java 1.7 support about caffeine

[ben-manes]

No, sorry. The library is intended for server-side Java 1.8+. The API and implementation take advantage of features that cannot be easily back ported (e.g. using retrolambda). I'd recommend,

• If you are on Android then prefer LruCache.
• If you want features, then use Guava's cache.
• If performance matters most, then use ConcurrentLinkedHashMap.

[guidomedina]

Hi Ben, sorry for using this issue to ask a quick question:

I have been using Guava and now for small caches I just switched to Caffeine as I don't need Guava features, there are some concurrent maps I'm using for very specific caches which are growing to a finite level not a big number, mostly load almost all at start up and then just read, am I better of with ConcurrentHashMapV8 or ConcurrentLinkedHashMap ?

I mean, most of my caches are say 95%+ times read ops hence I'm using computeIfAbsent when writing and few times computeIfPresent for specialized writes.

[ben-manes]

You're welcome to send me a private email, or create / add onto an issue. As long as you get answers I don't mind what mechanism.

I'm a little confused with your last statement about compute methods, which are Java 8 only. Or do you mean that you'd use ConcurrentHashMapV8 for a back port of that feature to Java 7?

[guidomedina]

Ah, I confused you there when I mentioned the CHMv8, what I meant was that I'm already using Java 8 and AFAIK its CHM is better than older versions (v8) correct?

So what I'm trying to figure out is Java 8 CHM performance vs CLHM and Caffeine counter parts (if any present which I didn't find), for specific expiring/loading caches I'm already using Caffeine cache which I migrated from Guava's cache.

[ben-manes]

Oh okay. Adding this to the Java 7 issue was confusing.

CLHM doesn't handle computations natively, so you would fallback to the looping retrials from the default method. Or you could use a decorator, but more work than you'd prefer.

CLHM is built on CHMv8 backport and Caffeine on CHM, so the raw hash table will always win. I included that as "unbounded" in my benchmark results. Note that CHM is pessimistic about computations, so I pre-screen to avoid unnecessary locking. That has a big benefit for fewer than 32-cores and is less pronounced on larger machines. Caffeine has an unbounded version that is CHM with prescreening, interface adaptions, and optional listener/writer/stats.

I think the answer you're looking for is that CHM is fine to use and really great.