Comparisons: Add Java

[joshtriplett]

Still incredibly popular for some applications, including large-scale servers. Need to get some experts in this area to help make the comparison.

[davidbarsky]

Some brief, unedited thoughts as a person who works with large Java deployments/servers:

• Java's garbage collector, as nice as it is, introduces all sorts of unpredictable memory behavior that manifest themselves in highly variable latencies. Even a single latency spike in a service, across a 20-service hop can impact a surprisingly large number of users.
• A big win of using a language like Rust is far more predicable memory usage, and therefore, latencies. As a colleague put it:

  I think the issue is java gets you within spitting distance of close to "optimal" performance but then leaves you high and dry when you try and cross the finish line. That and every instinct people seem to have about writing code in java is wrong for perf at that level, and the perf isn't reliable across the percentiles.

• Rust's pitch could be “memory usage is less variable and makes it easier to avoid making silly mistakes that hurt you in production.“
• More broadly, Java's GC (to the best of my recollection—this may change in Java 11) scales with heap size, not live set size. After a certain point, the number of GC pauses spike.
• No-pause garbage collection (or even the absence of a garbage collector!) would turn latency problems into throughput problems, which can be solved by “throw more/larger machines“ at the problem.

(Sorry for not editing this better, figured that some experience report is better than none.)

[scottmcm]

To pile onto what @davidbarsky said, I've seen the same doing services in C#: the GC is the most complex, least safe, and poorest-understood part of our services. And we hit a bug in the GC a while ago, which was an incredibly awkward thing to track down and work around.