3

I have a query that's taking quite a long time on my machine (7 minutes) to execute and I was wondering whether I could make it (significantly) quicker:

SELECT
    rec.[Id] AS RecordId,
    MIN(rec.[CreationDate]) AS RecordCreationDate,
    MIN(rec.[LastModified]) AS RecordLastModified,
    MIN(rec.[AssetType]) AS RecordAssetType,
    MIN(rec.[MasterFilename]) AS RecordMasterFilename,
    MIN(rec.[GameName]) AS RecordGameName,
    usr.[OrganizationName],
    COUNT(hist.[Id]) AS TimesDownloaded
FROM
(
    SELECT
        innerRec.Id,
        MIN(innerRec.CreationDate) AS CreationDate,
        MIN(innerRec.LastModified) AS LastModified,
        MIN(innerRec.AssetType) AS AssetType,
        MIN(innerRec.MasterFilename) AS MasterFilename,
        MIN(innerRec.GameName) AS GameName
    FROM
        [dbo].[Record] innerRec INNER JOIN [dbo].[RecordClassificationLink] innerLnk ON innerRec.Id = innerLnk.RecordId
    -- WHERE (classification ID is foo or bar)
    GROUP BY
        innerRec.Id
    -- HAVING COUNT(innerLnk.ClassificationId) = (number of specified classifications)
) rec
CROSS JOIN
    [dbo].[AdamUser] usr
LEFT JOIN
    (SELECT * FROM [dbo].[MaintenanceJobHistory] WHERE [CreatedOn] > '2016-01-01 00:00:00' AND [CreatedOn] < '2016-12-01 00:00:00') hist ON usr.Name = hist.AccessingUser AND rec.Id = hist.RecordId
GROUP BY
    rec.Id, usr.OrganizationName

What it's doing is extracting data to be put in an Excel spreadsheet report (whether a spreadsheet is a good presentation of this data is outside the scope of this question :-) )

The first subquery pulls out records optionally filtered by a list of classification IDs. These are then cross joined with the user table, because each user table row actually contains the info we really need for this: the user's organization name. I then left join the maintenance job history table (storing an entry for each record download) in order to create multiple rows if a record has been accessed multiple times, then group by record ID and organization name to get a "number of record downloads per organization" count as TimesDownloaded.

The code that is reading this output then populates an associative array whose key is OrganizationName and whose value is TimesDownloaded, creating the equivalent of a dynamic PIVOT where each record row contains one column per organization each containing the count of the number of record downloads.

As you can imagine this runs quite slowly on a large dataset, as I said above; the one I'm working with has ~38000 Records and ~1000 users, meaning the cross join results in ~38,000,000 rows, but it seems conceptually necessary.

Can this be made significantly more efficient? Would it be better if I did the PIVOT in dynamic SQL instead?

The DBMS I'm using is SQL Server 2014.

Here are the schema definitions for the tables:

CREATE TABLE [dbo].[AdamUser](
    [Id] [uniqueidentifier] NOT NULL,
    [Name] [nvarchar](200) NOT NULL,
    [UserGroupName] [nvarchar](50) NOT NULL,
    [OrganizationName] [nvarchar](50) NOT NULL,
PRIMARY KEY CLUSTERED 
(
    [Id] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
) ON [PRIMARY]

GO

CREATE UNIQUE NONCLUSTERED INDEX [UIX_AdamUser_Name] ON [dbo].[AdamUser]
(
    [Name] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, IGNORE_DUP_KEY = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
GO

CREATE TABLE [dbo].[MaintenanceJobHistory](
    [Id] [uniqueidentifier] NOT NULL,
    [Data] [xml] NOT NULL,
    [CreatedOn] [datetime] NOT NULL,
    [Type] [nvarchar](512) NOT NULL,
    [RecordId] [uniqueidentifier] NOT NULL,
    [AccessingUser] [nvarchar](200) NOT NULL,
 CONSTRAINT [PK_MaintenanceJobHistory] PRIMARY KEY CLUSTERED 
(
    [Id] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
) ON [PRIMARY] TEXTIMAGE_ON [PRIMARY]

GO

CREATE NONCLUSTERED INDEX [IX_MaintenanceJobHistory_CreatedOn] ON [dbo].[MaintenanceJobHistory]
(
    [CreatedOn] ASC
)
INCLUDE (     [Id],
    [RecordId],
    [AccessingUser]) WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, SORT_IN_TEMPDB = OFF, DROP_EXISTING = OFF, ONLINE = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON)
GO

CREATE TABLE [dbo].[Record](
    [Id] [uniqueidentifier] NOT NULL,
    [CreationDate] [datetime] NOT NULL,
    [LastModified] [datetime] NOT NULL,
    [AssetType] [nvarchar](max) NULL,
    [MasterFilename] [nvarchar](max) NULL,
    [GameName] [nvarchar](max) NULL,
 CONSTRAINT [PK_Record] PRIMARY KEY CLUSTERED 
(
    [Id] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
) ON [PRIMARY] TEXTIMAGE_ON [PRIMARY]

GO

CREATE TABLE [dbo].[RecordClassificationLink](
    [Id] [uniqueidentifier] NOT NULL,
    [RecordId] [uniqueidentifier] NOT NULL,
    [ClassificationId] [uniqueidentifier] NOT NULL,
 CONSTRAINT [PK_RecordClassificationLink] PRIMARY KEY CLUSTERED 
(
    [Id] ASC
)WITH (PAD_INDEX = OFF, STATISTICS_NORECOMPUTE = OFF, IGNORE_DUP_KEY = OFF, ALLOW_ROW_LOCKS = ON, ALLOW_PAGE_LOCKS = ON) ON [PRIMARY]
) ON [PRIMARY]

GO

Here's the execution plan: https://www.brentozar.com/pastetheplan/?id=Sy6LlXDXg

Typical output: Typical output

However this output gets turned into an Excel spreadsheet along the lines of the following by the calling program (hence its being like a PIVOT operation):

.----------------------------------------------------------------------.
| Filename | Creation Date | #times downloaded by: CompanyA | CompanyB | ...
| foo.png  | 1/2/3         |                       0        | 2        |
| bar.png  | 1/3/4         |                       3        | 1        |
...

UPDATE:

I ended up making things significantly more efficient by moving the PIVOT operation into the SQL query itself; that way, SQL Server only has to output the number of rows in the Record table rather than that multiplied by the number of organizations (not too bad until you get to hundreds of organizations, at which point it's a huge number). The operation still takes a few minutes, but it's much more bearable. Here's the query I finally decided to use:

SELECT *
FROM (
    SELECT
        rec.[Id]                         AS RecordId,
        'Org_' + usr.[OrganizationName]  AS OrganizationNamePrefixed,
        COUNT(hist.[Id])                 AS TimesDownloaded  -- To be aggregated by PIVOT
    FROM (
        SELECT
            innerRec.[Id]
        FROM
            [dbo].[Record] innerRec
        INNER JOIN
            [dbo].[RecordClassificationLink] innerLnk ON innerLnk.[RecordId] = innerRec.[Id]
        -- WHERE (classification ID is foo or bar), for optional classification filtering
        GROUP BY
            innerRec.[Id]
        -- HAVING COUNT(innerLnk.ClassificationId) = (number of specified classifications), for optional classification filtering
    ) rec
    CROSS JOIN [dbo].[AdamUser] usr
    LEFT JOIN (
        SELECT * FROM [dbo].[MaintenanceJobHistory] WHERE [CreatedOn] > 'eg. 2016-01-01 12:00:00' AND [CreatedOn] < 'eg. 2016-12-01 12:00:00'
    ) hist ON hist.[AccessingUser] = usr.[Name] AND hist.[RecordId] = rec.[Id]
    GROUP BY
        rec.[Id], usr.[OrganizationName]
) srcTable
PIVOT  -- Pivot around columns outside aggregation fn, eg. heading column [OrganizationNamePrefixed] & all other columns: [RecordId]
(
    MIN(srcTable.[TimesDownloaded]) FOR [OrganizationNamePrefixed] IN (...list of ~200 columns dynamically generated...)
) pivotTable
INNER JOIN [dbo].[Record] outerRec ON outerRec.[Id] = pivotTable.[RecordId]

I added various indexes and also made the PIVOT as efficient as possible by only selecting the aggregation column, the headings column, and the necessary other column(s) to pivot around. Lastly, I re-JOIN the Record table using the RecordId PK to get the extra record info per row.

  • Can you give an example of the results you expect? Is it a "cross tab" of Records and Users? (with "Number of downloads" or null as the sum that appears in each cell?) – seventyeightist Dec 8 '16 at 19:54
  • Try with option (querytraceon 4199, querytraceon 9481) --> Enable all optimizer fixes and use legacy CE. I see it using parallelism and hash join with good seeks & no spool operator. I dont have the data to play, but using fake stats with ROWCOUNT=10000000 I am getting better plan. Obviously, adding appropriate indexes will give it more boost. – Kin Shah Dec 8 '16 at 22:18
  • Sorry .. I see table spool .. but I would still test out with the trace flags as I mentioned above. – Kin Shah Dec 8 '16 at 22:49
3

The stuff that is probably taking a lot of time is a large number of Sort operations in your query plan. You can preempt those by sorting the data yourself, in the form of indexes.

Here are some index suggestions that I think would get you started:

CREATE INDEX IX_RecordId
    ON dbo.RecordClassificationLink (RecordId);

CREATE INDEX IX_MaintenanceJobHistory_ByUser
    ON dbo.MaintenanceJobHistory (AccessingUser, RecordId);

CREATE UNIQUE INDEX UIX_AdamUser_OrgnName
    ON dbo.AdamUser (OrganizationName, Name);

Then, you could modify your query a little bit to help the optimizer make some smart choices, like aggregating certain data streams before they're joined and create a much larger product that will take more time to aggregate:

SELECT
    rec.[Id] AS RecordId,
    MIN(rec.[CreationDate]) AS RecordCreationDate,
    MIN(rec.[LastModified]) AS RecordLastModified,
    MIN(rec.[AssetType]) AS RecordAssetType,
    MIN(rec.[MasterFilename]) AS RecordMasterFilename,
    MIN(rec.[GameName]) AS RecordGameName,
    usr.[OrganizationName],
    SUM(ISNULL(hist._count, 0)) AS TimesDownloaded
FROM
(
    SELECT
        innerRec.Id,
        MIN(innerRec.CreationDate) AS CreationDate,
        MIN(innerRec.LastModified) AS LastModified,
        MIN(innerRec.AssetType) AS AssetType,
        MIN(innerRec.MasterFilename) AS MasterFilename,
        MIN(innerRec.GameName) AS GameName
    FROM [dbo].[Record] innerRec
    INNER JOIN [dbo].[RecordClassificationLink] innerLnk ON innerRec.Id = innerLnk.RecordId
    -- WHERE (classification ID is foo or bar)
    GROUP BY
        innerRec.Id
    -- HAVING COUNT(innerLnk.ClassificationId) = (number of specified classifications)
) AS rec
CROSS JOIN [dbo].[AdamUser] AS usr
LEFT JOIN
    (SELECT AccessingUser, RecordId, COUNT(*) AS _count
     FROM [dbo].[MaintenanceJobHistory]
     WHERE [CreatedOn] > '2016-01-01 00:00:00' AND
           [CreatedOn] < '2016-12-01 00:00:00'
     GROUP BY AccessingUser, RecordId
    ) AS hist ON
        usr.Name = hist.AccessingUser AND
        rec.Id = hist.RecordId
GROUP BY
    rec.Id, usr.OrganizationName;

I've done the following:

  • the hist subquery is aggregated on AccessingUser, RecordId, and I've created a COUNT(*) AS _count. This query uses the new index IX_MaintenanceJobHistory_ByUser to perform very efficiently without any memory grant or hash tables.
  • Because of this, I've replaced COUNT(hist.Id) with SUM(ISNULL(hist._count, 0)) AS TimesDownloaded
  • The new index on dbo.RecordClassificationLink helps perform a smooth join with the Record table, but if you add your WHERE and HAVING, that index won't help you.
  • The new index on dbo.AdamUser also improves performance by eliminating a Sort operator - because you aggregate on the OrganizationName column, might as well have your data sorted on this from the get-go.

In my mind, that should give you the same result, but it's late here, so you'll have to verify the results yourself. :)

Here's my query plan: First plan

EDIT: You could also simplify the rec part a bit - probably be a bit easier to read:

SELECT
    rec.[Id] AS RecordId,
    rec.[CreationDate] AS RecordCreationDate,
    rec.[LastModified] AS RecordLastModified,
    rec.[AssetType] AS RecordAssetType,
    rec.[MasterFilename] AS RecordMasterFilename,
    rec.[GameName] AS RecordGameName,
    usr.[OrganizationName],
    SUM(ISNULL(hist._count, 0)) AS TimesDownloaded
FROM
(
    SELECT
        Id,
        CreationDate,
        LastModified,
        AssetType,
        MasterFilename,
        GameName
    FROM [dbo].[Record]
    WHERE Id IN (
        SELECT RecordId
        FROM [dbo].[RecordClassificationLink]
        --- WHERE ClassificationId=...
        )
) AS rec
CROSS JOIN [dbo].[AdamUser] AS usr
LEFT JOIN
    (SELECT AccessingUser, RecordId, COUNT(*) AS _count
     FROM [dbo].[MaintenanceJobHistory]
     WHERE [CreatedOn] > '2016-01-01 00:00:00' AND
           [CreatedOn] < '2016-12-01 00:00:00'
     GROUP BY AccessingUser, RecordId
    ) AS hist ON
        usr.Name = hist.AccessingUser AND
        rec.Id = hist.RecordId
GROUP BY
    usr.OrganizationName,
    rec.[Id],
    rec.[CreationDate],
    rec.[LastModified],
    rec.[AssetType],
    rec.[MasterFilename],
    rec.[GameName],
    usr.[OrganizationName];

... and the plan looks a tiny bit better as well (look for the red Lazy spool at the bottom of the original plan, that is now gone.)

New plan

  • IX_RecordId is invalid because that column isn't guaranteed to be unique; records can have multiple classifications. – Jez Dec 8 '16 at 21:52
  • Ah, of course. I've changed it to non-unique. Doesn't really affect the plan. – Daniel Hutmacher Dec 8 '16 at 21:56
  • Daniel, FYI : dba.stackexchange.com/q/162801/1186 – Aaron Bertrand Jan 31 '17 at 22:11

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.