Batching work in SQL Server

In one of our ASP.NET Core applications, I added a new feature to cleanup old data. My implementation was simple and used a BackgroundService to run a cleanup script on periodic intervals:

	using Microsoft.ApplicationInsights;
	using Microsoft.AspNetCore.Mvc.Infrastructure;
	using Microsoft.Extensions.Hosting;
	using Microsoft.Extensions.Options;
	using System;
	using System.Security.Permissions;
	using System.Threading;
	using System.Threading.Tasks;
	using System.Timers;
	using VLM.DocumentStorage.API;
	using VLM.DocumentStorage.Infrastructure;
	using VLM.SOFACore.Data;
	namespace VLM.DocumentStorage.Domain
	{
	public class CleanupService : BackgroundService
	{
	private readonly IUnitOfWorkFactory _unitOfWorkFactory;
	private readonly IDocumentRepository _documentRepository;
	private readonly TelemetryClient _telemetryClient;
	private readonly IOptionsMonitor<CleanupOptions> _options;
	public CleanupService(IUnitOfWorkFactory unitOfWorkFactory, IDocumentRepository documentRepository, TelemetryClient telemetryClient, IOptionsMonitor<CleanupOptions> options)
	{
	_unitOfWorkFactory = unitOfWorkFactory;
	_documentRepository = documentRepository;
	_telemetryClient = telemetryClient;
	_options = options;
	}
	protected override async Task ExecuteAsync(CancellationToken stoppingToken)
	{
	using PeriodicTimer timer = new PeriodicTimer(TimeSpan.FromMinutes(_options.CurrentValue.IntervalInMinutes));
	while (!stoppingToken.IsCancellationRequested && await timer.WaitForNextTickAsync(stoppingToken))
	{
	try
	{
	if(_options.CurrentValue.Enabled)
	await Cleanup();
	}
	catch (Exception ex)
	{
	_telemetryClient.TrackException(ex);
	}
	}
	}
	private async Task Cleanup()
	{
	using var uow = _unitOfWorkFactory.Create();
	_telemetryClient.TrackTrace("Cleanup - Calling cleanup.");
	var nrOfAffectedRecords=await _documentRepository.Cleanup();
	_telemetryClient.TrackTrace($"Cleanup - {nrOfAffectedRecords} cleaned up.");
	}
	}
	}

view raw CleanupService.cs hosted with ❤ by GitHub

	using NHibernate.Linq;
	using System;
	using System.Collections.Generic;
	using System.Linq;
	using System.Threading.Tasks;
	using VLM.DocumentStorage.Domain;
	using VLM.SOFACore.Configuration;
	using VLM.SOFACore.NHibernate;
	namespace VLM.DocumentStorage.Infrastructure
	{
	public class DocumentRepository : NHibernateRepository, IDocumentRepository
	{
	public DocumentRepository(ISessionProvider sessionProvider) : base(sessionProvider)
	{
	}
	public async Task<int> Cleanup()
	{
	var updateQuery= """
	DELETE
	FROM DocumentVersion dv
	WHERE dv.UpdatedOn < DATEADD(month, -6, GETDATE())
	""";
	//Delete data older than 6 months
	return await this.Session.CreateSQLQuery(updateQuery)
	.ExecuteUpdateAsync();
	}
	}
	}

view raw DocumentRepository.cs hosted with ❤ by GitHub

All worked fine during development and testing, but when I deployed it to production it brought the whole application to a halt.

What was happening?

First, as this was the first time the script was run on production, there was a lot of old data. So while the query executed and completed quite fast on other environments, on production it impacted millions of rows.

What made the problem even worse is that the table that should be cleaned up contained a large amount of binary data. This made the transaction log grow in size and further increased the query duration.

My first attempt to improve the performance of this query was to delete the data based on the primary key. A suggestion I found here: How to Delete Large Amounts of Data – SQLServerCentral

However the impact of this change was minimal and the query still timed out.

I had not enough time to find a solution that allowed me to fix the issue at the application level. So I ‘hacked’ together a solution where I executed the query in small batches:

	DECLARE @i INT = 1;
	DECLARE @COUNT INT = 1;
	DECLARE @BeforeDate DateTime = DATEADD(month, -6, GETDATE());
	SET @COUNT = (SELECT COUNT(*) FROM dbo.DocumentVersion);
	WHILE (@i <= @Count)
	BEGIN
	WAITFOR DELAY '00:00:01'
	DELETE TOP 100
	FROM DocumentVersion dv
	WHERE dv.UpdatedOn < @BeforeDate;
	SET @i = @i + 1;
	END

view raw Delete.sql hosted with ❤ by GitHub

Using the query above the delete is executed in small batches of 100 rows with an interval of 1 second between every batch.

The total query execution time is still long but at least my application & database remained accessible during the execution.

Anyone with a suggestion for a better solution?