ruby-grape / grape / 35499232060

99%

master: 99%

Pull #2972

github

Stop the request path resolving params it was not asked for

Three places where a request resolved params nobody had asked for, or
resolved the same ones again.

**A nested params scope is resolved from the scope above it.**
`ParamsScope#params` walks back to the root on every call, and
`#should_validate?` asked every scope on the chain for its params in turn
-- so a scope nested four deep resolved ten scopes to answer one question,
and did it again for each of its validators. The cost grows with the square
of the nesting depth. A scope that declares no dependency and has none
above it can only be asked whether its own params were given: `given` is
the one thing that makes a scope's answer depend on anything else, through
the qualifying params it stores per request, and such a chain stores none.
So the chain is resolved once from the root down, each scope from the one
above it, and checked from there -- every check is pure, so the order it
runs in does not change the answer. `#params` skips the qualifying-params
lookup and the `presence` test that follows it on the same grounds.
`#all_element_blank?` is only ever asked about params that are not blank,
and a Hash that is not blank is never all blank: what `all?` yields for it
is a [key, value] pair, an Array of two, which `blank?` is false for.

**The query string is parsed for a `format` param only when it could name
one.** `?format=json` overrides the format an API pins, so the formatter
read it out of the query params -- parsing the whole string on every
request that had one, before the endpoint had a say. Parsing is the most
expensive thing such a request does, about 1.2 µs per pair, and an endpoint
that never reads its params paid all of it for one lookup that could not
have hit. Rack splits the query string on `&` and `=` and then unescapes
each key, and unescaping only ever rewrites `%XX` (or a `+`, which becomes
a space), so a string that spells neither `format` nor a percent-escape
cannot hold... (continued)

1464 of 1521 branches covered (96.25%)

Branch coverage included in aggregate %.

26 of 26 new or added lines in 4 files covered. (100.0%)

4325 of 4325 relevant lines covered (100.0%)

28771.75 hits per line