I've written an application with a SQL Server backend that collects and stores and extremely large amount of records. I've calculated that, at the peak, the average amount of records is somewhere in the avenue of 3-4 billion per day (20 hours of operation).

My original solution (before I'd done the actual calculation of the data) was to have my application inserting records into the same table that is queried by my clients. That crashed and burned fairly quickly, obviously, because it's impossible to query a table that's having that many records inserted.

My second solution was to use 2 databases, one for data received by the application and one for client-ready data.

My application would receive data, chunk it into batches of ~100k records and bulk-insert into the staging table. After ~100k records the application would, on the fly, create another staging table with the same schema as before, and begin inserting into that table. It would create a record in a jobs table with the name of the table that has 100k records and a stored procedure on the SQL Server side would move the data from the staging table(s) to client-ready production table, and then drop the table temporary table created by my application.

Both databases have the same set of 5 tables with the same schema, except the staging database which has the jobs table. The staging database has no integrity constraints, key, indexes etc... on the table where the bulk of records will reside. Shown below, the table name is SignalValues_staging. The goal was to have my application slam the data into SQL Server as quickly as possible. The workflow of creating tables on the fly so they can easily be migrated works pretty well.

The following is the 5 relevant tables from my staging database, plus my jobs table:

Staging tables The stored procedure I have written handles the moving of the data from all of the staging tables and inserting it into production. Below is the part of my stored procedure that inserts into production from the staging tables:

-- Signalvalues jobs table.
      ,ROW_NUMBER() OVER (ORDER BY JobId) AS 'RowIndex'
INTO #JobsToProcess
    SELECT JobId 
           ,SignalValueStagingTableName AS 'TableName'
           ,(DATEDIFF(SECOND, (SELECT last_user_update
                              FROM sys.dm_db_index_usage_stats
                              WHERE database_id = DB_ID(DB_NAME())
                                AND OBJECT_ID = OBJECT_ID(SignalValueStagingTableName))
                     ,GETUTCDATE())) SecondsSinceLastUpdate
    FROM SignalValueJobs
) cte
WHERE cte.ProcessingComplete = 1
   OR cte.SecondsSinceLastUpdate >= 120


DECLARE @currentTable NVARCHAR(128) 
DECLARE @processingParam BIT
DECLARE @sqlStatement NVARCHAR(2048)
DECLARE @paramDefinitions NVARCHAR(500) = N'@currentJob UNIQUEIDENTIFIER, @processingComplete BIT'
DECLARE @qualifiedTableName NVARCHAR(128)

WHILE @i > 0

    SELECT @jobParam = JobId, @currentTable = TableName, @processingParam = ProcessingComplete
    FROM #JobsToProcess 
    WHERE RowIndex = @i 

    SET @qualifiedTableName = '[Database_Staging].[dbo].['+@currentTable+']'

    SET @sqlStatement = N'

        --Signal values staging table.
        SELECT svs.* INTO #sValues
        FROM '+ @qualifiedTableName +' svs
        INNER JOIN SignalMetaData smd
            ON smd.SignalId = svs.SignalId  

        INSERT INTO SignalValues SELECT * FROM #sValues

        SELECT DISTINCT SignalId INTO #uniqueIdentifiers FROM #sValues

        DELETE c FROM '+ @qualifiedTableName +' c INNER JOIN #uniqueIdentifiers u ON c.SignalId = u.SignalId

        DROP TABLE #sValues
        DROP TABLE #uniqueIdentifiers

        IF NOT EXISTS (SELECT TOP 1 1 FROM '+ @qualifiedTableName +') --table is empty
            -- processing is completed so drop the table and remvoe the entry
            IF @processingComplete = 1 
                DELETE FROM SignalValueJobs WHERE JobId = @currentJob

                IF '''+@currentTable+''' <> ''SignalValues_staging''
                    DROP TABLE '+ @qualifiedTableName +'

    EXEC sp_executesql @sqlStatement, @paramDefinitions, @currentJob = @jobParam, @processingComplete = @processingParam;

    SET @i = @i - 1

DROP TABLE #JobsToProcess

I use sp_executesql because the table names for the staging tables come as text from the records in the jobs table.

This stored procedure runs every 2 seconds using the trick I learned from this dba.stackexchange.com post.

The problem I cannot for the life of me resolve is the speed at which the inserts into production are performed. My application creates temporary staging tables and fills them with records incredibly quickly. The insert into production cannot keep up with the amount of tables and eventually there's a surplus of tables into the thousands. The only way I've ever been able to keep up with the incoming data is to remove all keys, indexes, constraints etc... on the production SignalValues table. The problem I then face is that the table ends up with so many records it becomes impossible to query.

I've tried partitioning the table using the [Timestamp] as a partitioning column to no avail. Any form of indexing at all slows the inserts so much that they can't keep up. In addition, I'd need to create thousands of partitions (one every minute? hour?) years in advance. I couldn't figure out how to create them on the fly

I tried creating partitioning by adding a computed column to the table called TimestampMinute whose value was, on INSERT, DATEPART(MINUTE, GETUTCDATE()). Still too slow.

I've tried making it a Memory-Optimized Table as per this Microsoft article. Maybe I don't understand how to do it, but the MOT made the inserts slower somehow.

I've checked the Execution Plan of the stored procedure and found that (I think?) the most intensive operation is

SELECT svs.* INTO #sValues
FROM '+ @qualifiedTableName +' svs
INNER JOIN SignalMetaData smd
    ON smd.SignalId = svs.SignalId

To me this doesn't make sense: I've added wall-clock logging to the stored procedure that proved otherwise.

In terms of time-logging, that particular statement above executes in ~300ms on 100k records.

The statement

INSERT INTO SignalValues SELECT * FROM #sValues

executes in 2500-3000ms on 100k records. Deleting from the table the records affected, per:

DELETE c FROM '+ @qualifiedTableName +' c INNER JOIN #uniqueIdentifiers u ON c.SignalId = u.SignalId

takes another 300ms.

How can I make this faster? Can SQL Server handle into the billions of records per day?

If it's relevant, this is SQL Server 2014 Enterprise x64.

Hardware Configuration:

I forgot to include hardware in the first pass of this question. My bad.

I'll preface this with these statements: I know I am losing some performance because of my hardware configuration. I've tried many times but because of budget, C-Level, the alignment of the planets, etc... there's nothing I can do to get a better setup unfortunately. The server is running on a virtual machine and I can't even increase the memory because we simply don't have any more.

Here's my system information:

System Info

The storage is attached to the VM server via iSCSI interface to a NAS box (This will degrade performance). The NAS box has 4 drives in a RAID 10 configuration. They're 4TB WD WD4000FYYZ spinning disk drives with 6GB/s SATA interface. The server only has one data-store configured so tempdb and my database are on the same datastore.

Max DOP is zero. Should I change this to a constant value or just let SQL Server handle it? I read up on RCSI: Am I correct in assuming that the only benefit from RCSI comes with row updates? There will never be updates to any of these particular records, they'll be INSERTed and SELECTed. Will RCSI still benefit me?

My tempdb is 8mb. Based on the answer below from jyao, I changed the #sValues to a regular table to avoid tempdb altogether. Performance was about the same though. I will try increasing the size and growth of tempdb, but given that the size of #sValues will more or less always be the same size I don't anticipate much gain.

I have taken an execution plan that I've attached below. This execution plan is one iteration of a staging table -- 100k records. The execution of the query was fairly quick, around 2 seconds, but keep in mind that this is without indexes on the SignalValues table and the SignalValues table, the target of the INSERT, has no records in it.

Execution plan

  • 3
    Did you already experiment with delayed durability? Dec 10, 2015 at 21:30
  • 2
    What indexes were in place with slow production inserts?
    – paparazzo
    Dec 10, 2015 at 23:12
  • So far I don't think there is enough data here to find out what actually is consuming so much time. Is it CPU? Is it IO? Since you seem to be getting 30k rows per second it does not look like IO to me. Do I understand this right that you are quite close to achieving your perf goal? You need 50k rows per second, so one 100k batch every 2 seconds should suffice. Right now one batch seems to take 3 seconds. Post the actual execution plan of one representative run. Any suggestion that does not attack the most time consuming operations is moot.
    – usr
    Dec 15, 2015 at 0:01
  • I've posted the execution plan.
    – Brandon
    Dec 15, 2015 at 17:24

3 Answers 3


I've calculated that, at the peak, the average amount of records is somewhere in the avenue of 3-4 billion per day (20 hours of operation).

From your screenshot, you ONLY have 8GB memory total RAM and 6 GB allocated to SQL Server. This is way tooo low for what you are trying to achieve.

I suggest you to upgrade the memory to a higher value - 256GB and bump up your VM CPUs as well.

You need to invest in hardware at this point for your workload.

Also refer to data loading performance guide - it describes smart ways of efficiently loading the data.

My tempdb is 8mb.

Based on your edit .. you should have a sensible tempdb - preferably multiple tempdb data files equally sized along with TF 1117 and 1118 enabled instance wide.

I would suggest you to get a professional health check and start from there.

Highly recommend

  1. Bump up your server spec.

  2. Get a professional* person do a health check of your database server instance and follow the recommendations.

  3. Once a. and b. are done, then immerse yourself in query tuning and other optimizations like looking at wait stats, query plans, etc.

Note: I am a professional sql server expert at hackhands.com - a pluralsight company, but in no means suggesting you to hire me for help. I am merely suggesting you to take professional help based on your edits only.


  • I'm trying to put together a proposal (read: begging for) more hardware for this. With that in mind and your answer here, there is nothing else from a SQL Server configuration or query optimization standpoint that you would suggest to make this faster?
    – Brandon
    Dec 15, 2015 at 15:35

General advice for such problems with big-data, when facing a wall and nothing works:

One egg is going to be cooked 5 minutes about. 10 eggs will be cooked in same time if enough electricity and water.

Or, in other words:

First, look at the hardware; second, separate the process logic (data remodeling) and do it in parallel.

It is quite possible to create custom vertical partitioning dynamically and automated, per table count and per table size; If I have Quarter_1_2017, Quarter_2_2017, Quarter_3_2017, Quarter_4_2017, Quarter_1_2018... and I don't know where my records are and how much partitions I have, run same query(s) against all custom partitions in same time, separate sessions and assembly the result to be processed forward for my logic.

  • The OP's issue appears to be handling the insertion and access to newly entered data, more than processing data from weeks or months ago. OP mentions partitioning data by the minute on his time stamp (so 60 partitions, splitting current data into separate buckets); splitting by quarter wouldn't be likely to be much help. Your point is well taken in general, but is unlikely to help someone in this specific situation.
    – RDFozz
    Feb 8, 2018 at 16:00

I will do the following check / optimization:

  1. Ensure both the data and logs file of the production database does not grow during the insert operation (pre-grow it if needed)

  2. Do not use

    select * into [dest table] from [source table];

    but instead, pre-define the [dest table]. Also instead of dropping the [dest table] and recreate it, I will truncate table. This way, if needed, instead of using temp table, I'd use regular table. (I may also create the index on [dest table] to facilitate the performance of the join query)

  3. Instead of using dynamic sql, I'd rather use hard-coded table names with some coding logic to choose which table to operate.

  4. I will also monitor the memory, CPU and disk I/O performance to see whether there are resource starvations during big workload.

  5. Since you mentioned you can handle the insertion by dropping the indexes on the production side, I'd check whether there are many page splits occurring, if so, I'd decrease the fillfactor of the indexes and rebuild the indexes before I consider to drop the indexes.

Good luck and love your question.

  • Thanks for the answer. I had set the database size to 1gb and grow by 1gb anticipating that growth operations would take some time, that did help with speed initially. I will try to implement the pre-growth today. I did implement the [dest] table as a regular table but didn't see much performance gain. I haven't had much time the past few days but I will try to get to the other ones today.
    – Brandon
    Dec 14, 2015 at 16:34

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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