The Document Attachments were stored in our AX 2012 database, and were occupying several hundred GBs. As part of the migration to D365FO we had to export these files in bulk. The most efficient way is to run the batch task in parallel. But this data is tilted in a way that if you would group the tasks based on their creation date, it the first tasks would barely have any records to process while the last ones would go on forever. The solution for this is called data histogram equalization.

It would be a big challenge to code this in X++, but SQL Server has a function for doing this exactly: NTILE.

The following direct query is able to process the data histogram, and then return 10 buckets of Record Identifier ranges of roughly equal size:

WITH documents (Bucket, RecId)
AS (
       SELECT NTILE(10) OVER( ORDER BY DocuRef.RecId) AS Bucket
    FROM docuRef
        INNER JOIN docuType
            ON  (docuType.dataAreaId = docuRef.RefCompanyId OR DOCUREF.REFCOMPANYID = '')
                AND docuType.TypeId = docuRef.TypeId
        INNER JOIN docuValue
            ON  docuValue.RecId          = docuRef.ValueRecId                
       WHERE  docuType.TypeGroup         = 1 -- DocuTypeGroup::File
        AND docuType.FilePlace           = 0 -- DocuFilePlace::Archive
        AND docuType.ArchivePath  <> ''
              AND docuValue.Path                = ''
       ,count(*) AS Count
       ,(SELECT MIN(RecId) FROM documents D WHERE D.Bucket = documents.Bucket) AS RecId_From
       ,(SELECT MAX(RecId) FROM documents D WHERE D.Bucket = documents.Bucket) AS RecId_To
       FROM documents
       GROUP BY Bucket

Here are the results for spawning 10 batch job tasks to do parallel execution based on the RecId surrogate key index, with ~57230 rows in each bucket. This allows you to evenly distribute the load for data processing in parallel.

data histogram equalization

If we would export our document attachments sequentially, it would take roughly 40 hours total. By utilizing the data histogram equalization we could get it down to be under 3 hours. That is amazing!

It is a great way to split ranges of data for parallel processing, to keep in mind for the future.

You can learn more about the functionality here: