[client][performance] Regression after #446

[bsless]

We hit an edge case at work where

keepalives.remove(job.addr.toString() + job.host);

hogged the client's event loop and caused a server to fall over with relatively light load.
Looks like this change is related to #446

The root of the problem is in how PersistentConn implements equals, which is by:

return (addr.toString() + host).equals(obj) || key.equals(obj);

So for each object that checks for equality against a PersistentConn, we do a not insignificant amount of string allocation. Address and host can be compared directly. Same for key.

Possible fix:
edited
Disregard previous suggestion, this can be fixed by implementing equals differently for PersistentConn
The priority queue uses an object's equals. In our case the object can be one of three - PersistentConn, SelectionKey, Request

We can check by instanceof and dispatch to a specialized equality implementation for each case.

wdyt?

[bsless]

See draft PR for potential solution

[ptaoussanis]

@bsless Hi Ben, thanks for pinging about this - and apologies for the regression.

To help make this issue easier for me and others to parse (and therefore review or comment on), can you please provide some more context?

Define a class which will hold addr and job and defines equals when they're both equal

For example here- could you please describe what you believe the problem actually is, and why you believe this approach would solve the problem?

Thanks! 🙏

[bsless]

@ptaoussanis I realized this can be solved by just implementing equals dynamically, edited original issue and the PR contains the proposed solution