Introspection query normalization via apollo-rs

[o0Ignition0o]

Introspection query normalization via apollo-rs

[Geal]

could we parse the query with apollo-rs and hash the result?

[abernix]

For caches right now I believe we just use a hash (SHA-256, perhaps?) of the string of the raw operation sent by the client. (I also don't know that variations in ignored tokens — e.g., spaces, etc. — are particularly prevalent in the wild right now or something we need to optimize for.

[o0Ignition0o]

We might wanna double check with @lrlna, if I got their presentation right apollo rs keeps all of the tokens around, so I'm not sure this would suffice.

OTOH I cannot find any normalization step happening on queries in federation, (poke @abernix maybe you know where/if it happens ?)

I think we can start by hashing the parsed result, and we can then iterate on it and normalize the query before we parse it via apollo-rs!

[lrlna]

@o0Ignition0o yes, the AST is lossless: i.e. we keep all the whitespace and other ignored tokens. Eventually, once we have semantic analysis in place, you should be able to store the schema object which would mark ignored tokens as insignificant, therefore two queries that only differ in whitespace would be identical. This is, however, a few months out.

On the other hand, if a query has changed, even if it's just whitespace, how do you currently find out if that's the only thing that has changed, therefore is syntactically identical to what's already in a cache? Don't you have to do a brand new parse anyways?

[o0Ignition0o]

I suspect we would have to indeed :(

(I just found the conversation we had on proj-router about that, and it seems to be what stripIgnoredCharacters does)

If a couple of \n and whitespaces strip is good enough we might be able to do that, otherwise i'm not sure that's too useful for our caching usecase

[Geal]

On the other hand, if a query has changed, even if it's just whitespace, how do you currently find out if that's the only thing that has changed, therefore is syntactically identical to what's already in a cache? Don't you have to do a brand new parse anyways?

My (untested) assumption is that query planning and introspection could cost a lot more than the parsing and normalization step

[abernix]

(Of course, we can eventually and there are a variety of reasons to do so, particularly as the weight of what we cache increases!)

[abernix]

(Continuing my parentheticals, rather than just editing the previous post): I should also note that the current cache store for these things are all bound LRU caches.

[abernix]

Closing as I know our more-practical nera-term answer this is going be "do introspection in Rust", which is in-progress thanks to @SimonSapin in apollographql/apollo-rs#758.