One of the oft-touted benefits of Go is that applications written in it are easily deployed because they are statically complied. A lot of this benefit goes away if you need to manage the location and permissions on a bunch of files needed to run a web application.
The solution is to compile any necessary files into the application binary itself. This can be done in Go by using a byte slice literal containing the string representation of the bytes in a file.
fileData := byte("\x1f\x8b\ ... \x0f\x00\x00")
The biggest downsides to this approach are:
- Larger binaries
- For my current project Lex Library I’m seeing a 20 MB executable without embedded assets, and a 21 MB executable with them.
- Longer compilation
- This is mostly mitigated by the latest caching in the Go compiler.
- Increased memory during compilation
- This may bite you if you are developing on low memory devices, but for me personally this hasn’t been a concern.
You are essentially making a trade-off between development time and operational management time. If your application’s intended audience is the general public (or at least the really geeky public that host their own web apps), this trade off is more than worthwhile.
Perhaps the first library, or at least the first really popular one, to handle embedding static assets in Go applications
is jteeuwen’s Go-BinData. It’s a command line application that you pass a
directory path to and it generates a
.go file with your assets embedded in it.
Unfortunately, jteeuwen seems to have dropped off the face of the planet, deleting all of this repositories from github as he left. Luckily for all of us, his code was open-source and so was forked by the community at large. You can find several, well-maintained forks of his code on github. My fork of choice was shuLhan’s, but I have since moved onto other options, the reasons for which, I’ll get into below.
More details on the jteewen repo are here: https://github.com/jteeuwen/go-bindata/issues/5
Since jteewen’s library there have been many, many alternatives. Below is a non-comprensive list I put together while researching this myself:
- vfsgen - https://github.com/shurcooL/vfsgen
- go.rice - https://github.com/GeertJohan/go.rice
- statik - https://github.com/rakyll/statik
- esc - https://github.com/mjibson/esc
- go-embed - https://github.com/pyros2097/go-embed
- go-resources - https://github.com/omeid/go-resources
- packr - https://github.com/gobuffalo/packr
- statics - https://github.com/go-playground/statics
- templify - https://github.com/wlbr/templify
- gnoso/go-bindata - https://github.com/gnoso/go-bindata
- shuLhan/go-bindata - https://github.com/shuLhan/go-bindata
- fileb0x - https://github.com/UnnoTed/fileb0x
- gobundle - https://github.com/alecthomas/gobundle
- parcello - https://github.com/phogolabs/parcello
The main goal of this post is to help you sort through differences and help you determine which features you should consider when choosing one of these libraries.
Separating the Wheat from the Chaff
With so many options, it can be overwhelming to determine which library will work best for your purposes. While your application may have different requirements than mine, chances are if it’s a web application, there will be a lot of overlap. Hopefully this comparison of the libraries below will be useful if you need to make this same decision.
As mentioned before, one of the downsides to embedding static files is that it increases the size of the executable. You can alleviate some of that by using a library that compresses the files before embedding them. It is usually worth the small cost of decompression to save space as well as less memory usage when compiling. Also, static web files tend to compress really well (other than images), which is why most web browsers support receiving gzip compressed files. This leads directly to my next criterion.
If your static files are stored in your executable with gzip compression, and you are going to potentially serve up those files to your client with gzip compression, why not just send them the already compressed file data? The ideal library would give you the option, at runtime, to retrieve the already compressed file without decompressing it first.
Loading from the local File System
When you are developing your web application, anything that adds time or friction between when you make a change and when you see a change in your app, should be limited. If you have to rebuild your statically embedded go file every time you make a change to CSS or HTML, you’ll quickly be looking for alternatives.
The ideal library should allow you to easily switch between a development build, where the files are loaded locally and on-the-fly, and your production build, where the files are fully embedded into executable and ready for distribution.
This criterion took me by surprise, and I didn’t consider it originally when I started development on Lex Library. As mentioned earlier, my first choice was a go-bindata fork by shuLhan. I chose it mainly because I was already familiar with the original go-bindata library, and this fork looked well maintained.
The library was working great, but, out of the blue, my CI builds started failing. Like every developer when their tests start failing, my first thought was what changed. I immediately checked my last commit, and tried to figure out how that change broke my template handling. After scratching my head for a while, unable to find a culprit, I re-ran my test suite against the last successful build and found that those too were failing. That told me the change had to be environmental, and not in my code. However, that raised more questions.
I run my Continuous Integration tests in Docker containers. The environments should be
self-contained, pristine, and reproducible. Somewhere my assumptions about that were wrong. Stepping through my
Dockerfile, I found my culprit:
RUN go get -u github.com/shuLhan/go-bindata/...
There was a small update to the go-bindata library that broke how I was passing in the path to my static files. All of a sudden my embedded files weren’t on the paths I was expecting. Now, this could be blamed on several factors such as the fact that go get always gets the default branch. In the end, it came down to the simple truth that a process which was generating code in my application wasn’t being versioned and tracked inside my git repo. External changes on a dependency of mine could break my builds or my tests without me knowing.
One way to fix this issue was to store a copy of the pre-compiled go-bindata executable in my git repo, but:
- It’s usually not a good idea to store binary blobs in git.
- I’d have to manually update it every time there was a bug fix.
- It would make it much harder to build on any other platform other than the one I develop on.
Alternately, I could find a library that didn’t require a stand-alone executable, and relied entirely on code that
is vendored in my git repo. This meant a library that supported
go generate, and more specifically, one that didn’t
rely on an external executable to run.
As I mentioned before, you may have different requirements than I do, so in my comparison table below I included a few additional criteria that you may find useful.
If you have a lot of different folders and files to manage, it can be easier to manage them via a config file checked into your source control.
Using a library that satisfies the http.FileSystem interface can make it much easier to work with the embedded files.
Greater than 200 Github Stars
While this is a bit arbitrary, and it’s not necessarily a measure of quality, a repo’s number of stars can be a good indicator of an active library, or at least one that’s used in a lot of places. This, in turn, could mean that a lot of people are battle testing it, and / or submitting bug reports. The library I chose, just barely made it past this arbitrary number of stars, so keep that in mind.
|Library||Compression||Opt. decompression||Local FS||go generate||No EXE||Config File||http.FS||> 200 Stars|
* Additional code required
My experience with these libraries vary from writing and deploying applications with them, to simply looking through
README and documentation. If you find any inaccuracies in the table, please let me know in the comments and
I’ll update them.
The comparison table makes it pretty clear why I ended up going with vfsgen, and I
highly recommend it, especially if you require reproducible builds. A close runner up was
fileb0x, but unfortunately, it’s usage of
go generate required a stand-alone
executable to run.