[Regression]: Webkit 6 times CPU load since 1.59.0

[nihiloid]

Last Good Version

1.58.2

First Bad Version

1.59.0

Steps to reproduce

Playwright	WebKit	Runs	WebKit CPU-sec (median)	Range	Session sec (median)	CPU per sec (median)	Peak % (median)	Node CPU-sec (median)
1.50.1	2123	5	5.2	4.3-7.1	43	0.123	119	1.6
1.51.1	2140	5	10.3	1.7-26.3	91	0.085	134	2.4
1.52.0	2158	5	5.3	1.0-60.7	43	0.116	129	1.4
1.53.2	2182	5	4.1	3.9-47.5	44	0.099	119	1.5
1.54.2	2191	5	5.1	4.0-9.8	43	0.104	124	1.5
1.55.1	2203	5	5.0	1.1-12.2	39	0.103	124	1.4
1.56.1	2215	5	5.3	4.0-24.7	46	0.094	119	1.5
1.57.0	2227	5	6.3	4.4-9.4	43	0.108	119	1.6
1.58.2	2248	5	5.2	4.3-13.0	44	0.103	124	1.8
1.59.0	2272	5	36.3	31.2-156.2	44	0.737	373	1.3
1.59.1	2272	5	88.4	27.9-129.8	45	0.638	381	1.3
1.60.0	2287	5	28.5	26.1-63.1	43	0.640	366	1.3
1.61.1	2311	5	30.5	25.0-187.0	48	0.527	365	1.4
1.62.1	2336	5	72.7	30.8-242.0	47	0.659	383	1.4

Grouped by WebKit revision

WebKit	Playwright versions	median CPU per sec	range	median peak
18.2-18.5	1.50.1, 1.51.1, 1.52.0, 1.53.2	0.107	0.085-0.123	~120%
26.0	1.54.2, 1.55.1, 1.56.1, 1.57.0, 1.58.2	0.103	0.094-0.108	~120%
26.4, 26.5	1.59.0, 1.59.1, 1.60.0, 1.61.1, 1.62.1	0.640	0.527-0.737	~375%

WebKit 26.4 and later use about 6 times the CPU of every earlier revision for
the same work, at the same session length.

1.59.0 and 1.59.1 ship the same WebKit build (2272), and both sit in the heavy
group at 0.737 and 0.638 CPU per sec. The step therefore comes with the WebKit
build, not with a change in the playwright-core layer between those two
patches. 1.58.2 (WebKit 2248, 26.0) is the last version before the step.

Method

• 14 Playwright builds, 5 counted runs each, one Docker container per version.
  13 minors from 1.50 up, plus 1.59.0 next to 1.59.1 to separate the
  playwright-core layer from the WebKit build.
• Workload: live run against real webite device descriptor pinned to one iPhone fixture.
• CPU sampled from /proc every 200 ms over the whole WebKit process tree,
  cross-checked against cgroup v2 cpu.stat for the container.
• Container pinned to --cpuset-cpus 0-3, so 400% is the ceiling.

Confidence

• The 6x step at WebKit 26.4 is far larger than any noise in this data, is
  reproduced by 5 independent Playwright versions across 25 runs, and shows in
  the cgroup counter as well as the /proc sampler. It is real.

Expected behavior

No elevated CPU load

Actual behavior

Webkit with Playwright patches consumes 6 times CPU load since 1.59.0 with bundled Webkit 26.4

Additional context

MacOS 26.6 M4Max running Docker container

Environment

System:                                                                                    
      OS: Linux 6.10 Debian GNU/Linux 12 (bookworm) 12 (bookworm)                              
      CPU: (16) arm64 unknown                                                                  
      Memory: 6.40 GB / 7.65 GB                                                                
      Container: Yes                                                                           
    Binaries:                                                                                  
      Node: 22.23.2 - /usr/bin/node                                                            
      npm: 12.0.2 - /usr/bin/npm                                                               
    Languages:                                                                                 
      Bash: 5.2.15 - /usr/bin/bash                                                             
    npmPackages:                                                                               
      playwright: 1.59.0 => 1.59.0

[dcrousso]

thanks for the detailed report!

the CPU increase looks real, but the current data does not isolate it to the WebKit build

1.59.0 and 1.59.1 using revision 2272 only rules out a change between those patch releases, since both playwright-core and WebKit changed between 1.58.2 and 1.59.0

the iPhone device descriptors also changed their userAgent from Version/26.0 to Version/26.4 in 1.59.0, so the live site may be serving or running different code

could you rerun 1.58.2 and 1.59.0 with the exact same userAgent

could you also run 1.58.2 with WK_DISABLE_FRAME_SESSIONS=1, since frame sessions were enabled for revision 2248 and disabled starting with revision 2256

please also share the site and script, or a self contained reproduction, and split the CPU usage by WebKit process

if those controls still show revision 2248 staying low and revision 2272 staying high, the next step is bisecting revisions 2249 through 2272

[nihiloid]

Here is what we have so far.

Our harness sets userAgent explicitly on the context, identical for both versions:

Mozilla/5.0 (iPhone; CPU iPhone OS 18_7 like Mac OS X) AppleWebKit/605.1.15 (KHTML, like Gecko) Version/26.0 Mobile/15E148 Safari/604.1

CPU split by process, headful under xvfb, no init scripts:

	total CPU-s	MiniBrowser	WebKitWebProcess	NetworkProcess
1.58.2 (2248), 1 nav	5.9	0.7	4.6	0.6
1.59.0 (2272), 1 nav	11.8	7.0	4.3	0.6
1.58.2 (2248), 10 navs	41.3	2.9	33.6	4.7
1.59.0 (2272), 10 navs	84.4	50.3	30.2	4.4

The web process is flat across both revisions. The entire delta is in MiniBrowser, and it scales per navigation: ~0.24 CPU-s per extra navigation on 2248 versus ~4.81 on 2272.

It only appears headful. Headless, same comparison: 1.638 → 1.491 CPU-s per second, i.e. no difference. That is consistent with the cost sitting in the window/compositing path.

WK_DISABLE_FRAME_SESSIONS=1 on 1.58.2 cannot launch:

browser.newPage: Protocol error (Console.enable): 'Console' domain was not found

we are finishing a self-contained harness - plain Playwright WebKit, one public URL, no init scripts, no proxy, no interception, CPU sampled from /proc and grouped by process role. We will post it shortly so you can run it directly.

[nihiloid]

Bisecting found it - the step is at 2258.

2.6x on CPU per sec, and MiniBrowser jumps 0.7 → 6.7 (10x) in a single revision step

2257 is the one revision the CDN does not serve (400), so the change landed in 2257 or 2258

[nihiloid]

I've put test harness with comprehensive docs into https://github.com/nihiloid/pw-webkit-cpuload, hope it helps to diagnose

[dcrousso]

this should be fixed in #42438

[nihiloid]

Devin Rousso (@dcrousso) This hasn't been deployed as part of 1.63.0? The issue is still persisting in it

[dcrousso]

working on a few more upstream fixes WebKit/WebKit#73786 WebKit/WebKit#73787 that should further help resolve this