Clojurians Log v2

Interesting project! I was surprised to see jetty score so low and http-kit so well on the benchmarks, do you know if anyone has done more comprehensive benchmarking of clojure http servers?

Love the name 😄

http-kit itself does quite a few benchmarks, but between Ring Jetty and itself. I don't believe it benchmarks other Ring adapters. There were also a couple of Ring adapters I could still add, but they don't have releases on Clojars which makes it more difficult. While developing Capra I benchmarked frequently against http-kit, and was surprised by just how quick it is. It does take some shortcuts that Capra doesn't; but overall its a very efficient codebase.

We've used http-kit for years, it's pretty good. But I'm very keen to follow this project 🙌

The problem with jetty is log level, default debug, they're all rendered but not emitted, takes about 95% off your throughput Found by profiling because I wondered why jetty was slow

https://clojurians.slack.com/archives/C03S1KBA2/p1691305935346429?thread_ts=1691305650.647229&cid=C03S1KBA2

https://clojurians.slack.com/archives/C03L9H1FBM4/p1681247812313709?thread_ts=1681223084.938609&cid=C03L9H1FBM4

I'll try that and see if it affects Jetty's performance in the benchmarks. Thanks for your insight, Ben.

that is a MVP level top tip right there, thanks @ben.sless. I've always stuck with jetty cause it's just so rock solid.

why did you want to avoid proxy?

I do think it's a bit irresponsible of jetty to set this log level by default on the assumption users will configure the entire mess of Java logging

I've added a logback.xml file to the base src directory:

capra.benchmark=> (require '[ :as io])
nil
capra.benchmark=> (slurp (io/resource "logback.xml"))
"<root level='ERROR'></root>\n"

I'm not sure this is the complete xml

At work, we use Jetty because New Relic supports it, but does not support http-kit. We tried http-kit but lost a lot of valuable information for monitoring so we switched back. I suspect the same would apply to Capra, since New Relic wouldn't know how to instrument the code to enable observability.

See https://logback.qos.ch/manual/configuration.html

(and I see another announcement thread has turned to performance! 🙂 )

Everyone blaming the nerd instead of the sniper

Point taken though

I spent a lot of time testing performance of Capra, so this thread feels appropriate.

I was just echoing Rich's frustration with some of the recent threads. Ben knows I'm just teasing him.

I tried this logback.xml next:

<?xml version="1.0" encoding="UTF-8"?>
<configuration>
  <appender name="STDOUT" class="ch.qos.logback.core.ConsoleAppender">
    <encoder>
      <pattern>%d{HH:mm:ss.SSS} [%thread] %-5level %logger{36} - %msg%n</pattern>
    </encoder>
  </appender>
  <root level="error">
    <appender-ref ref="STDOUT" />
  </root>
</configuration>

Regarding the Jetty benchmarks, I didn't find them too surprising. A lot of adapters based on Java web servers hover around the 30K req/s mark, and the Ring Jetty adapter is somewhat less efficient because it passes through Ring to Servlets to Jetty. rj9a also uses Jetty 12, but is faster because it doesn't use Jetty's Servlet compatibility layer.

Though of course, there's other reasons for using Jetty, as @seancorfield pointed out. And once you throw in a handler that actually does work, the difference between adapters is a lot smaller.

I think it's very appealing to have more pure Clojure alternatives, so I applaud the effort.

lol hey @weavejester why did you want to avoid proxy?

It's slower than reify. So I use reify to create a custom IInputStream instance, and then a custom class ProxyInputStream to turn it into a full InputStream.

coolio, thanks

Managed to find the logback.xml from the original thread

<?xml version='1.0' encoding='UTF-8'?>
<configuration scanPeriod='5 seconds' scan='true'>
    <appender class='ch.qos.logback.core.ConsoleAppender' name='STDOUT'>
        <filter class='ch.qos.logback.classic.filter.ThresholdFilter'>
            <level>
                ERROR
            </level>
        </filter>
        <encoder>
            <pattern>
                %date %highlight(%-5level) %logger{36}: %msg%n
            </pattern>
            <charset>
                UTF-8
            </charset>
        </encoder>
    </appender>
    <root level='ERROR'>
    </root>
    <logger name='hello'>
        <appender-ref ref='STDOUT'>
        </appender-ref>
    </logger>
    <logger name='user'>
        <appender-ref ref='STDOUT'>
        </appender-ref>
    </logger>
</configuration>

I put it under src/logback.xml

I added the logback.xml you posted to the classpath, but it hasn't improved the Jetty benchmark significantly. It may be that the logging only becomes a significant factor when there's more simultaneous connections, or perhaps later versions of Jetty fix the default behaviour.

I'll do some investigating and follow up in #performance

The Capra benchmarks are all in the repository, so feel free to use them if that's useful.