Performance

Background

Muuntaja has been built with performance in mind, while still doing mostly everything in Clojure:

• single middleware/interceptor for all formats (instead of stacked middleware)
• avoid run-time regexps
• avoid dynamic bindings
• avoid Clojure (map) destructuring
• (Java-backed) memoized content negotiation
• Protocols over Multimethods
• Records over Maps
• use field access instead of lookups
• keep all core functions small enough to enable JVM Inlining
• unroll generic functions (like get-in)
• use streaming when possible

The codebase contains the performance test, done using Criterium under the perf Leiningen profile with a 2013 Mackbook pro.

NOTE: Tests are not scientific proof and may contain errors. If you have idea how to test things better, please poke us.

Middleware

Muuntaja is tested against the current common ring-based formatters. It's fastest in all tests.

[ring/ring-json "0.4.0"] & [ring-transit "0.1.6"]

• ok performance by default, but only provide a single format - Stacking these separate middleware makes the pipeline slower.

[ring-middleware-format "0.7.0"]

• has really bad defaults:
  • with 1K JSON, Muuntaja is 10-30x faster (depending on the JSON encoder used)
  • with 100k JSON, Muuntaja is still 2-4x faster
• with tuned r-m-f options
  • with <1K messages, Muuntaja is still much faster
  • similar perf on large messages

JSON

Transit

Interceptors

Pedestal:

• io.pedestal.http.content-negotiation/negotiate-content for content negotiation
• io.pedestal.http.body-params/body-params for decoding the request body
• io.pedestal.http/json-body, io.pedestal.http/transit-json-body etc. to encode responses

Muuntaja:

• muuntaja.interceptor/format-negotiate for content negotiation
• muuntaja.interceptor/format-request for decoding request body
• muuntaja.interceptor/format-response to encode responses
• muuntaja.interceptor/format all on one step