

Among other differences, for the 16000 distinct value segment we see:

Bitpack Data Header:

Bitpack Entry Size = 16
Bitpack Unit Count = 0
Bitpack MinId = 3
Bitpack DataSize = 0

But for the 17000 distinct value segment we see:

Bitpack Data Header:
Bitpack Entry Size = 16
Bitpack Unit Count = 262144
Bitpack MinId = 3
Bitpack DataSize = 2097152

Perhaps bitpack compressed data is not eligible for aggregate pushdown?

It’s important to note that segments are not compressed independently in SQL Server, despite the data being stored at a column level. For the AP_1_RG table we’re eligible for pushdown when aggregating by ID1 or ID2. However, if I truncate the table and change the data slightly:

TRUNCATE TABLE dbo.AP_1_RG;

INSERT INTO dbo.AP_1_RG WITH (TABLOCK)
SELECT
t.RN % 8000
, t.RN % 8001 -- was previously 8000
, 0
FROM
(
	SELECT TOP (1048576) ROW_NUMBER()
		OVER (ORDER BY (SELECT NULL)) RN
	FROM master..spt_values t1
	CROSS JOIN master..spt_values t2
	CROSS JOIN master..spt_values t3
) t
OPTION (MAXDOP 1);

Now both columns are no longer eligible for aggregate pushdown. A REBUILD operation on the table does not help. Tricking SQL Server into assigning more memory to the columnstore compression also does not help.

Pushdown Surprises

---

During testing I was surprised by a few queries that supported aggregate pushdown. The most surprising was that the following query can get aggregate pushdown:

SELECT MAX(ID + ID2)
FROM dbo.AP_1_RG;

I have no idea how it works, but it does. For a few others, TABLESAMPLE does not prevent aggregate pushdown from happening. In addition, GROUP BY CUBE, GROUPING SETS, and ROLLUP are supported as well.

Changes with SQL Server 2017

---

I ran the same series of tests on SQL Server 2017 RC1. The only difference I observed was that I could no longer get a trivial plan without batch mode aggregation. This change was announced here by Microsoft.

Final Thoughts

---

As you can see, there are many undocumented restrictions around aggregate pushdown. These limits may be changed or go away as Microsoft continues to update SQL Server. Pushdown with GROUP BY is supported for some queries against some tables, but eligibility appears to be based on how the data is compressed, which cannot be predicted ahead of time. In my opinion, this makes it rather difficult to count on in practice.

Thanks for reading!

Going Further

---

If this is the kind of SQL Server stuff you love learning about, you’ll love my training. I’m offering a 75% discount to my blog readers if you click from here. I’m also available for consulting if you just don’t have time for that and need to solve performance problems quickly.

This is part 1 of a series on columnstore index partitioning. The focus of this post is on rowgroup elimination.

Defining Fragmentation for CCIs

---

There are many different ways that a CCI can be fragmented. Microsoft considers having multiple delta rowgroups, deleted rows in compressed rowgroups, and compressed rowgroups less than the maximum size all as signs of fragmentation. I’m going to propose yet another way. A CCI can also be considered fragmented if the segments for a column that is very important for rowgroup elimination are packed in such a way that limits rowgroup elimination. That sentence was a bit much but the following table examples should illustrate the point.

The Test Data

---

First we’ll create a single column CCI with integers from 1 to 104857600. I’ve written the query so that the integers will be loaded in order. This is more commonly done using a loop and I’m relying slightly on undocumented behavior, but it worked fine in a few tests that I did:

DROP TABLE IF EXISTS dbo.LUCKY_CCI;

CREATE TABLE dbo.LUCKY_CCI (
	ID BIGINT NOT NULL,
	INDEX CCI CLUSTERED COLUMNSTORE
);

INSERT INTO dbo.LUCKY_CCI WITH (TABLOCK)
SELECT TOP (100 * 1048576) ROW_NUMBER()
	OVER (ORDER BY (SELECT NULL)) RN
FROM master..spt_values t1
CROSS JOIN master..spt_values t2
CROSS JOIN master..spt_values t3
ORDER BY RN
OPTION (MAXDOP 1);

On my machine, the table takes less than 40 seconds to populate. We can see that the minimum and maximum IDs for each rowgroup are nice and neat from the sys.column_store_segments DMV:

This table does not have any rowgroup elimination fragmentation. The segments are optimally constructed from a rowgroup elimination point of view. SQL Server will be able to skip every rowgroup except for one or two if I filter on a sequential range of one million integers:

SELECT COUNT(*)
FROM dbo.LUCKY_CCI
WHERE ID BETWEEN 12345678 AND 13345677;

STATISTICS IO output:

Table ‘LUCKY_CCI’. Segment reads 2, segment skipped 98.

Erik Darling was the inspiration for this next table, since he can never manage to get rowgroup elimination in his queries. The code below is designed to create segments that allow for pretty much no rowgroup elimination whatsoever:

DROP TABLE IF EXISTS dbo.UNLUCKY_CCI;

CREATE TABLE dbo.UNLUCKY_CCI (
	ID BIGINT NOT NULL,
	INDEX CCI CLUSTERED COLUMNSTORE
);

INSERT INTO dbo.UNLUCKY_CCI WITH (TABLOCK)
SELECT rg.RN
FROM
(
	SELECT TOP (100) ROW_NUMBER()
		OVER (ORDER BY (SELECT NULL)) RN
	FROM master..spt_values t1
) driver
CROSS APPLY (
	SELECT driver.RN

	UNION ALL

	SELECT 104857601 - driver.RN

	UNION ALL

	SELECT TOP (1048574) 100 + 1048574 * (driver.RN - 1)
		+ ROW_NUMBER() OVER (ORDER BY (SELECT NULL))
	FROM master..spt_values t2
	CROSS JOIN master..spt_values t3
) rg
ORDER BY driver.RN
OPTION (MAXDOP 1, NO_PERFORMANCE_SPOOL);

Each rowgroup will contain an ID near the minimum (1 – 100) and an ID near the maximum (104857301 – 104857400). We can see this with the same DMV as before:

This table has pretty much the worst possible rowgroup elimination fragmentation. SQL Server will not be able to skip any rowgroups when querying almost any sequential range of integers:

SELECT COUNT(*)
FROM dbo.UNLUCKY_CCI
WHERE ID BETWEEN 12345678 AND 13345677;

STATISTICS IO output:

Table ‘UNLUCKY_CCI’. Segment reads 100, segment skipped 0.

Measuring Rowgroup Elimination Fragmentation

---

It might be useful to develop a more granular measurement for rowgroup elimination instead of just “best” and “worst”. One way to approach this problem is to define a batch size in rows for a filter against the column and to estimate how many total table scans would need to be done if the entire table was read in a series of SELECT queries with each processing a batch of rows. For example, with a batch size of 1048576 query, 1 would include the IDs from 1 to 1048576, query 2 would include IDs from 1 + 1048576 to 1048576 * 2, and so on. I will call this number the rowgroup elimination fragmentation factor, or REFF for short. It roughly represents how many extra rowgroups are read that could have been skipped with less fragmentation. Somewhat-tested code to calculate the REFF is below:

SET TRANSACTION ISOLATION LEVEL READ UNCOMMITTED;

DECLARE @batch_size INT = 1048576;
DECLARE @global_min_id BIGINT;
DECLARE @global_max_id BIGINT;
DECLARE @segments BIGINT;

DROP TABLE IF EXISTS #MIN_AND_MAX_IDS;
CREATE TABLE #MIN_AND_MAX_IDS (
	min_data_id BIGINT,
	max_data_id BIGINT
);
CREATE CLUSTERED INDEX c ON #MIN_AND_MAX_IDS
(min_data_id, max_data_id);

INSERT INTO #MIN_AND_MAX_IDS
SELECT css.min_data_id, css.max_data_id
FROM sys.objects o
INNER JOIN sys.columns c ON o.object_id = c.object_id
INNER JOIN sys.partitions p ON o.object_id = p.object_id
INNER JOIN sys.column_store_segments css
	ON p.hobt_id = css.hobt_id
	AND css.column_id = c.column_id
INNER JOIN sys.dm_db_column_store_row_group_physical_stats s
	ON o.object_id = s.object_id
	AND css.segment_id = s.row_group_id
	AND s.partition_number = p.partition_number
WHERE o.name = 'LUCKY_CCI'
AND c.name = 'ID'
AND s.[state] = 3;

SET @segments = @@ROWCOUNT;

SELECT
  @global_min_id = MIN(min_data_id)
, @global_max_id = MAX(max_data_id)
FROM #MIN_AND_MAX_IDS;

DROP TABLE IF EXISTS #BUCKET_RANGES;
CREATE TABLE #BUCKET_RANGES (
	min_data_id BIGINT,
	max_data_id BIGINT
);

 -- allows up to 6.4 million pieces
INSERT INTO #BUCKET_RANGES
SELECT
  @global_min_id + (RN - 1) * @batch_size
, @global_min_id + RN * @batch_size - 1
FROM
(
	SELECT ROW_NUMBER() OVER
	(ORDER BY (SELECT NULL)) RN
	FROM master..spt_values t1
	CROSS JOIN master..spt_values t2
) t
WHERE -- number of buckets
t.RN <= CEILING((1.0 * @global_max_id
	- (@global_min_id)) / @batch_size);

 -- number of total full scans
SELECT 1.0 * SUM(cnt) / @segments
FROM (
	SELECT COUNT(*) cnt
	FROM #BUCKET_RANGES br
	LEFT OUTER JOIN #MIN_AND_MAX_IDS m
		ON m.min_data_id <= br.max_data_id AND m.max_data_id >= br.min_data_id
	GROUP BY br.min_data_id
) t;

For the LUCKY_CCI table, if we pick a batch size of 1048576 we get a REFF of 1.0. This is because each SELECT query would read exactly 1 rowgroup and skip the other 99 rowgroups. Conversely, the UNLUCKY_CCI table has a REFF of 100.0. This is because every SELECT query would need to read all 100 rowgroups just to return 1048576 rows. In other words, each query does a full scan of the table and there are 100 rowgroups in the table.

What about DUIs?

---

Suppose our lucky table runs out of luck and starts getting DUIs. As rows are deleted and reinserted they will be packed into new rowgroups. We aren’t likely to keep our perfect rowgroups for every long. Below is code that randomly deletes and inserts a range of 100k integers. There is some bias in which numbers are processed but that’s ok for this type of testing. I ran the code with 600 loops which means that up to 60% of the rows in the table could have been moved around. There’s nothing special about 60%. That just represents my patience with the process.

DECLARE @target_loops INT = 600,
@batch_size INT = 100001,
@current_loop INT = 1,
@middle_num BIGINT,
@rows_in_target BIGINT;

BEGIN

SET NOCOUNT ON;

SELECT @rows_in_target = COUNT(*)
FROM dbo.LUCKY_CCI;

DROP TABLE IF EXISTS #num;
CREATE TABLE #num (ID INT NOT NULL);

INSERT INTO #num WITH (TABLOCK)
SELECT TOP (@batch_size) -1 + ROW_NUMBER()
	OVER (ORDER BY (SELECT NULL)) RN
FROM master..spt_values t1
CROSS JOIN master..spt_values t2;

WHILE @current_loop <= @target_loops
BEGIN
	SET @middle_num = CEILING(@rows_in_target * RAND());

	DELETE tgt WITH (TABLOCK)
 	FROM dbo.LUCKY_CCI tgt
 	WHERE tgt.ID BETWEEN
 	CAST(@middle_num - FLOOR(0.5*@batch_size) AS BIGINT)
	AND
	CAST(@middle_num + FLOOR(0.5*@batch_size) AS BIGINT);

 	INSERT INTO dbo.LUCKY_CCI WITH (TABLOCK)
 	SELECT @middle_num + n.ID - FLOOR(0.5 * @batch_size)
	FROM #num n
 	WHERE n.ID BETWEEN 1 AND @rows_in_target
	OPTION (MAXDOP 1);

	SET @current_loop = @current_loop + 1;
END;

END;

Our table now has a bunch of uncompressed rowgroups and many compressed rowgroups with lots of deleted rows. Let’s clean it up with a REORG:

ALTER INDEX CCI ON dbo.LUCKY_CCI
REORGANIZE WITH (COMPRESS_ALL_ROW_GROUPS = ON);

This table has a REFF of about 41 for a batch size of 1048576. That means that we can expect to see significantly worse rowgroup elimination than before. Running the same SELECT query as before:

Table ‘LUCKY_CCI’. Segment reads 51, segment skipped 75.

The LUCKY_CCI table clearly needs to be renamed.

The Magic of Partitioning

---

A partition for a CCI is a way of organizing rowgroups. Let’s create the same table but with partitions that can hold five million integers. Below is code to do that for those following along at home:

CREATE PARTITION FUNCTION functioning_part_function
(BIGINT)
AS RANGE RIGHT
FOR VALUES (
  0
, 5000000
, 10000000
, 15000000
, 20000000
, 25000000
, 30000000
, 35000000
, 40000000
, 45000000
, 50000000
, 55000000
, 60000000
, 65000000
, 70000000
, 75000000
, 80000000
, 85000000
, 90000000
, 95000000
, 100000000
, 105000000
); 

CREATE PARTITION SCHEME scheming_part_scheme
AS PARTITION functioning_part_function
ALL TO ( [PRIMARY] );

DROP TABLE IF EXISTS dbo.PART_CCI;

CREATE TABLE dbo.PART_CCI (
ID BIGINT NOT NULL,
INDEX CCI CLUSTERED COLUMNSTORE
) ON scheming_part_scheme(ID);

INSERT INTO dbo.PART_CCI WITH (TABLOCK)
SELECT TOP (100 * 1048576) ROW_NUMBER()
	OVER (ORDER BY (SELECT NULL)) RN
FROM master..spt_values t1
CROSS JOIN master..spt_values t2
CROSS JOIN master..spt_values t3
ORDER BY RN
OPTION (MAXDOP 1);

5 million was a somewhat arbitrary choice. The important thing is that it’s above 1048576 rows. Generally speaking, picking a partition size below 1048576 rows is a bad idea because the table’s rowgroups will never be able to reach the maximum size.

Using the sys.dm_db_column_store_row_group_physical_stats dmv we can see that rowgroups were divided among partitions as expected:

This table has a REFF of 1.9 for a batch size of 1048576. This is perfectly reasonable because the partition size was defined as 5 * 1000000 instead of 5 * 1048576 which would be needed for a REFF of 1.0.

Now I’ll run the same code as before to do 60 million DUIs against the table. I expect the partitions to limit the damage because row cannot move out of their partitions. After the process finishes and we do a REORG, the table has a REFF of about 3.7. The rowgroups are definitely a bit messier:

Running our SELECT query one last time:

SELECT COUNT(*)
FROM dbo.Part_CCI
WHERE ID BETWEEN 12345678 AND 13345677;

We see the following numbers for rowgroup elimination:

Table ‘PART_CCI’. Segment reads 4, segment skipped 3.

Only four segments are read. Many segments are skipped, including all of those in partitions not relevant to the query.

Final Thoughts

---

Large, unpartitioned CCIs will eventually run into issues with rowgroup elimination fragmentation if the loading process can delete data. A maintainance operation that rebuilds the table with a clustered rowstore index and rebuilds the CCI with MAXDOP = 1 is one way to restore nicely organized rowgroups, but that is a never-ending battle which grows more expensive as the table gets larger. Picking the right partitioning function can guarantee rowgroup elimination fragmentation.

Thanks for reading!

Going Further

---

If this is the kind of SQL Server stuff you love learning about, you’ll love my training. I’m offering a 75% discount to my blog readers if you click from here. I’m also available for consulting if you just don’t have time for that and need to solve performance problems quickly.