These are to sidestep issues on some platforms. Generally certainly fine if
they don't take effect. Caught thanks to tor-relays@ beta tester on FreeBSD...

  nyx: environment corrupt; missing value for
  Traceback (most recent call last):
    File "./run_nyx", line 14, in <module>
      nyx.main()
    File "/usr/home/ryan/nyx/nyx/__init__.py", line 147, in main
      nyx.starter.main()
    File "/usr/home/ryan/nyx/stem/util/conf.py", line 289, in wrapped
      return func(*args, config = config, **kwargs)
    File "/usr/home/ryan/nyx/nyx/starter.py", line 90, in main
      os.putenv('LANG', 'C')  # make subcommands (ps, netstat, etc) provide
  english results
  OSError: [Errno 14] Bad address

Hey, it's only a couple orders of magnitude.

... *sigh*.

Why did we divide by that? Weird. Because of it we always showed tor's cpu
usage as 0%.

We shouldn't silently swallow errors. Opting for notice runlevel so they're
visible to users by default so they can be reported.

The 'width' attribute already takes into account the initial 'x' offset, but in
the conditional we didn't account for that causing connection details to just
about never be shown.

Fixing a couple easy to make gotchas in our --log argument. It's especially
confusing becuase nyx expected 'NYX_WARNING' whereas it displayed 'NYX_WARN'.

Typo Roger spotted in the comments.

Lighten the work needed by the ConsensusTracker by having it process unparsed
consensus content. We already do this when starting up, so this actually
simplifies our code quite a bit.

If the dialog's title is shorter than the content (which happens when we lack
descriptors) the dialog crops the title.

It's rare, but once I spotted a "[0 duplicates hidden]" in my logs. Reason is
that the message deduplicated, but then the log got long enough that the
duplicate was dropped.

Now that stem can cache our address we can drop our our_address() helper.

Calling the endpoint of our own circuits 'exit' worries folks on occasion,
concerned that they might accidently be running an exit relay. Renaming
the termination point to 'end' to hopefully avoid those concerns...

  https://trac.torproject.org/projects/tor/ticket/12956

Replacing the textual listing with images to give the page a friendlier, more
approachable feel.

Rather than directing users toward the nyxrc.sample file providing them with a
webpage describing the options. Also making it more discoverable by putting it
in the 'more' menu.

Lots of breakages have slipped in while attempting to improve our performance.
Most were timing related due to our switch to static sleeps.

Oops! Forgot to add a couple small image resources.

Simple page listing our relay support channels (email list, irc, trac, and me).

When looking at our site I just realized it's hard to figure out how to report
bugs. It's in the FAQ, but who's gonna read that?

Adding a 'more' dropdown menu with useful, but less common sections folks will
want (changelog, bug tracker, and contact info for now). The contact page
doesn't exist yet - that's next.

God the styling on this was a pita. I'm really not sure why this negative
margin is necessary but without it the rest of the nav menu gets pushed down.
Stared for hours at this hoping to get something simpler but time to raise the
flag and commit what works.

I really hate the lack of a shared header, but can't say that alone is worth
something like Hyde. :/

Connections shouldn't use a lru_cache. If the cache size is less than the
number of connections we have at any point then we consume extra cpu, but if
it's too high we consume more and more memory unnecessarily until it's filled.

Instead evicting from the cache when a connection hasn't been referenced in a
while.

Interesting graph panel timing bug when draw's concurrently invoked before
we've set stats_paused. Only seen it once, but simple to fix.

  File "/home/atagar/Desktop/nyx/nyx/curses.py", line 737, in draw
    func(_Subwindow(subwindow_width, subwindow_height, curses_subwindow))
  File "/home/atagar/Desktop/nyx/nyx/panel/graph.py", line 569, in _draw
    stat = self._stats_paused[self._displayed_stat]
  TypeError: 'NoneType' object has no attribute '__getitem__'

Interesting. The lru_cache for from_connection works, but the ones for the
methods don't. This causes truly horrid performance when you have a ton of
connections since it's recalculating the type and lines with every redraw.

When paused daemon updates turned into a busy loop, pegging the cpu. Actually,
I'm unsure why we're checking the paused status at all since daemons still run
in the background.

Replacing stat cloning with a simple lock bought us some performance
improvement but it also introduced a deadlocking issue. If redraw is called
while we're updating a stat the interface freezes. This can be easily
  sidestepped by only locking for the stat update and not including the redraw
  call.

Tor's DisableDebuggerAttachment feature breaks all methods we attempt to get
tor's cwd (/proc, pwdx, and lsof). However, this is fine since we only need the
cwd to expand relative paths which doesn't come up much (iirc tor makes more
effort nowadays to give only absolute paths).

Skipping cwd resolution if a path's relative so we avoid the doomed system
calls.

Stem transparently caches immutable information but our address is a rather
strange critter. It can change but hardly ever does, and is used quite a bit.

Nyx forced Stem to treat this as being cachable but when our address is
unavailable this still accounted for numerous requests. Instead fetching
the our address at a fixed rate of once a minute.

This provides two benefits...

* Avoid pointless requests while there's no network available.
* This way we'll evenually update our address if it changes while running.

When the interpreter panel has content the scollbar doesn't report itself as
being at the bottom.

Threading conditionals are the proper method of doing an interruptable sleep,
so nyx can shut down right away when quit. However, each of these conditionals
impose a surprisingly high tax in terms of baseline cpu usage. Fully half the
cpu when idle is just spent on checking conditionals. Guess they're implemented
with some sort of busy puller...

We don't need interruptability. It's nice, but having our interface take a
fraction of a second is well worth halving our cpu usage. When idle this change
drops cpu usage from 2.1% to 1.1% for me.

  Pause   Cpu
  0.1s    1.4%
  0.2s    1.3%
  0.3s    1.2%
  0.4s    1.1%
  0.5s    1.1%
  0.6s    1.1%

No detectable difference when I lack any connections, but constructing this
lists with every connection entry is silly (even if it is cached). Constructing
a set once instead for constant time lookups.

Turns out having a daemon panel reacquire a lock every 0.2 seconds is fairly
taxing, even if the panel isn't doing anything. That, and having the log panel
every 0.3 seconds is kinda silly. This drops the log panel usage when idle by
80% (0.5% to 0.1% on my system).

There's no point in fetching graph title stats every second. These change on an
hourly basis. Caching this and skipping vline when there's nothing to display
drops our cpu usage when idle by 70% (from 1% to 0.3% on my system).

Unnecessarily providing the 'forced' flag when redrawing can cause screen
flickering, but omitting it causes actions to not be visible. For instance,
when the relay is inactive changing pages takes seconds to appear.

My relay complains with...

  19:13:35 [NYX_INFO] Failed three attempts to determine the process using
  active ports (unrecognized output from lsof (lines are expected to have ten
  fields: COMMAND PID   USER   FD   TYPE  DEVICE SIZE/OFF NODE NAME): COMMAND
  PID   USER   FD   TYPE  DEVICE SIZE/OFF NODE NAME)

This is because there's a single space after 'COMMAND' rather than two, causing
us not to omit the title line from parsing. Making this less strict.

Finally got it! The connection panel's performance relies on heavily reusing
cached resources. However, on my moderately busy relay I have 2000 connections
whereas lur_cache's default size is only 250. This caused us to constantly
evict and recreate resources every five seconds. This change drops the nyx cpu
usage for me from 28% to 16%.

At only INFO runlevel cpu usage on my relay jumps from 13% to 20%. This is
entirely coming from deduplication. We made this far more efficient (linear
from n^2) but it's pretty easy to finish dropping this down to a constant
time operation.

Our graph panel updates and redraws itself every second, making it a
non-trivial load on cpu. Two actions in particular account for most of
it...

  * Cloning stats on every redraw. This constantly consumes 0.3% cpu on my
    relay. Being lock-free is nice, but in this case a lock really makes
    sense.

  * Performing a flurry of addstr calls for each cell in the graph. Doing
    a single vline call for each column instead drops cpu by another 0.4%.

On my relay this drops cpu usage of the graph panel from 1% to 0.4% (60%
less)...

  * 0.3% cpu usage when only the header panel is rendered
  * 1.3% cpu usage when the graph panel was rendered too
  * 0.7% cpu usage when the graph panel was rendered with these changes

If the nyxrc disables all our panels except the header then we crash with...

  Traceback (most recent call last):
    File "/home/ubuntu/nyx/run_nyx", line 14, in <module>
      nyx.main()
    File "/home/ubuntu/nyx/nyx/__init__.py", line 137, in main
      nyx.starter.main()
    File "/home/ubuntu/nyx/stem/util/conf.py", line 289, in wrapped
      return func(*args, config = config, **kwargs)
    File "/home/ubuntu/nyx/nyx/starter.py", line 94, in main
      nyx.curses.start(nyx.draw_loop, acs_support = config.get('acs_support', True), transparent_background = True, cursor = False)
    File "/home/ubuntu/nyx/nyx/curses.py", line 216, in start
      curses.wrapper(_wrapper)
    File "/usr/lib/python2.7/curses/wrapper.py", line 43, in wrapper
      return func(stdscr, *args, **kwds)
    File "/home/ubuntu/nyx/nyx/curses.py", line 214, in _wrapper
      function()
    File "/home/ubuntu/nyx/nyx/__init__.py", line 155, in draw_loop
      interface = nyx_interface()
    File "/home/ubuntu/nyx/nyx/__init__.py", line 217, in nyx_interface
      Interface()  # constructor sets NYX_INTERFACE
    File "/home/ubuntu/nyx/nyx/__init__.py", line 574, in __init__
      visible_panels = self.page_panels()
    File "/home/ubuntu/nyx/nyx/__init__.py", line 638, in page_panels
      return [self._header_panel] + self._page_panels[page_number]
  IndexError: list index out of range

Truncation of '.0' from values had a bug where it made units disappear...

  Bandwidth (limit: 5/s, burst: 1/s, measured: 10.3 KB/s):