Compilation latency dominated by addOverload operations when using extension libraries

[jpbetz]

Originally found by @DangerOnTheRanger and @liggitt

In Kubernetes we use the same set of libraries for all CEL expressions, but have to pay the cost of registering them into an environment each time a CEL expression is compiled. This appears to dominate the compilation cost.

cel-go/cel/env.go

Line 163 in ffa9b84

err = ce.Add(e.declarations...)

[jpbetz]

cc @TristonianJones @DangerOnTheRanger @liggitt

[TristonianJones]

@jpbetz This doesn't surprise me. Why can't you use a statically configured environment?

[jpbetz]

Why can't you use a statically configured environment?

Maybe we can? We vary the variable decls used across the compiled CEL expressions, which is what I think got us hung up here.

I'll have a look at #347 PR that you shared offline with me and see if that can be use to eliminate the bulk of the cost.

[TristonianJones]

@jpbetz Take a look at the draft state for #347. I've added some options to early validate declarations and to ensure that validated declarations are copied from one Env to the next when the Env.Extend() call is used.

This means that if you do an init() of the base environment for K8s you should see a 7-10x performance boost and a dramatic reduction in memory allocations:

func init() {
        // maybe panic if this errors
        baseEnv, _ = cel.NewEnv(
		cel.HomogeneousAggregateLiterals(),
                 library.ExtensionLibs...,
                 cel.EarlyValidateDeclarations(true),
	)
}

Then use the baseEnv.Extend with the per-compile options you need in your use case.

Before PR:

cpu: Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz
BenchmarkNewEnv-8   	     430	   2674244 ns/op	  574888 B/op	   20228 allocs/op

After PR:

cpu: Intel(R) Core(TM) i7-6700HQ CPU @ 2.60GHz
BenchmarkNewEnv-8             	     994	   1064345 ns/op	  369540 B/op	   12238 allocs/op
BenchmarkEnvExtend-8          	    7965	    154726 ns/op	   42053 B/op	     661 allocs/op
BenchmarkEnvExtendNoDecls-8   	    9135	    114062 ns/op	   34511 B/op	     401 allocs/op

What do you think? If you think this will work, I'll finish up the tests tomorrow and we'll do another release. I'm actually tempted to make a base copy of the stdEnv internal to CEL to speed up the simple extension cases as well.

[jpbetz]

Fixed by #347

See kubernetes/kubernetes#108954 for benchmarks.