Echoing other's praise for call-by-name week, here's the start-up time for

pants --no-pantsd version

on my real-world work codebase: (Measured with

hyperfine -L version '2.27.0a0,2.28.0.dev0,2.28.0.dev1' -w 2 -r 5 'PANTS_VERSION={version} pants --no-pantsd version'

)

Benchmark 1: PANTS_VERSION=2.27.0a0 pants --no-pantsd version
  Time (mean ± σ):      5.580 s ±  0.253 s    [User: 5.164 s, System: 0.297 s]
  Range (min … max):    5.206 s …  5.839 s    5 runs
 
Benchmark 2: PANTS_VERSION=2.28.0.dev0 pants --no-pantsd version
  Time (mean ± σ):      4.783 s ±  0.185 s    [User: 4.467 s, System: 0.267 s]
  Range (min … max):    4.587 s …  5.039 s    5 runs
 
Benchmark 3: PANTS_VERSION=2.28.0.dev1 pants --no-pantsd version
  Time (mean ± σ):      2.938 s ±  0.056 s    [User: 2.726 s, System: 0.183 s]
  Range (min … max):    2.856 s …  3.001 s    5 runs
 
Summary
  PANTS_VERSION=2.28.0.dev1 pants --no-pantsd version ran
    1.63 ± 0.07 times faster than PANTS_VERSION=2.28.0.dev0 pants --no-pantsd version
    1.90 ± 0.09 times faster than PANTS_VERSION=2.27.0a0 pants --no-pantsd version

5.6s -> 2.9s 🎉

That is very nice 🙂

The call-by-name wins keep coming: on my work repo, the start-up time has continued to drop since I last checked: • 5.3s (2.27) • 2.0s (2.28.0a0) • 1.9s (2.29.0dev0) Well done everyone.

hyperfine -L version '2.27.0,2.28.0.dev1,2.28.0a0,2.29.0.dev0' -w 2 -r 5 'PANTS_VERSION={version} pants --no-pantsd version'
Benchmark 1: PANTS_VERSION=2.27.0 pants --no-pantsd version
  Time (mean ± σ):      5.317 s ±  0.018 s    [User: 5.037 s, System: 0.259 s]
  Range (min … max):    5.297 s …  5.340 s    5 runs
 
Benchmark 2: PANTS_VERSION=2.28.0.dev1 pants --no-pantsd version
  Time (mean ± σ):      2.841 s ±  0.008 s    [User: 2.663 s, System: 0.162 s]
  Range (min … max):    2.834 s …  2.853 s    5 runs
 
Benchmark 3: PANTS_VERSION=2.28.0a0 pants --no-pantsd version
  Time (mean ± σ):      2.035 s ±  0.013 s    [User: 1.889 s, System: 0.128 s]
  Range (min … max):    2.019 s …  2.046 s    5 runs
 
Benchmark 4: PANTS_VERSION=2.29.0.dev0 pants --no-pantsd version
  Time (mean ± σ):      1.871 s ±  0.043 s    [User: 1.732 s, System: 0.117 s]
  Range (min … max):    1.843 s …  1.944 s    5 runs
 
Summary
  PANTS_VERSION=2.29.0.dev0 pants --no-pantsd version ran
    1.09 ± 0.03 times faster than PANTS_VERSION=2.28.0a0 pants --no-pantsd version
    1.52 ± 0.03 times faster than PANTS_VERSION=2.28.0.dev1 pants --no-pantsd version
    2.84 ± 0.07 times faster than PANTS_VERSION=2.27.0 pants --no-pantsd version

Whoa, that's a big drop in 2.28 - was that the union rule change?

I think that big drop was just all the porting everyone did in CBN week

Here's the measurements for all recent releases, as a starting point for bisecting (if you care about the specifics)

Version      mean   min    max
2.27.0       5.335  5.301  5.419
2.28.0.dev0  4.618  4.571  4.648
2.28.0.dev1  2.877  2.832  3.003
2.28.0.dev2  2.742  2.720  2.780
2.28.0.dev3  2.745  2.734  2.759
2.28.0.dev4  2.748  2.707  2.835
2.28.0a0     2.045  2.027  2.072
2.29.0.dev0  1.843  1.834  1.860

polymorphic call-by-name should contribute even further, but only once we port more of the union code to use it

Interesting. My 7.5s test was roughly at the end of call-by-name week I think - and I didn't think we had a substantial drop since then. Either way, stoked for the results 🤷

My memory is that 2.27.0a0 (created on May 13) had a few minor CBN changes, but most of the work for the CBN week landed after 2.27.x branch was cut... as the numbers suggest, especially in time for 2.28.0.dev0 (May 15) and 2.28.0.dev1 (May 20).

We just upgraded our repo from Pants 2.27 to 2.28 and saw startup times drop by 10x from ~30s to ~3s. We have a decent amount of custom pants plugins, so maybe that has something to do with it? In any case, this is great! Thanks all!

And once I figure out how to enable the migration tool, you should be able to cut that down further by migrating your custom plugins