Add endpoints for memory: absolute and relative

[dcorbacho]

Add endpoints for memory: absolute and relative
Solves #161

@michaelklishin This PR is missing the tests for relative memory, as I need confirmation of:

1. Is it fine to report relative memory as an integer value?
2. If memory usage is very low, but not 0, should it be reported as 0 or 1? Meaning that the component is indeed doing something.

[michaelklishin]

@dcorbacho I'd say integer values (as %) is good enough. Reporting a value as 0 would indeed be weird, so I'm OK with using 1% for those cases.

[michaelklishin]

The relative endpoint returns a total value that seems absolute. If other numbers are meant to add up to 100, I'd say we should return 100 for total:

{
    "memory": {
        "atom": 1,
        "binary": 1,
        "code": 24,
        "connection_channels": 0,
        "connection_other": 1,
        "connection_readers": 0,
        "connection_writers": 0,
        "mgmt_db": 1,
        "mnesia": 1,
        "msg_index": 1,
        "other_ets": 2,
        "other_proc": 16,
        "other_system": 54,
        "plugins": 3,
        "queue_procs": 1,
        "queue_slave_procs": 0,
        "total": 100594608
    }
}

@dcorbacho WDYT?

[dcorbacho]

I think that you are right. I'll fix it now

[gerhard]

We've just come across this part of rabbitmq/rabbitmq-server#1259 and noticed 2 things:

1. Relative memory usage adds up to 106% in a just started RabbitMQ node and 115% in one of the 3.6 LRE nodes. Would it be worth returning fractions with single precision rather than rounded integers? Single precision rounding will skew the total by at most 1.7% (17 values * 0.1)
2. The /api/nodes/NODE/memory API endpoint is not documented in /api. Should it be?

[michaelklishin]

@gerhard we can switch to floats with, say, 2 decimal places. @dcorbacho @kjnilsson WDYT?

[gerhard]

That would be perfect. Can this ship in 3.6 or is it a breaking change?

[kjnilsson]

2 decimal places sounds reasonable.

[michaelklishin]

@gerhard most JSON parsers should parse integers much like floats. Computations on floats can work differently depending on the language. I think it's acceptable given other changes around VM memory consumption calculations.