Go embedded structs and pointer receivers

While working on a PR for the Azure storage GO library I discovered (yet another) corner of Go that confused the hell out of me. I know enough about embedded types/structs to use them effectively (so I thought) and I know the difference between receivers and pointer receivers. But when using them together you really need to make sure you’re careful.

The specifics of the problem I hit was that in the Azure GO library we have the struct called “Entity” which is used for table storage. Now, for purposes I wont go into I needed to create a another struct called BatchEntity which was declared as such:

type BatchEntity struct {
Entity
otherstuff string
}

Now, the Entity struct has a very useful JSON formatting method with the signature:

func (e *Entity) MarshalJSON() ([]byte, error)

Please please PLEASE note that this is a pointer receiver! Now I’d spend an embarrassingly long time trying to figure out why when I tried to marshal my BatchEntity class, why it didn’t call this MarshalJSON method as I expected. Of course it turned out that when I passed the marshalling method BatchEntity.Entity (referencing the embedded type) it would *somehow* (I haven’t figure out the specifics yet) treat it as a regular receiver and not call the MarshalJSON method since that specifically had a pointer receiver. I still need to dig into the details but this is a big warning to me, when dealing with embedded types make sure you check the receivers of the base structs. Would have saved me hours!!!

Kind of surprised there wasn’t a compile warning/error around that….   but I need to dig further.

AzureCopy Go now with added CopyBlob flag

Azurecopy (Go version) 0.2.2 has now been released. Major benefit is now that when copying to Azure we can use the absolutely AWESOME CopyBlob functionality Azure provides. This allows blobs to be copied from S3 (for example) to Azure without having to go via the machine executing the instructions (and use my bandwidth!)

An example of copying from S3 to Azure is as simple as:

azurecopycommand_windows_amd64.exe  -S3DefaultAccessID=”S3 Access ID” -S3DefaultAccessSecret=”S3 Access Secret” -S3DefaultRegion=”us-west-2″ -dest=”https://myaccount.blob.core.windows.net/mycontainer/” -AzureDefaultAccountName=”myaccount” -AzureDefaultAccountKey=”Azure key” -source=https://s3.amazonaws.com/mybucket/ –copyblob

The key thing is the –copyblob flag. This tells AzureCopy to do it’s magic!

By default AzureCopy-Go will copy 5 blobs concurrently, this is to not overload your own bandwidth, but with Azure CopyBlob feature feel free to crank that setting up using the –cc flag (eg add -cc=20) Smile

Azure Functions (and Go)

I’m a HUGE Azure Webjob fan, they’re so useful for a wide variety of scenarios. Being able to setup a Webjob listening to an input queue and just throw messages at it is such a nice way to do adhoc data processing. Azure Functions is a nice improvement over Webjobs (but basically do the same thing). Whereas Webjobs are constantly running in the context of you Application Service (and potentially using up your App Service quotas), Azure Functions aren’t tied to App Service plans and best of all, you’re only charged for when they’re actually processing something. Setup them up and if they don’t process anything for a month, you get charged nothing. You don’t need to tear them down and restart them, they’re just there ready and waiting.

There are plenty of tutorials/pages on Azure Functions, so I wont go into the general details of how they work, but what I will touch on it how to run Go in them. Now, Go within Azure Functions isn’t a first class citizen. Javascript and C# are first class citizens; Azure Functions know how to execute those and it does that well. One nice feature Microsoft has provided is the ability to run batch files. Batch files can in run anything <queue evil laugh/>

What I’m describing isn’t specific to Go, infact I blatantly stole it from an Azure website describing how to do this with Java. But I’m not a Java fan-boy…

Firstly, setting up an Azure Function is pretty straight forward (under a minute to complete). Essentially go to the Azure Functions Portal (you can also do it via the regular Azure Portal) , give your function a name and region then hit “create”. It’s THAT easy!

You’ll get redirected to the Azure Portal, select “Create your own custom function”

 

customfunction

 

Then you have a LOT of options for triggering the processing of data. Expand to see all the options, then select QueueTrigger-Batch (if you want to use queues that is, which I do).

 

queuetrigger

Now that we have configured the function will really be triggered by a batch file we can make it in turn call our Go compiled binary (yes, I keep talking about Go here, but in reality this will work for any executable regardless of the language it was coded in).

azurefunction1

The input queue message is converted into an environment variable (see line 2), then we call azurefunction.exe. Where does Azure Functions know about azurefunction.exe?  Click on the “view files” then “upload”, then you can upload any files required for processing the messages.

One last thing to do in the portal is to specify how the Azure Function returns its results. In my case I also want the output to be via Azure Queues. To enable this, select Integrate on the left menu, select New Output and then Azure Queue Storage.

 

output

 

Now, EVERYTHING so far has had nothing to do with Go. So now we’ve done the Azure plumbing, let’s actually Go do some work (BOOM BOOM!)

In this example my super sophisticated Azure Function will count the number of characters in the input message and send that result to the output message. Rocket Surgery for the win!

The entire source for my Go application is:

 

package main

import (
“fmt”
“log”
_ “net/http/pprof”
“os”
“strconv”
)

func main() {
inputMessage := os.Getenv(“inputMessage”)
outputMessageLocation := os.Getenv(“outputQueueItem”)

if inputMessage != “” {
// just write the length to the queue location. Just to prove this works.
cacheFile, err := os.OpenFile(outputMessageLocation, os.O_WRONLY|os.O_CREATE, 0)
if err != nil {
log.Fatalf(“Unable to open file %s %s”, outputMessageLocation, err)
}
l := len(inputMessage)
s := strconv.Itoa(l)
cacheFile.WriteString(s)
}
}

=============================

The code is extremely simple. The input messages comes from the “inputMessage” environment variable. We get the location of the output file via the “outputQueueItem” environment variable. We create a file at that location, write the results (length of input) and hey presto, one usable Azure Function!

Now, we compile this and upload via the Azure portal (mentioned earlier). Now for testing!

Select “Develop” on the left hand menu on the Azure Portal (see screenshot below) and then on the “test” button on the right hand side. This will allow you put some test inputs on the queue and check the results.

develop

 

Enter some test data. In the above case I just entered “abcde” then hit RUN. You can see some of the debugging in the lower window. The input has been read, data processed and result thrown onto a queue (check your azure queue with your favourite azure viewing tool).

Now, given this is all just a batch file wrapper, you can see that the time to execute isn’t brilliant. You can see the entire process was about 750ms, and 600ms of that seems to be just the firing up of the azurefunction.exe executable from the batch file. Still, I think that being able to set these up, and “almost” fire and forget about them is really nice. You’re only charged for the time they’re running.

I’ll definitely be tinkering with these more, I can see them being hugely useful!

Lets GO OS crazy with AzureCopy-Go

Being able to copy from one cloud provider to another is useful, but if everything is purely serial (ie one blob at a time) the time taken to copy everything might be less than stellar. I’ve now released a new version of AzureCopy-Go (0.2.1) which allows concurrently copying of blobs. The default is 5, but using the –cc flag (concurrent copying) it can be expanded up to 1000 (arbitrary max limit). So far, so good!

Also, for this release I’ve build the binaries using the AMAZING Gox project. This allows for easy cross compiling for Go. So we now have Linux, FreeBSD, NetBSD, OpenBSD, Darwin (MacOS) and Windows binaries. For the most part we have 3 variations of each platform binary, one each for ARM, AMD64 and 386.

I knew how to get cross compiling with Go on Linux/MacOS but could never get it working on Windows (current main OS). Gox is definitely a time saver and is so damned easy to use.

Please give AzureCopy-Go 0.2.1 a try if you have any S3 <—> Azure migration needs. More features being worked on every few days.

Azurecopy (GO version) pre-release

As mentioned in previous posts, I’ve been writing a GO version of AzureCopy so people would have something that works cross platform (Linux, MacOS and Windows). Today I’ve released the first pre-release just to test the waters. It supports Windows only (simply due to not having compiled up the other platforms yet), and only supports local filesystem, S3 and Azure Blob Storage.

Baby steps.

The plan is to build for Linux,OSX, then start adding other cloud platforms. Meanwhile the original Azurecopy (Windows only, full dotnet framework) will still be developed (mainly from a Nuget/library point of view).  If you just need an executable to perform copying, then I suggest using this newer version.

Some examples of using this  newer version:

image1

In this case we’re just listing the contents of my testken123 (super secret) bucket. My AccessID and AccessSecret are passed in via command line options. The output format is in a basic tree structure (will add in a bog-standard list soon). In the above case, the top of the tree is “testken123” which is the bucket name. Under that we have 2 virtual directories (remember Azure/S3 etc do not really have directories but fake it by using / as a delimiter). In this case we see there is a blob called “ken1/test1” which treats the “ken1” part as a directory and “test1” as the blob name. Same applies for all the other results.  Simple enough.

Then we have:

image2

In this case we’re copying from my local filesystem (c:\temp\data\s3\) into the S3 bucket testken123. The console output is just to show what is going to be copied. Output will be modified to show progress.

Finally we have:

image3

That’s coping from Azure Blob Storage to S3. Same deal, basic output.

For every command it is possible to pass the “-debug” flag. This makes things VERY verbose but is extremely useful for figuring out issues.

This is just a first step, pre-release, uber new version. Please give it a go and let me know if there are any issues. The plan is to start cranking out changes pretty frequently.

0.1.0 version

AzureCopy GO

The Go version of AzureCopy is slowly making progress. So far I’ve just been focusing on local filesystem and Azure (since I can do those while offline on the train commute thanks to the Azure Storage Emulator). The next plan is for S3 integration, primarily because S3 –> Azure seems to be the big use case for the original AzureCopy.

I’m planning on frequent releases once the basic S3 code is added (hopefully within the next few days). Not all features from the original AzureCopy will be available, but will simply be focusing on 1) list content and 2) copy content. There will be a few new additions such as a “don’t overwrite” flag so copies can be continued after being stopped (has been requested by a few people).

Ofcourse, the original AzureCopy will still be developed (mainly from a Nuget packaging point of view) but if you just need a command line tool to copy (and maybe need it on multiple platforms) then this new version is probably the way to go.

Hopefully the S3 code will drop in a few days then I’ll have a first binary release for Linux, MacOS and Windows, and see how things proceed from there.

Lets Go Profiling

Recently I’ve been going through some experimenting with the Go profiler (on Windows, YMMV for other platforms). This is both the best and worse profiler I’ve ever used. Firstly, I want to address the bad part (since that’s relatively small compared to the great stuff).

Not so good

The profiler is a sampling profiler (as many are these days) but I’ve found that in many runs of the profiler, entire chunks of code are being missed. In my particular case I’m trying to profile the put/get queue methods of the Azure SDK. I have a single method that puts a few thousand messages onto the queue then retrieves them.

In the resulting profile run the puts are recorded but the gets are completely missed. I rerun the test (literally “up arrow, enter”) and I’ll get both puts and gets. Turns out one recommended way to improve the situation is to allow the profiler to run for a longer period of time. This definitely improved things but I still often had cases where entire code branches were simply missed :/

I haven’t tried it on any other platform than Windows, so maybe this is a common issue (doubt it) or if it’s just an OS issue.

The good stuff.

The simple command of “web” is fricken awesome. It produces a nice SVG that gives you a nice tree of function calls with the appropriate statistics (memory/CPU usage).

The way we generate the profile information is by running the go executable in testing and benchmark mode. For example, for profiling the Azure SDK (queuing in particular) I run the command:

go test –bench=. –benchtime=10s –run=XXX –cpuprofile=prof10s.cpu

What this does is run the test files (ie the *_test.go files) and also runs the benchmarking methods in those files (any method starting with “Benchmark” which accept the parameter *testing.B).

I’ve also told it to run for a minimum of 10 seconds. The result is a profile file called prof10s.cpu.

To load it simple run: go tool pprof  .\storage.test.exe .\prof10s.cpu

Note, storage.test.exe was produced during the test/benchmark run.

Now things get interesting. Say, to get to the top 10 CPU hungry function calls, the command “top –cum 10” can be used. In my case the results were:

top10

This information, although in theory is useful it’s showing me results for functions that setup the tests, outputted results etc, but not really the function that actually put some messages onto the queue. For that, we can specify where the profiler should focus on, such as :

top –cum PutMessage

top10-putmessage

Here we can see more useful function calls such as the ones in storage.Client.*, these are the ones I want to see the performance of.

Now, I’m a simple soul… (just ask my wife) and pictures paint such a nicer view on things. For that we can use the “web” command, or more specifically for this case “web PutMessage”. This generates a lovely SVG which makes things really clear to the user. A small snippet of the SVG looks like this:

 

topweb

This is just a small snippet of a far larger graph. But you can clearly see the major code paths of the “PutMessage” function. Where the time goes (ms times) but also the nice big bold boxes that subtly shout “LOOK HERE!!!”. This is very useful!

Others have blogged far more extensively than I ever will about the subject (eg Dave Cheney). I’m only starting out on Go let alone profiling, but it’s a seriously nice place to be.