Monday, February 24, 2020

Strongly Typed, Dependency Managed Azure in C#: Introducing Cake.AzureCLI

The story, nay legend, of providing strongly typed, cross platform, dependency managed access to all 2,935 Azure CLI commands in C#.

You can now have strongly typed, cross platform, dependency managed access to all 2,935 Azure CLI commands in C#, with full intellisense including examples. That's because I just published a Cake plugin for AzureCLI called Cake.AzureCli.

This blog post is a little about what it is and how to use it, but it's more about how I built it. That's because I had a blast solving this problem and my solution might even entertain you: parsing thousands of help files through the CLI, storing results in 16 meg intermediate JSON files, and code generating 276K lines of code with T4 templates.

In the process I apparently also broke Cake's static site generator.


But first, I suppose the most relevant information is the what and the how.

Have & Eat Your Cake.AzureCLI

This plugin runs in Cake. If you aren't familiar with cake, please check out Code Hour Episode 16, Intro to Cake, where I go over what it is and why you should care.

If you don't have a spare hour right now: it's a dependency management system (like make, ant, maven, or rake) except in C#. It also has a huge plugin ecosystem, one that's now "slightly" larger with access to all of Azure CLI.

Right, you didn't watch the video, and you're still skeptical, right? You're wondering what was wrong with the official azure-sdk nuget plugins. The answer is: they aren't Cake enabled and so they don't support dependency management. If that statement isn't meaningful to you, please, watch just the "Scripts" section of my talk starting at 9:08.

Now that you're 100% convinced let's dig in. Using Cake.AzureCLI is as simple as adding a preprocessor directive to pull it from NuGet:
#addin "nuget:?package=Cake.AzureCli&version=1.2.0"
And then accessing commands like Az().. So a simple program to log in and list all your resource groups might look like this:

var target = Argument("target""Default"); var username = Argument<string>("username"null); var password = Argument<string>("password"null); Task("Login")    .Does(() => {    // 'az login' is accessed via Az().Login()    Az().Login(new AzLoginSettings {       Username = username,       // all commands can be customized if necessary with a ProcessArgumentBuilder       Arguments = new ProcessArgumentBuilder()          // anything appended with .AppendSecret() will be rendered as [REDACTED]           //    if cake is run with `-verbosity=diagnostic`          .Append("--password").AppendSecret(password)    }); }); Task("ListResourceGroups")    .IsDependentOn("Login"// yayy dependency management!    .Does(() => {    // listing names of all resource groups    Information("Resource Groups:");    // all results are strongly typed as dynamic if results are json    dynamic allResourceGroups = Az().Group.List(new AzGroupListSettings());    foreach (var resourceGroup in allResourceGroups) {       Information(;    } }); RunTarget(target);
And that should hopefully provide enough background to go create sql instances, scale up or down kubernetes clusters, and provision VM's with dependency management, from the comfort of a language you know and love.

The Making Of

"But Lee, I'm dying to know, how did you build this work of art?"
Oh, I'm so very glad you asked. Writing something this large by hand was obviously not going to work. Plus it needs to be easy to update when Azure team releases new versions. Code generation it was. And I always wanted to learn T4 templates.

I first came up with a data structure, always a solid place to start. I wanted something that would support Azure CLI, but that could also be used to generically represent any CLI tool, because ideally this solution could work for other CLI programs as well. I came up with this:

A Program contains a single root Group (az). Groups can contain other Groups recursively (e.g. az contains az aks which contains az aks nodepool). Groups can contain Commands (e.g. az aks contains az aks create). And for documentation Commands can have Examples and Arguments.

It's basically a tree, with Commands as leafs, and so will work nicely in json. But how to populate it?

Fill er up

"Well, that's not how I would have done it"

said a skeptical co-worker when I told him I was executing thousands of az [thing] --help commands and parsing the results. See, AzureCLI was written in Python and is open source, so theoretically I could have downloaded their source and generated what I needed from there in Python.

But I really wanted a more generic approach that I or someone else could apply to any CLI program. So I parsed each "xyz --help" into an intermediary object: a Page. That's basically just a collection of headers, name-value pairs, and paragraphs. Then I converted pages to groups or commands and recursed to produce a 350,385 line, 15 megabyte behemoth.

Incidentally a fun side-effect of this approach is you can see all the changes across Azure CLI version changes e.g. this commit shows changes from 2.0.77 to 2.1.0 (although GitHub doesn't like showing diffs across 16 Meg files in the browser for some reason, can't imagine why).

T4 Templates

I'd never used T4 templates. Turns out they're super awesome. Well, super powerful, and pretty awesome anyway. They are a little annoying when every time you hit save or tab off a .tt file it takes 13 seconds to generate your 178 thousand lines of code -- even on an 8 core i9 with 64 gigs of ram and an SSD. Oh, and then at that scale Visual Studio seems to crash periodically, although I'm sure Resharper doesn't help.

But whatever, they work. And this part is cool: If you set hostspecific=true, then you can access this.Host to get the current directory, read a json file, then deserializing it to model objects that live in a .Core project that you can reference inside of the tt file yet not reference inside of your main project (Cake.AzureCli). If you're interested check out

What to generate was interesting too. The easy part is exposing a method to cake. You just write an extension method like this:

public static class AzAliases {     [CakeMethodAlias]     public static AzCliGroup Az(this ICakeContext context)     {         return new AzCliGroup(context);     } }
And Cake is good to go. But what about generating 2,935 extension methods? Turns out, not such a great idea. The intellisense engine in Visual Studio code is powered by OmniSharp. As awesome as OmniSharp is, it just isn't quite powerful enough to generate intellisense quickly or accurately with that architecture. However, if you group commands into "namespaces" like Az().Aks.Create() instead of AzAksCreate(), then you get nice intellisense at every level:


While this project may not solve world hunger just yet, I do hope it'll make someone's life a little easier. More importantly, I hope this technique will entertain or better yet inspire someone (you) to create something cool. If it does, please let me know about it in the comments or on twitter.

Wednesday, January 22, 2020

Conquer ASP.Net Boilerplate Query Performance in LINQPad, (Announcing LINQPad.ABP)

Ever made it to production only to realize your code fails miserably at scale?  When performance problems rear their gnarly head and there name is EntityFramework, there is but one blade to slice that gordeon knot: LINQPad.  
However, if you're using the ASP.Net Boilerplate (ABP) framework, the situation is a tad more dire. That's because ABP uses a repository pattern, which, it turns out, is less than compatible with LINQPad's DataContext centric approach.
In this post I'll describe two ways to use LINQPad with ABP apps to solve performance problems:
1. Rewrite repository pattern based queries to run in LINQPad with a data context.  This works well for small problems.
2. Enable LINQPad to call directly into your code, support authentication, multi-tenancy, and the unit of work and repository patterns with the help of a NuGet packge I just released called LINQPad.ABP. This helps immensely for more complex performance problems.

Repository Pattern 101

The Repository and Unit of Work patterns used in ASP.Net Boilerplate apps are a wonderful abstraction for simplifying unit testing, enabling annotation based transactions, and handling database connection management.  As described in this article from the Microsoft MVC docs:
The repository and unit of work patterns are intended to create an abstraction layer between the data access layer and the business logic layer of an application. Implementing these patterns can help insulate your application from changes in the data store and can facilitate automated unit testing or test-driven development (TDD)
That document gives a nice diagram to show the difference between an architecture with and without a repository pattern:

However, these abstractions present a problem for LINQPad.  When using LINQPad to diagnose performance problems it would often be super convenient to call directly into your app's code to see the queries translated into SQL and executed.  However, even if LINQPad understood dependency injection, it would have no idea how to populate a
 IRepository, what to do with a [UnitOfWork(TransactionScopeOption.RequiresNew)] attribute, or what value to return for IAbpSession.UserId or IAbpSession.TenantId.  Fortunately, I just released a NuGet Package to make that easy. 
But first, the simplest way to solve the problem for single queries is just to rewrite the query without a repository pattern and paste it into LINQPad.

Undoing Repository Pattern

This is where this blog post gets into the weeds.  If you'd rather watch me perform the following steps please check out my latest episode of Code Hour:

Otherwise, here it is in written form:
Step one is to enable ABP'S data context to support a constructor that takes a connection string.  If you add the following code to your DataContext:
        private string _connectionString;

        /// For LINQPad

        public MyProjDbContext(string connectionString)
            : base(new DbContextOptions())
            _connectionString = connectionString;

        protected override void OnConfiguring(DbContextOptionsBuilder optionsBuilder)
            if (_connectionString == null)
                base.OnConfiguring(optionsBuilder); // Normal operation

            // We have a connection string
            var dbContextOptionsBuilder = new DbContextOptionsBuilder();


Then in LINQPad you can
  1. Add a connection
  2. "Use a typed data context from your own assembly"
  3. Select a Path to Custom Assembly like "MyProjSolution\server\src\MyProj.Web.Host\bin\Debug\netcoreapp2.1\MyProj.EntityFrameworkCore.dll"
  4. Enter a Full Type Name of Typed DbContext like "MyProj.EntityFrameworkCore.MyAppDbContext"
  5. LINQPad should instantiate your DbContext via a constructor that accepts a string, then provide your connection string

Now, when you start a new query you can write:
var thing = this.Things.Where(t => t.Id == 1);

And if you run it you'll see the resulting SQL statement.
Not bad.  If you paste in any real code you'll need to add using statements and replace _thingRepository.GetAll() with this.Things and you'll be translating LINQ to SQL in no time.

Pretty cool.  It certainly works for simple queries.
However, in my experience performance problems rarely crop up in the easy parts of the system.  Performance problems always seem to happen in places where multiple classes interact because there was simply too much logic for the author to have stuffed it all into one class and have been able to sleep at night.

Enter: LINQPad.ABP

To enable LINQPad to call directly into ABP code you'll need to set up dependency injection, define a module that's starts up your core module, specify a current user and tenant to impersonate, and somehow override the default unit of work to use LINQPad's context rather than ABP's.  That's a lot of work.
Fortunately, I just published LINQPad.ABP, an Open Source NuGet Package that does all this for you.  To enable it:
  1. In LINQPad add a reference to "LINQPad.ABP"
  2. Add a custom Module that's dependent on your project's specific EF Module
  3. Create and Cache a LINQPad ABP Context
  4. Start a UnitOfWork and specify the user and tenant you want to impersonate
The code will look like this:
// you may need to depend on additional modules here eg MyProjApplicationModule
// this is a lightweight custom module just for LINQPad
public class LinqPadModule : LinqPadModuleBase
    public LinqPadModule(MyProjEntityFrameworkModule abpProjectNameEntityFrameworkModule)
        // tell your project's EF module to refrain from seeding the DB
        abpProjectNameEntityFrameworkModule.SkipDbSeed = true;

    public override void InitializeServices(ServiceCollection services)
        // add any custom dependency injection registrations here

async Task Main()
    // LINQPad.ABP caches (expensive) module creation in LINQPad's cache
    var abpCtx = Util.Cache(LinqPadAbp.InitModule(), "LinqPadAbp");

    // specify the tenant or user you want to impersonate here
    using (var uowManager = abpCtx.StartUow(this, tenantId: 5, userId: 8))
        // retrieve what you need with IocManager in LINQPad.ABP's context
        var thingService = abpCtx.IocManager.Resolve();
        var entity = await thingService.GetEntityByIdAsync(1045);

That may look like a lot of code, but truest me, it's way simpler than it would otherwise be.  And now you can call into your code and watch every single query and how it gets translated to SQL.  


I hope this helps you track down a hard bug faster.  If maybe then please subscribe, like, comment, and/or let me know on twitter.