feat(cbor): add Smithy RPCv2 CBOR protocol generator

[kuhe]

Adds a protocol generator for the trait smithy.protocols#rpcv2Cbor

[kuhe]

jest has problems passing the JSON data between workers when it includes BigInt's, apparently?

[trivikr]

LGTM. I didn't go through CBOR encode/decode source code, as it has unit tests.

[mtdowling]

Would it help to store the result of typeof into variable and reuse that rather than this many if/else statements that call it over and over? I also see typeof input === "string" above too.

I also think V8 would optimize a switch statement into jump statements. I doubt it would do that for these if/else statements.

[kuhe]

This testing shows if-statements to be faster https://jsperf.app/kuyaba/2.

typeof variable === "..." is likely optimized in a specific manner. Even storing typeof variable instead of calling it multiple times de-optimizes the code. The typeof operator is likely a property on the native object and not a function call.

[mtdowling]

Very weird!

[mtdowling]

Is this some kind of global buffer that all cbor encoding writes to? Forgive my JS ignorance, but is that dangerous if someone uses something like workers? Is this does use global write state, I'm curious why we wouldn't use an isolated state for each encoding process and do something like buffer pooling to borrow/return up to N already allocated buffers.

[kuhe]

Yes, and the global encoding buffer is inaccessible to consumers, so it cannot be passed to workers and won't be subject to concurrent operations. Workers will each have a copy of that buffer.

Because the encoding function is synchronous and the runtime single-threaded, the buffer is in use by only one encoding process at a time. A copy is made when the result is handed over to the outgoing request. This copy is done very quickly and represents <1% of the time spent.

I believe using a pool of buffers would not improve the speed or memory usage, even if it could avoid that copy operation, due to the speed of that copy operation and the overhead of managing the buffer pool.

[mtdowling]

is there any way to use something like buffer pooling?

[kuhe]

I don't know of a specific Node.js API that maps to the term "buffer pooling". If you mean using for example creating a manager class that allocates and reclaims separate buffers, I don't believe it will improve the perf profile over using a single buffer, for reasons given below in the comment thread about the global buffer.

Initial development was done with individual buffers per encoding process, and iterative development showed that allocation/reclamation was costly and would unpredictably trigger GC. Even a well managed pool shouldn't beat the simplicity of a single buffer.

Others will review