importtime-waterfall

Generate waterfalls from -Ximporttime tracing.

install

pip install importtime-waterfall

note: importtime-waterfall requires python3.7+

usage

importtime-waterfall provides a single executable by the same name.

importtime-waterfall takes a module name as a positional argument. This is the module that will be profiled.

--include-interpreter-startup

Include tracing information of modules that are always imported as part of interpreter startup. These are usually not interesting and so they are left out by default.

--har

Output as an HTTP Archive or "HAR" file. Yes these aren't actual HTTP requests but it's an easy way to get a visualization using a standardized data format.

HAR unfortunaly doesn't have microsecond resolution so all times in the HAR output are * 1000 (1 μs => 1ms).

The easiest way to use the output of this is to paste it into a har viewer.

I use the following:

$ importtime-waterfall importtime_waterfall --har | xclip -selection c

xclip takes the output and puts it onto the clipboard. Alternatively, you can redirect to a file (> foo.har) and upload it that way.

Once pasted into the viewer you can inspect the output.

The blocked import time is represented as "waiting" (purple) and the self time is represented as "receiving" (grey). Generally when looking for slow modules look for ones with large grey chunks.

--graph

(this is the default display). Display the output as a tree. This doesn't really add much on top of what python -Ximporttime already displays (but was useful for developing / debugging this tool). I guess it's in human order instead of reversed so that's something 🤷.

Times displayed next to the module names are self-times in μs.

$ importtime-waterfall importtime_waterfall
importtime_waterfall (419)
  argparse (864)
    re (599)
      enum (661)
      sre_compile (270)
        _sre (109)
        sre_parse (336)
          sre_constants (339)
      copyreg (161)
    gettext (1056)
      locale (820)
  datetime (768)
    time (234)
    math (57)
    _datetime (154)
  json (254)
    json.decoder (446)
      json.scanner (481)
        _json (193)
    json.encoder (443)
  subprocess (628)
    signal (1030)
    errno (101)
    _posixsubprocess (40)
    select (51)
    selectors (543)
      collections.abc (184)
    threading (578)
      traceback (394)
        linecache (162)
          tokenize (911)
            token (178)
      _weakrefset (217)
  typing (1469)

success stories

I used this to find a 24% speedup in flake8's startup.

nitty-gritty how it works

importtime-waterfall imports the profiled module in a subprocess while setting the -Ximporttime flag. importtime-waterfall takes picks the best-of-5 (by total time) and uses that result. It parses the "import time:" lines and then outputs.