hlship
commented
9 years ago As a note, when running ab, I don't see any signs of failures:
Server Software: Jetty(9.2.6.v20141205)
Server Hostname: localhost
Server Port: 8080
Document Path: /auth/users
Document Length: 1271 bytes
Concurrency Level: 20
Time taken for tests: 10.003 seconds
Complete requests: 157
Failed requests: 0
Keep-Alive requests: 157
Total transferred: 239268 bytes
HTML transferred: 199547 bytes
Requests per second: 15.70 [#/sec] (mean)
Time per request: 1274.274 [ms] (mean)
Time per request: 63.714 [ms] (mean, across all concurrent requests)
Transfer rate: 23.36 [Kbytes/sec] received
Connection Times (ms)
min mean[+/-sd] median max
Connect: 0 0 0.2 0 1
Processing: 891 1208 213.2 1150 2089
Waiting: 891 1208 213.2 1150 2089
Total: 891 1208 213.3 1150 2089
Percentage of the requests served within a certain time (ms)
50% 1149
66% 1239
75% 1278
80% 1316
90% 1428
95% 1666
98% 1952
99% 2039
100% 2089 (longest request)The output would be different if the server dropped any connections, or returned any non-2xx responses.
mpenet
niwinz
My handler returns a core.async channel containing a Ring response map.
I've found that if my client terminates early, I start seeing exceptions in my console:
I discovered this using Apache bench (the ab command), using the -t option. I would guess that after it hits its limit, it just closes the HTTP connection(s).
The UUID (f80d055a-7dc1-4823-899e-b2eee7ae8e73) is assigned to each new request.
In fact, using ab -c 20, I see 19 of these exceptions in my log; I see a burst of 19 of these exceptions, a pause of about 20 seconds, then another burst. The UUIDs (assigned by my code as the request is received by my Ring handler) do not repeat, this means that something inside Jet and/or Jetty is retrying them.
I'm tracking down a bit more about how this happens, but it really looks like it would be nice if this exception could be caught and either ignore, or perhaps written to a channel associated with the original request (i.e., so that I could write asynchronous code that would know how to recover from the response being non-deliverable).