Performance hit switching to sub pages

[davetapley]

I'm working on an MRE, but wanted to put this out there in case anyone has any ideas 🙏🏻

So I have something like this:

@page('/')
async def my_page():
    ui.label("My data")
    resp = await load_some_data()
    data = from_json(Data, resp.text)
    row_data = [...] # comprehension that manipulates data
    ui.aggrid( {'rowData': row_data,...

It's not the most efficient code ever, but I get the aggrid within approx. 2 seconds which is good enough for me.

---

Today I started refactoring to sub pages.

Immediately I noticed something strange: Unlike in the async sub pages demo, I would see nothing for approx. 8 seconds, then suddenly get the label and the aggrid at the same time.

Okay, so figure I should probably be using run.cpu_bound, so I update it to this:

def make_rows(data):
    return [...] # comprehension that manipulates data

async def my_page():
    ui.label("My data")
    resp = await load_some_data()
    data = await run.cpu_bound(from_json, Data, resp.text)
    row_data = await run.cpu_bound(make_rows, data)
    ui.aggrid( {'rowData': row_data,...

Great, I now get the expected behavior, but it still takes approx. 8 seconds to see the aggrid (each await seems to take a couple of seconds) 🤯

---

I went back to my original implementation but with using run.io, and it still loads in approx. 2 secs.

I tried creating an MRE using a fake endpoint and sleep to simulate the processing, but I can't reproduce.

[rodja]

That looks strange. Could you share the exact timing measurements? For example by adding logging output?
Also you can try to add a asyncio.sleep(0) after your initial label. That way you yield the execution and label can render. Also from_json and make_rows seems to need computations. You should use run.cpu_bound if that is the case (see https://github.com/zauberzeug/nicegui/wiki/FAQs#what-can-i-do-to-improve-performance).

[davetapley]

Whoops, io instead of cpu was just a typo when I was writing that pseudo code, I do have cpu_bound in my code 😊

I'm going to sanitize the data and make an MRE with the actual code (since trying to create one from scratch isn't working).
In the mean time here's some logs:

Before (with @page):

2025-08-22 00:22:54,334 INFO jefdaq.gui.main: start
2025-08-22 00:22:54,356 INFO httpx: HTTP Request: GET http://localhost:8082/cache/sensors.json "HTTP/1.1 200 OK"
2025-08-22 00:22:54,365 INFO jefdaq.gui.main: got
2025-08-22 00:22:56,780 INFO jefdaq.gui.main: from_json
2025-08-22 00:22:56,810 INFO jefdaq.gui.main: to_row_data
2025-08-22 00:22:56,810 INFO jefdaq.gui.main: done

After with sub pages:

2025-08-22 00:25:57,827 INFO jefdaq.gui.main: start
2025-08-22 00:25:57,848 INFO httpx: HTTP Request: GET http://localhost:8082/cache/sensors.json "HTTP/1.1 200 OK"
2025-08-22 00:25:57,854 INFO jefdaq.gui.main: got
2025-08-22 00:26:04,051 INFO jefdaq.gui.main: from_json
2025-08-22 00:26:04,715 INFO jefdaq.gui.main: to_row_data
2025-08-22 00:26:04,716 INFO jefdaq.gui.main: done

Almost all the delay is introduced between got (the raw JSON) and from_json (parsed to dataclass with serde).

[davetapley]

Aha, found it: If I take my original (pre sub pages) code and literally just change:

 sensors = from_json(Sensors, resp.text)

to:

sensors = await run.cpu_bound(from_json, Sensors, resp.text)

I see the same jump:

2025-08-22 19:20:06,000 INFO jefdaq.gui.main: got
2025-08-22 19:20:16,200 INFO jefdaq.gui.main: from_json

[davetapley]

I wonder if it's the overhead of pickle-ing resp.text? It is a 5MB string 😬

[davetapley]

... and sure enough using run.io_bound is comparable to my original version, which makes my typo ironic 😆

2025-08-22 20:23:29,903 INFO jefdaq.gui.main: start
2025-08-22 20:23:29,924 INFO httpx: HTTP Request: GET http://localhost:8082/cache/sensors.json "HTTP/1.1 200 OK"
2025-08-22 20:23:29,930 INFO jefdaq.gui.main: got
2025-08-22 20:23:32,343 INFO jefdaq.gui.main: from_json
2025-08-22 20:23:32,372 INFO jefdaq.gui.main: to_row_data
2025-08-22 20:23:32,372 INFO jefdaq.gui.main: done

So I'm happy that it's process vs thread overhead.

---

One last thing: I tried await asyncio.sleep(0) just after the label (and not using run) and it still didn't render the label until the subsequent await were all complete 🤔