C#’s foreach ruined my afternoon

Published on: 2017-03-21

How I stumbled into a breaking change between C# 4.0 & 5.0

Forest Fire” by CIFOR is licensed under CC BY-NC-ND 2.0

The other afternoon I ran into some nightmarish debugging with the following code:

private static void StartThreads()
{
    var values = new List<int>() { 1, 2, 3 };
    var threads = new List<Thread>();
    foreach (var value in values)
    {
        Thread t = new Thread(() => Run(value));
        threads.Add(t);
        t.Start();
    }
    // Wait for all threads to complete
    foreach (Thread t in threads)
        t.Join();
}
private static void Run(int value)
{
    Console.Write(value.ToString());
}

(I know, I know, I wish I could be using TPL but in this case I couldn’t)

On my local machine, the code above ran and gave me my expected console output of 123 (your results may vary depending on what order the threads run in).

When I ran this code on my server however, the output was 333.

<begin pulling out hair>

Long story short, after a couple hours of investigation I figured out that the way a foreach loop works under the hood in C# ≥ 5.0, which is what I run on my local machine, works differently than a foreach loop in C# < 5.0, which is what I had on my server.

From the C# 4.0 spec, a foreach loop is really a while loop behind the scenes, meaning the code above really translates into something like this:

private static void StartThreads()
{
    var values = new List<int>() { 1, 2, 3 };
    var threads = new List<Thread>();
    IEnumerator<int> e = ((IEnumerable<int>)values).GetEnumerator();
    try
    {
        int v; // DECLARED OUTSIDE OF THE LOOP!!!
        while (e.MoveNext())
        {
            v = (int)(int)e.Current;
            Thread t = new Thread(() => Run(v));
            threads.Add(t);
            t.Start();
        }
    }
    finally
    {
        if (e != null) ((IDisposable)e).Dispose();
    }
    // Wait for all threads to complete
    foreach (Thread t in threads)
        t.Join();
}

The important thing to note in the above code is that int v gets declared outside of the while loop.

In the C# 5.0 spec, that int v gets declared inside the loop (causing it to get recreated with every iteration):

private static void StartThreads()
{
    var values = new List<int>() { 1, 2, 3 };
    var threads = new List<Thread>();
    IEnumerator<int> e = ((IEnumerable<int>)values).GetEnumerator();
    try
    {
        while (e.MoveNext())
        {
            int v; // C# 5.0 DECLARED INSIDE THE LOOP
            v = (int)(int)e.Current;
            Thread t = new Thread(() => Run(v));
            threads.Add(t);
            t.Start();
        }
    }
    finally
    {
        if (e != null) ((IDisposable)e).Dispose();
    }
    // Wait for all threads to complete
    foreach (Thread t in threads)
        t.Join();
}

Because my local machine and server were running different versions of .NET, the same exact code was producing totally different results.

Eventually I found Eric Lippert’s article about the matter. Since I’m still fairly new to the world of .NET, I wasn’t around for the big debate that took place in his comment’s section regarding which should be the correct implementation. However, it is interesting to note that the C# devs decided to switch the logic on how the foreach loop operates so late in the game.

My eventual workaround for the .NET 3.5/C# 4.0 server was to assign the int to a newly created variable inside the foreach:

foreach (var value in values)
{
    var tempValue = value; // THE FIX
    Thread t = new Thread(() => Run(tempValue));
    threads.Add(t);
    t.Start();
}

As frustrating it may be to debug problems like this, it is nice to learn a little bit more of the language’s history and idiosyncrasies.

Thanks for reading. You might also enjoy following me on Twitter.

Want to learn even more SQL?

Sign up for my newsletter to receive weekly SQL tips!

JSON Support Is The Best New Developer Feature in SQL 2016 — Part 4: Performance Comparisons

Published on: 2017-02-21

This is the fourth article in my series about learning how to use SQL Server 2016’s new JSON functions. If you haven’t already, you can read Part 1 — Parsing JSON, Part 2 — Creating JSON, and Part 3 — Updating, Adding, and Deleting JSON.


Additional performance comparisons available in an updated post.

We’ve finally come to my favorite part of analyzing any new software feature: performance testing. SQL Server 2016’s new JSON functions are great for parsing JSON data, creating JSON data, and modifying JSON data, but are they efficient?

Today we’ll examine three areas of SQL Server JSON performance:

  1. How to maximize performance for the new SQL Server JSON functions
  2. How the new SQL Server JSON functions compare against what was previously available in SQL Server
  3. How the new SQL JSON functions compare against Newtonsoft’s Json.NET

Maximizing SQL Server JSON Function Performance

I wanted to use a sizable data set in order to test the performance of the new JSON functions in SQL Server 2016. I found arthurkao’s car year/make/model data on GitHub and decided this ~20k element JSON array would be perfect for performance testing purposes. For my tests I’ll be using both the original JSON string as well as a SQL table that I created from the original JSON array:

-- Car data source: https://github.com/arthurkao/vehicle-make-model-data
IF OBJECT_ID('dbo.Cars') IS NOT NULL 
BEGIN
	DROP TABLE dbo.Cars;
END
CREATE TABLE dbo.Cars
(
	Id INT IDENTITY(1,1),
	CarDetails NVARCHAR(MAX)
);
-- See https://gist.github.com/bertwagner/1df2531676112c24cd1ab298fc750eb2 for the full untruncated version of this code
DECLARE @cars nvarchar(max) = '[ {"year":2001,"make":"ACURA","model":"CL"}, {"year":2001,"make":"ACURA","model":"EL"},...]';

INSERT INTO dbo.Cars (CarDetails)
SELECT value FROM OPENJSON(@cars, '$');

SELECT * FROM dbo.Cars;
/* 
Output:
Id          CarDetails
----------- ----------------------------------------------
1           {"year":2001,"make":"ACURA","model":"CL"}
2           {"year":2001,"make":"ACURA","model":"EL"}
3           {"year":2001,"make":"ACURA","model":"INTEGRA"}
...
*/

Unlike XML in SQL Server (which is stored in it’s own datatype), JSON in SQL Server 2016 is stored as an NVARCHAR. This means instead of needing to use special indexes, we can use indexes that we are already familiar with.

To maximize performance, we can use Microsoft’s recommendation of adding a computed column for one of the JSON properties and then indexing that computed column:

-- Remember to turn on "Include Actual Execution Plan" for all of these examples

-- Before we add any computed columns/indexes, let's see our execution plan for our SQL statement with a JSON predicate
SELECT * FROM dbo.Cars WHERE JSON_VALUE(CarDetails, '$.model') = 'Golf'
/*
Output:
Id          CarDetails
----------- --------------------------------------------------
1113        {"year":2001,"make":"VOLKSWAGEN","model":"GOLF"}
2410        {"year":2002,"make":"VOLKSWAGEN","model":"GOLF"}
3707        {"year":2003,"make":"VOLKSWAGEN","model":"GOLF"}
...
*/
-- The execution plan shows a Table Scan, not very efficient

-- We can now add a non-persisted computed column for our "model" JSON property.
ALTER TABLE dbo.Cars
ADD CarModel AS JSON_VALUE(CarDetails, '$.model');

-- We add the distinct to avoid parameter sniffing issues.  
-- Our execution plan now shows the extra computation that is occuring for every row of the table scan.
SELECT DISTINCT * FROM dbo.Cars WHERE JSON_VALUE(CarDetails, '$.model') = 'Golf'
SELECT DISTINCT * FROM dbo.Cars WHERE CarModel = 'Golf'

Non-persisted computed columns (like in the example above) do not take up any additional space in the table. You can verify this for yourself by running sp_spaceused 'dbo.Cars' before and after adding the non-persisted column to the table.

Having a computed column doesn’t add any performance to our query on its own but it does now allow us to add an index to our parsed/computed JSON property.

Having the computed column doesn’t improve performance — we are still seeing a Table Scan

The clustered index that we add next stores pointers to each parsed/computed value causing the table not to take up any space and only causes the SQL engine to recompute the columns when the index needs to be rebuilt:

-- Add an index onto our computed column
CREATE CLUSTERED INDEX CL_CarModel ON dbo.Cars (CarModel)

-- Check the execution plans again
SELECT DISTINCT * FROM dbo.Cars WHERE JSON_VALUE(CarDetails, '$.model') = 'Golf'
SELECT DISTINCT * FROM dbo.Cars WHERE CarModel = 'Golf'
-- We now get index seeks!

And the resulting execution plan now shows both queries (the one using JSON_VALUE() in the WHERE clause directly as well the one calling our computed column) using index seeks to find the data we are looking for:

Yay index seeks!

Overall, adding computed columns to our table adds no overhead in terms of storage space and allows us to then add indexes on JSON properties which improve performance significantly.

SQL Server 2016 JSON vs SQL Server pre-2016 JSON

As I’ve mentioned before, the best option for processing JSON data in SQL Server before 2016 was by using Phil Factor’s amazing JSON parsing function. Although the function works well, it is limited by what SQL Server functionality was available at the time and therefore wasn’t all that efficient.

-- Let's compare how quick Phil Factor's JSON parsing function does against the new SQL 2016 functions
-- Phil's parseJSON function can be downloaded from https://www.simple-talk.com/sql/t-sql-programming/consuming-json-strings-in-sql-server/

SELECT years.StringValue AS Year, makes.StringValue AS Make, models.StringValue AS Model FROM dbo.parseJSON(@cars) models
INNER JOIN dbo.parseJSON(@cars) years ON models.parent_ID = years.parent_ID
INNER JOIN dbo.parseJSON(@cars) makes ON models.parent_ID = makes.parent_ID
WHERE models.NAME = 'model' AND models.StringValue = 'Golf' AND years.NAME = 'year' AND makes.NAME = 'make'

The above query should work for getting the data we need. I’m abusing what the parseJSON function was probably built to do (I don’t think it was intended to parse ~20k element JSON arrays), and I’ll be honest I waited 10 minutes before killing the query. Basically, trying to parse this much data in SQL before 2016 just wasn’t possible (unless you wrote CLR).

Compared to the following queries which is using our indexed computed column SQL Server 2016 is able to return all of the results to us in 1 ms:

-- Indexed computed column returns results in ~1ms
SELECT * FROM dbo.Cars WHERE CarModel = 'Golf'

SQL Server 2016 JSON vs Newtonsoft’s Json.NET

In cases like the above where parsing JSON in SQL Server was never an option, my preferred method has always been to parse data in C#. In particular, Newtonsoft’s Json.NET is the standard for high performance JSON parsing, so let’s take a look at how SQL Server 2016 compares to that.

The following code shows 6 tests I ran in SQL Server 2016:

-- Turn on stats and see how long it takes to parse the ~20k JSON array elements
SET STATISTICS TIME ON

-- Test #1
-- Test how long it takes to parse each property from all ~20k elements from the JSON array
-- SQL returns this query in ~546ms
SELECT JSON_VALUE(value, '$.year') AS [Year], JSON_VALUE(value, '$.make') AS Make, JSON_VALUE(value, '$.model') AS Model FROM OPENJSON(@cars, '$') 

-- Test #2
-- Time to deserialize and query just Golfs without computed column + index
-- This takes ~255ms in SQL Server
SELECT * FROM OPENJSON(@cars, '$') WHERE JSON_VALUE(value, '$.model') = 'Golf'

-- Test #3
-- Time it takes to compute the same query for Golf's with a computed column and clustered index 
-- This takes ~1ms on SQL Server
SELECT * FROM dbo.Cars WHERE CarModel = 'Golf'

-- Test #4
-- Serializing data on SQL Server takes ~110ms
SELECT * FROM dbo.Cars FOR JSON AUTO

-- What about serializing/deserializing smaller JSON datasets?
-- Let's create our smaller set
DECLARE @carsSmall nvarchar(max) = '[ {"year":2001,"make":"ACURA","model":"CL"}, {"year":2001,"make":"ACURA","model":"EL"}, {"year":2001,"make":"ACURA","model":"INTEGRA"}, {"year":2001,"make":"ACURA","model":"MDX"}, {"year":2001,"make":"ACURA","model":"NSX"}, {"year":2001,"make":"ACURA","model":"RL"}, {"year":2001,"make":"ACURA","model":"TL"}]';

-- Test #5
-- Running our query results in the data becoming deserialized in ~0ms
SELECT JSON_VALUE(value, '$.year') AS [Year], JSON_VALUE(value, '$.make') AS Make, JSON_VALUE(value, '$.model') AS Model FROM OPENJSON(@carsSmall, '$') 
--30ms in sql

-- Test #6
-- And serialized in ~0ms
SELECT TOP 7  * FROM dbo.Cars FOR JSON AUTO

And then the same tests in a C# console app using Json.Net:

static void Main(string[] args)
        {
            string cars = @"[ {""year"":2001,""make"":""ACURA"",""model"":""CL""}, ... ]";
            Stopwatch stopwatch = new Stopwatch();

            // Test #1
            stopwatch.Start();
            var deserializedCars = JsonConvert.DeserializeObject<IEnumerable<Car>>(cars);
            stopwatch.Stop();
            long elapsedMillisecondsDeserialize = stopwatch.ElapsedMilliseconds;

            // Test #2 & #3
            stopwatch.Restart();
            var queriedCars = JsonConvert.DeserializeObject<IEnumerable<Car>>(cars).Where(x=>x.model == "Golf");
            stopwatch.Stop();
            long elapsedMillisecondsQuery = stopwatch.ElapsedMilliseconds;

            // Test #4
            stopwatch.Restart();
            var serializedCars = JsonConvert.SerializeObject(deserializedCars);
            stopwatch.Stop();
            long elapsedMillisecondsSerialize = stopwatch.ElapsedMilliseconds;

            // smaller data
            string carsSmall = @"[ {""year"":2001,""make"":""ACURA"",""model"":""CL""}, {""year"":2001,""make"":""ACURA"",""model"":""EL""}, {""year"":2001,""make"":""ACURA"",""model"":""INTEGRA""}, {""year"":2001,""make"":""ACURA"",""model"":""MDX""}, {""year"":2001,""make"":""ACURA"",""model"":""NSX""}, {""year"":2001,""make"":""ACURA"",""model"":""RL""}, {""year"":2001,""make"":""ACURA"",""model"":""TL""}]";

            // Test #5
            stopwatch.Restart();
            var deserializedCarsSmall = JsonConvert.DeserializeObject<IEnumerable<Car>>(carsSmall);
            stopwatch.Stop();
            long elapsedMillisecondsDeserializeSmall = stopwatch.ElapsedMilliseconds;

            // Test #6
            stopwatch.Restart();
            var serializedCarsSmall = JsonConvert.SerializeObject(deserializedCarsSmall);
            stopwatch.Stop();
            long elapsedMillisecondsSerializeSmall = stopwatch.ElapsedMilliseconds;
        }

And the results compared side by side:

Essentially, it seems like Json.Net beats SQL Server 2016 on larger JSON manipulations, both are equal with small JSON objects, and SQL Server 2016 has the advantage at filtering JSON data when indexes are used.

Conclusion

SQL Server 2016 is excellent at working with JSON. Even though Json.NET beats SQL Server 2016 at working with large JSON objects (on the magnitude of milliseconds), SQL Server is equally fast on smaller objects and is advantageous when JSON data needs to be filtered or searched.

I look forward to using the SQL Server 2016 JSON functions more in the future, especially in instances where network I/O benefits me to process JSON on the SQL Server or when working with applications that cannot process JSON data, like SQL Server Reporting Services.

Thanks for reading. You might also enjoy following me on Twitter.

Want to learn even more SQL?

Sign up for my newsletter to receive weekly SQL tips!

XmlReader vs XmlDocument Performance

Published on: 2016-11-24

Running C# Tests for Code Performance

Recently I have been working on a project where I needed to parse XML files that were between 5mb and 20mb in size. Performance was critical for the project, so I wanted to make sure that I would parse these files as quickly as possible.

The two C# classes that I know of for parsing XML are XmlReader and XmlDocument. Based on my understanding of the two classes, XmlReader should perform faster in my scenario because it reads through an XML document only once, never storing more than the current node in memory. On the contrary, XmlDocument stores the whole XML file in memory which has some performance overhead.

Not knowing for certain which method I should use, I decided to write a quick performance test to measure the actual results of these two classes.

The Data

In my project, I knew what data I needed to extract from the XML up front so I decided to configure test in a way that mimics that requirement. If my project required me to run recursive logic in the XML document, needing a piece of information further down in the XML in order to know what pieces of information to pull earlier on from the XML, I would have set up an entirely different test.

For my test, I decided to use the Photography Stack Exchange user data dump as our sample file since it mimics the structure and file size of one my actual project’s data. The Stack Exchange data dumps are great sample data sets because they involve real-world data and are released under a Creative Commons license.

The Test

The C# code for my test can be found in its entirety on GitHub.

In my test I created two methods to extract the same exact data from the XML; one of the methods used XmlReader and the other XmlDocument.

The first test uses XmlReader. The XmlReader object only stores a single node in memory at a time, so in order to read through the whole document we need to usewhile(reader.Read()) in order to loop all of the nodes. Inside of the loop, we check if each node is an element that we are looking for and if so then parse out the necessary data:

public static void XmlReaderTest(string filePath)
{
    // We create storage for ids of all of the rows from users where reputation == 1
    List<string> singleRepRowIds = new List<string>();

    using (XmlReader reader = XmlReader.Create(filePath))
    {
        while (reader.Read())
        {
            if (reader.IsStartElement())
            {
                if (reader.Name == "row" && reader.GetAttribute("Reputation") == "1")
                {
                    singleRepRowIds.Add(reader.GetAttribute("Id"));
                }
            }
        }
    }
}

On the other hand, the code for XmlDocument is much simpler: we load the whole XML file into memory and then write a LINQ query to find the elements of interest:

public static void XmlDocumentTest(string filePath)
{
    List<string> singleRepRowIds = new List<string>();

    XmlDocument doc = new XmlDocument();
    doc.Load(filePath);

    singleRepRowIds = doc.GetElementsByTagName("row").Cast<XmlNode>().Where(x => x.Attributes["Reputation"].InnerText == "1").Select(x => x.Attributes["Id"].InnerText).ToList();
}

After writing these two methods and confirming that they are returning the same exact results it was time to pit them against each other. I wrote a method to run each of my two XML parsing methods above 50 times and to take the average elapsed run time of each to eliminate any outlier data:

public static double RunPerformanceTest(string filePath, Action<string> performanceTestMethod)
{
    Stopwatch sw = new Stopwatch();

    int iterations = 50;
    double elapsedMilliseconds = 0;

    // Run the method 50 times to rule out any bias.
    for (var i = 0; i < iterations; i++)
    {
        sw.Restart();
        performanceTestMethod(filePath);
        sw.Stop();

        elapsedMilliseconds += sw.ElapsedMilliseconds;
    }

    // Calculate the average elapsed seconds per run
    double avergeSeconds = (elapsedMilliseconds / iterations) / 1000.0;

    return avergeSeconds;
}

Results and Conclusions

Cutting to the chase, XmlReader performed faster in my test:

Performance test results.

Now, is this ~.14 seconds of speed difference significant? In my case, it is, because I will be parsing many more elements and many more files dozens of times a day. After doing the math, I estimate I will save 45–60 seconds of parsing time for each set of XML files, which is huge in an almost-real-time system.

Would I have come to the same conclusion if blazing fast speed was not one of my requirements? No, I would probably go the XmlDocument route because the code is much cleaner and therefore easier to maintain.

And if my XML files were 50mb, 500mb, or 5gb in size? I would probably still use XmlReader at that point because trying to store 5gb of data in memory will not be pretty.

What about a scenario where I need to go backwards in my XML document — this might be a case where I would use XmlDocument because it is more convenient to go backwards and forwards with that class. However, a hybrid approach might be my best option if the data allows it: if I can use XmlReader to get through the bulk of my content quickly and then load just certain child trees of elements into XmlDocument for easier backwards/forwards traversal, then that would seem like an ideal scenario.

In short, XmlReader was faster than XmlDocumet for me in my scenario. The only way I could come to this conclusion though was by running some real world tests and measuring the performance data.

So should you use XmlReader or XmlDocument in your next project? The answer is it depends.

Thanks for reading. You might also enjoy following me on Twitter.

Want to learn even more SQL?

Sign up for my newsletter to receive weekly SQL tips!