kernels and composition, patterns

[nytlabs]

kernels and composition, patterns

[mikedewar]

so does a user have two interactions available to them: group and compose? Or does the group operation compose those bits that can be composed and group what remains?

[nikhan]

I think they are separate actions. I am not a fan of it, but the actions are very different. I am not sure about the idea of hiding what is composed and what is not.

[mboggie]

Butting in...

I assume there's a performance benefit to composition vs grouping? The functional differences are small (and mostly, removals [no inspection]) between the two, so the strongest argument I could see for building composites would be a significant performance benefit.

There may also be an argument for composition as simplification, but finding the line between simplification and obfuscation is tricky.

On Nov 30, 2014, at 10:22, nik hanselmann [email protected] wrote:

I think they are separate actions. I am not a fan of it, but the actions are very different. I am not sure about the idea of hiding what is composed and what is not.

—
Reply to this email directly or view it on GitHub.

[nikhan]

The main functional differences are small, but possibly really important.

Consider a group of blocks who have the job to get a value from a store, add 1 to that value, and set the value back in the store. In the situation of doing this with a DB, or an in-memory store, the requirement of having that operation be atomic is really important -- you don't want some other operation swooping in and changing the value while that operation is running. By composing the multiple kernels into one block, all those operations become fused into a single operation -- allowing control over what is atomic and what is not.

The hope is that there is also a performance benefit, but as of right now, the sketch that I put in /thinking/kernels is ~2500 ns/op, whereas the same operation in Go w/o all the kernel abstraction is 130 ns/op. Adding two random floats in JS in Safari takes 34 ns/op. So... there is a little ways to go on that 🚀

I agree about the tenuous line between simplification and obfuscation -- the hope is that this is needed for expressiveness. I don't like that there are two ways to group things and that they behave very differently.

[nikhan]

• A block can only connect one resource
• A block can be composed of multiple kernels.
• A block must satisfy all inputs.

⬛ ⬛ ⬛ ⬛ ⬛

[nikhan]

⭐