[API-537] Add support for the binary metrics

[mdumandag]

Implemented the mechanism to compress the metrics we collect in a
way that is compatible with the server-side.

With this change, the latest Management Center versions are able to
show statistics regarding the Python client.

For backward compatibility, we also continue sending metrics in
a comma separated key-value format (client attributes).

In this PR, I also significantly cleaned up the statistics service.
In the old code, we were trying to reuse the results of some
calls. In the benchmarks I did, I saw that it didn't really improve
much. (New approach approx. 0.5 ms, old approach approx. 0.4 ms per
statistics collection) This really small difference does not justify
the need for complexity so I went with the simpler approach.

Also, aligned the log levels with the Java client for the various
warnings.

Also, bumped the version of psutil to 5.8.0.

[mdumandag]

Here is the benchmark I used to compare the old vs new approach

import timeit
import psutil


def f1():
    psutil.virtual_memory().total
    psutil.virtual_memory().free
    psutil.virtual_memory().used
    psutil.swap_memory().total
    psutil.swap_memory().free
    psutil.cpu_percent()
    psutil.cpu_count()
    p = psutil.Process()
    p.memory_info().rss
    p.num_fds()
    p.rlimit(psutil.RLIMIT_NOFILE)[1]
    p.cpu_times()
    p.create_time()


def f2():
    v = psutil.virtual_memory()
    v.total
    v.free
    v.used
    s = psutil.swap_memory()
    s.total
    s.free
    psutil.cpu_percent()
    psutil.cpu_count()
    p = psutil.Process()
    with p.oneshot():
        p.memory_info().rss
        p.num_fds()
        p.rlimit(psutil.RLIMIT_NOFILE)[1]
        p.cpu_times()
        p.create_time()


print(timeit.timeit(f1, number=10000))
print(timeit.timeit(f2, number=10000))

[srknzl]

Looking good left some minor comments

[srknzl]

Before merging, can you test that if metrics can be sent to management center on windows(python client runs on windows)? some of them like number of file descriptors may not work.

[mdumandag]

Yes, I have verified this, on Windows, this feature works with a subset of the stats we sent from Linux.

For example, file descriptor count is not sent on Windows, as there is no such thing in Windows, or the load average will not be sent, as it is not implemented there in the Python standard library.

[codecov-commenter]

Codecov Report

Merging #441 (7ea0d8c) into master (7d796f2) will decrease coverage by 0.66%.
The diff coverage is 96.37%.

❗ Current head 7ea0d8c differs from pull request most recent head ddfc21d. Consider uploading reports for the commit ddfc21d to get more accurate results

@@            Coverage Diff             @@
##           master     #441      +/-   ##
==========================================
- Coverage   86.62%   85.95%   -0.67%     
==========================================
  Files         336      340       +4     
  Lines       17398    17850     +452     
==========================================
+ Hits        15071    15343     +272     
- Misses       2327     2507     +180     

Impacted Files	Coverage Δ
hazelcast/statistics.py	41.80% <91.83%> (-26.98%)	⬇️
hazelcast/metrics.py	82.62% <98.87%> (ø)
hazelcast/reactor.py	77.51% <0.00%> (-6.70%)	⬇️
hazelcast/listener.py	68.60% <0.00%> (-1.17%)	⬇️
hazelcast/connection.py	82.73% <0.00%> (-0.19%)	⬇️
hazelcast/proxy/map.py	73.37% <0.00%> (-0.01%)	⬇️
hazelcast/aggregator.py	97.39% <0.00%> (ø)
hazelcast/protocol/codec/map_aggregate_codec.py	100.00% <0.00%> (ø)
... and 1 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 7d796f2...ddfc21d. Read the comment docs.