2.0.0-alpha* version out of memory

[chenjianye]

Library Version(s)

2.0.0-alpha2 ~ 2.0.0-alpha17

Problem

The server use Clojure 1.9 and spec, set JVM -Xmx=384m. Post data to server repeatedly, and then server terminate by "OutOfMemoryError" later.

And version 1.1.11 and 2.0.0-alpha1 have not this error.

[ikitommi]

Hi. Could you provide a minimalistic app that fails? Thanks.

[ikitommi]

haven't seen this, could you provide some failing sample?

[chenjianye]

I am so sorry that I was on vacation and didn't see the mail in time.
The example is here：https://github.com/chenjianye/Compojure-api-error-example

[asheldo]

The heap dump shows hundreds of thousands of spec_tools.Spec's -- but this is on 0.5.1 of spec-tools. It won't run under 0.6.1.

[ikitommi]

Thanks. The spec-memoization is clearly not working. On vacation now, can fix that easily when back on computer. To shield against regression, do you know a way to create a unit/system test that would tell if the number of Specs grows in use? I don't.

[asheldo]

Interesting. Just getting back into clj thinking, so maybe -- junit mock fmwk's would in Java space assert on "times(1)" the method calls.

[ikitommi]

back.

memoization fails as the (s/coll-of map?) returns different instance each time and misses the cache. Still, the cache should have been a fifo of 10000 entries, so it shouldn't cause OutOfMemoryError. I dropped the cache max size to 1000 anyway as starters, but not a real fix, in alpha-19.

I think a real fix would be to replace the Coercion implementation to the one from reitit - which is a polished version of the c-api original. With it, the coercers are created once per endpoint, and no memoization is needed. Currently, in c-api, it's resolved at request-time, memoization to keep it fast.

Could you test that the number of Specs floating doesn't go much over 1000 to verify that the fifo works as expected?

[asheldo]

Will do, and let you know -- thank you.