ruby-grape / grape / 33379390620

97%

master: 97%

Pull #2872

github

Include endpoint helpers once instead of on every request

Helpers are instance methods on the endpoint, because the route block is bound
to the endpoint and runs against it. They are also per-endpoint — including
them into Grape::Endpoint itself would leak one API's helpers into every
endpoint in the process — so they went onto the per-request copy:

    def call(env)
      dup.call!(env)
    end

    def call!(env)
      ...
      singleton_class.include(@helpers) if @helpers

That had to happen per request rather than once: Object#dup deliberately does
not carry a singleton class over, so including the helpers on the shared
endpoint would never reach the copies that actually serve requests.

The cost is not the include. `dup` plus `include` measures ~510 ns against
~110 ns for `dup` alone, which would be under 2% of a request. The expense is
that `singleton_class` on a fresh copy *creates a class*, and a class nothing
has called through yet has no method cache. Counting allocations: the dup, its
singleton class and the iclasses are 4 objects, and the first call to each
distinct method through that class costs 3 more — paid again on the next
request, for every helper, for the route block, and for every DSL method the
endpoint answers.

Include them once, at compile! time, into a subclass owned by this endpoint,
and copy the endpoint's state onto an instance of it. Each request then copies
an object whose class has been warm since boot, and request handling no longer
mutates a class at all.

On a small API with helpers, params and validations (median of 9 runs, Ruby
4.0.5, single-threaded Benchmark.ips):

                          YJIT                   no-YJIT          objects/req
    GET    40,853 -> 51,605  (+26%)   19,574 -> 21,368  (+9%)      154 -> 121
    POST   40,045 -> 51,009  (+27%)   19,657 -> 21,711  (+10%)     157 -> 123

An endpoint that declares no helpers is its own prototype: throughput and
allocation counts there are unchanged.

On... (continued)

1239 of 1330 branches covered (93.16%)

Branch coverage included in aggregate %.

10 of 10 new or added lines in 1 file covered. (100.0%)

3905 of 3973 relevant lines covered (98.29%)

27143.07 hits per line