Why tracer:http_headers_inject() instead of tracer:inject() passing type?

[james-callahan]

Other language seem to have a single inject method where you pass in the type. Why do we have different methods per type?

[rnburn]

There was a decision to get rid of the extra format parameter in the inject/extract method and bundle them into the carrier (see discussion on opentracing/opentracing-go#127).

The methods are similar to those in the C API, but I'm open to other designs.

[james-callahan]

So I sort of liked the more abstract inject API: lua has no 'standard' data type for http headers. Some frameworks use a simple table with key->value. Others add metamethods for case normalisation. Other environments use a richer object such as https://daurnimator.github.io/lua-http/0.2/#http.headers
This makes it unclear what should get provided to the http_headers_inject method.

[rnburn]

The carrier passed in just needs to store the key-values; it shouldn't do transformations. It's the responsibility of the tracer implementation to deal with the case sensitivity of header keys and ensure that they're valid.

[james-callahan]

The carrier passed in just needs to store the key-values;

That doesn't answer the issue though: e.g. do they get stored with carrier[key] = value or with carrier:add(key, value). And for an extractor: does it use e.g. carrier[key] or carrier:get(key)

[rnburn]

I would think carrier[key] = value and carrier[key] so that you can use a standard table for the carrier. You can always change the metamethods if you want to do something different.

[daurnimator]

I would think carrier[key] = value and carrier[key] so that you can use a standard table for the carrier. You can always change the metamethods if you want to do something different.

It would be following an anti-pattern to have a dummy table that has __newindex and __index methods instead of passing in a function that is a getter/setter.

[rnburn]

The PHP and JavaScript APIs both use associative arrays for carriers.

A carrier interface more like what's in Go is also a possibility. But given that you can always use an intermediary table to propagate to a custom data structure, is there much of a use case for custom carrier implementations?

[james-callahan]

is there much of a use case for custom carrier implementations?

I think so: e.g. in LuaJIT, iterating over a table (with pairs/next) bails out of the JIT compiler: if the HTTP implementation wants to be fast, then they can't use a table for storing headers.

A different use-case is in WAFs: you want to be able to track order of headers, and take action on multiple occurrences of the same field (sometimes you want to allow it; other times it needs to be blocked). Using a plain dictionary for this structure doesn't work.