[scosman]

What’s the case against cgo for SQLite? Just the usual cgo performance overhead?

It seems like a pretty good cgo use case: a decent amount of work, which is typically slow enough that cgo overhead isn’t perf critical (because DB usually means disk reads), a super robust and well tested C library with a super well maintained cgo wrapper (mattn).

[parhamn]

> Just the usual cgo performance overhead?

No, the performance is certainly orders of magnitude faster than serializing over std streams on a subprocess (c ffi calls in cgo are 10s of nanoseconds).

But one of the big draws of golang is the write-once-compile-anywhere toolchain and calls cgo makes that harder.

[Groxx]

To be a bit more specific here: pure Go binaries are trivial to cross-compile and they Just Work™ basically all the time. `GOOS="darwin" GOARCH="arm64" go build .` and you're done. Just iterate over the combinations you care about, they'll all work.

As soon as you or a library touches cgo, you have to deal with finding and setting up C cross-compilation tooling for your target(s), dynamic linking details, tons of stuff that may have nothing to do with your code or be an area you're an expert in as a Go developer.

[irusensei]

Golang works on Plan9. It can even bootstrap itself. A few months ago I was trying to setup some server software on 9Front for giggles and while most stuff worked I couldn't past the Sqlite CGO dependencies.

[ncruces]

If you still have that itch to scratch, you can try: https://github.com/ncruces/go-sqlite3

You'll need to use the sqlite3_nolock build tag; concurrent writes will quickly corrupt your database. SetMaxOpenConns(1) is your friend.

But it should work. I'm interested if it doesn't. Feedback appreciated.

[llimllib]

Very neat! Any idea how its performance compares to the modernc port?

[ncruces]

It's slower. But wazero is developing an optimizing compiler for amd64/arm64 (the current one is very naive) which I hope will close the gap on those platforms.

[ComputerGuru]

But compiling the non-go assistant process is not going to be any easier than cgo, right?

[cvilsmeier]

> But compiling the non-go assistant process is not going to be any easier than cgo, right?

Right. But you have to do it only once, or you can download a pre-built library from https://github.com/cvilsmeier/sqinn/releases (windows/amd64 or linux/amd64)

[theamk]

I don't think so.

The main go compilation must work in all sorts of environments: dev computers, CI runners.. it should be quick and automated to keep development fast. It should be easy, so everyone on your team can do this.

The assistant process is basically built once and then never changes, you just need to keep a binary somewhere (and they seem to be <1MB so you can check them into git directly). So a single person somewhere has to figure how to do a C build once, and everyone else can benefit. Have your someone ssh into CI runner directly and install gcc. Spend a day installing compiler and messing with Makefiles on exotic OS. You only do it once and you are good forever (or until you want to bump sqlite version)

[parhamn]

Well its a different access pattern. As the underlying library points out:

> It is used in programming environments that do not allow calling C API functions.

Also I guess which one is easier will be subjective. The steps are sorta similar:

Step 1) install sqlite or squinn on the base system (the latter might be harder)

Step 2) if sqllite use cgo, if squinn just use go (the former might be harder but more performant)

[ComputerGuru]

Which programming environments that do not allow calling C API functions also let you build/ship arbitrary C executables, though? (Genuinely curious what scenarios this unlocks.)

[cvilsmeier]

I use it for SQLite Database access in Go and Java. Java lets you theoretically interface with C code, but it's a lot of JNI/DLL/SO work. It's much easier for me to just os/exec (or Runtime/exec in Java) and send data back and forth. Your mileage may vary, of course.

[codetrotter]

In the mentioned https://gitlab.com/cznic/sqlite there would not be any assistant process, right?

[jerf]

No, but that has the disadvantage of being C compiled into Go, then being compiled into native executable.

I'm actually surprised by how readable this came out; props to the Go->C compiler author. But you can guess that pushing this sort of thing through the Go compiler is going to cause some slowdowns due to sheer paradigm mismatch: https://gitlab.com/cznic/sqlite/-/blob/master/lib/sqlite_lin...

[throwaway894345]

I don’t think the paradigm is particularly mismatched, right? If you translate C to Go, it would be pretty much best case for Go (neither language likes lots of small allocations). But Go lacks a lot of optimizations that most C compilers support, like return value optimization, aggressive inlining, etc. C also has lighter-weight function calls and so on that you pay for in Go.

Maybe this is what you mean by paradigm mismatch, but usually I would think of something like translating an allocation-heavy Java app into Go as paradigm mismatch.

[infogulch]

mattn's sqlite3 is probably the most ideal use case for cgo imaginable. But it can still be annoying to set up cgo to build across many platforms.

I think Go should just pull a Zig and just embed a full blown C compiler into go build.

[kunley]

Yeah, once I was on windows and couldn't get cgo sqlite working on a variety of mingw and alike compilers... felt like 90s.

Using this cznic's pure-Go sqlite saved the day.

[oefrha]

Pulling a zig isn't going to solve all problems, as the zig cross compiler itself runs into problems fairly often. I already use CC='zig cc -target x86_64-linux-musl' (or whatever the target is) with cgo to cross-compile mattn/go-sqlite3, and relatively recently it simply stopped working[1] and a workaround had to do be added to about every single project of mine using SQLite through cgo.

I also once tried to figure out a way to cross compile mattn/go-sqlite3 to 32bit Windows with zig, and failed.

The best way to make cross compiling painless is to not use cgo at all. Which is why I use modernc/sqlite whenever possible now.

Btw, bundling a C compiler is also much harder when you don't build on top of LLVM.

[1] https://github.com/mattn/go-sqlite3/issues/1164

[hnlmorg]

I’m going to add a counter argument to all the cgo views raised:

Personally I’ve had much more portability problems with a lot of the native Go ports of sqlite than I have with mattn’s cgo library.

I author a shell that targets most of the architectures and platforms supported by Go. At the request of some users, I added support for a Go native library because they didn’t want to install a C compilers as well as a Go compiler. I tried a few different sqlite ports (though off hand cannot recall which ones) and they all had massive limitations, like only compiling on Windows and Linux (in one example).

In the end, I gave up and reverted back to the cgo version with the option for other libraries hidden behind a compiler flag.

I found my build pipeline manages just fine with cross compiling and haven’t had any complaints (thus far) with the binaries bar one individual running an ancient version of CentOS.

Maybe I’ve been trying the wrong sqlite ports. But here lies the problem: with cgo I know I’m getting a stable, tested, library. With other ports it’s entirely a lottery with regards to how well maintained and tested it might be. For personal projects that’s a fine risk to take but for any larger open source (or even commercial) projects, that additional uncertainty is a risk that distracts me and the other contributors from working on the core part of the project. And thus defeats the connivance of using 3rd party libraries.

[ncruces]

Can you say more, like name your project?

I'm building https://github.com/ncruces/go-sqlite3 and am a community maintainer of https://wazero.io

I can cross compile SQLite into all platforms that Go OOB compiles too, with the caveat that any that aren't linux/windows/darwin/freebsd/illumos (CPU architecture doesn't matter) need a build flag because of file locking: https://github.com/ncruces/go-sqlite3/blob/main/vfs/README.m...

Anything that helps me test portability, or any feedback you might have on it, would be greatly appreciated.

[hnlmorg]

Sorry, just seen this. It is https://murex.rocks

sqlite3 only needs to run from one thread in my use case. Which might explain why I have fewer problems than most.

[hnlmorg]

Just reread your comment and realised I was confusing your library with mattm’s library of the same name.

[scosman]

Exactly. Good to know I didn’t miss a trade off in the mix.

If I’m choosing to use a C framework (SQLite) I’m okay signing up for the environment costs. Prefer that over abstractions in an intermediate layer that might not be maintained in a few years.

[scosman]

Just for the record, had to compile sqlite in gomobile literally the day after this comment and it was a big pain . But got it working and sticking with this approach.

[pram]

Didn't it? Isn't that what 5c, 6c, and 8c were.

[cvilsmeier]

Sqinn author here. Nothing against CGO, but I develop/deploy on Win/Linux, and cross-compiling CGO is very painful. Regarding performance: To my own surprise, Sqinn out-performs mattn (and others) for normal workloads, see https://github.com/cvilsmeier/sqinn-go-bench

[eatonphil]

I think it's a somewhat unfair (though who cares if it's unfair) comparison because you aren't using the database/sql interface and mattn does.

If you drop that interface, you get much better performance.

See: https://github.com/eatonphil/gosqlite for example.

Edit: Nevermind, you did include crawshaw (which doesn't use database/sql) in your benchmark!

[llimllib]

the other pain point I know of is that it's hard to cross-compile. You can do it with zig[1], but it's still not pleasant.

1: https://zig.news/kristoff/building-sqlite-with-cgo-for-every...

[tptacek]

Or with musl-cross:

https://github.com/FiloSottile/homebrew-musl-cross

It works pretty well! It's a thing you might keep in your back pocket to test builds from your ARM dev machine on a dev host, and then let the CI/CD system build the real version later.

[arp242]

And Zig only covers a pretty limited range of platforms to start with.