Streamlit startup time could be reduced from 1s to 400ms

[sfc-gh-tteixeira]

Problem

The time between streamlit run foo.py and having a blank app load in the browser feels a bit slow nowadays!

Based on a quick analysis, a simple way to cut startup time by 60% would be to lazy-load certain imports.

Methodology

1. On a new environment, install Streamlit
2. Call python -X importtime -c 'import streamlit' 2> latency.log
3. Look at latency.log by hand or with Tuna.

Interestingly, even if you run this multiple times, there's no real change in the numbers.

Machine: M2 Macbook Pro

Results

The result of step 2 above is attached as latency.log.

Also, here's a pretty chart:

Findings

Streamlit takes 1s to initialize:

import time:       641 |    1011102 | streamlit

Major culprits:

• Importing pandas inside type_util: 240ms

  import time:      2062 |     240916 |                           pandas
  import time:       927 |     281330 |                         streamlit.type_util
  import time:      2104 |     440877 |                       streamlit.runtime.state.session_state
  import time:       320 |     441196 |                     streamlit.runtime.state.safe_session_state
  import time:       175 |     442065 |                   streamlit.runtime.state
  import time:        11 |     442076 |                 streamlit.runtime.state.session_state
  import time:       444 |     465120 |             streamlit.runtime.app_session
  import time:       785 |     506798 |           streamlit.runtime.runtime
  import time:       146 |     506944 |         streamlit.runtime
  import time:        11 |     506954 |       streamlit.runtime.scriptrunner
  import time:       193 |     507147 |     streamlit.cursor
  

• Importing pandas.style in st.arrow: 209ms

  import time:       957 |     209657 |       pandas.io.formats.style
  import time:       278 |     209934 |     streamlit.elements.arrow
  

• Setting up the plotly theme: 49ms

  import time:     46609 |      49771 |       streamlit.elements.lib.streamlit_plotly_theme
  import time:      2810 |      54590 |     streamlit.elements.plotly_chart
  

• Importing altair inside of arrow_altair (mostly to set up theme): 48ms

  import time:       345 |      48145 |       altair
  import time:       124 |        124 |           streamlit.elements.lib
  import time:       264 |        388 |         streamlit.elements.lib.dicttools
  import time:       293 |        680 |       streamlit.elements.arrow_vega_lite
  import time:       300 |        300 |         streamlit.elements.form
  import time:       197 |        497 |       streamlit.elements.utils
  import time:       462 |      49782 |     streamlit.elements.arrow_altair
  

• Loading the validators library: 40ms

  import time:       309 |      40592 |       validators
  import time:       413 |      41005 |     streamlit.elements.media
  

• Importing requests inside streamlit.version: 52ms

  import time:       364 |      51844 |     requests
  import time:      1632 |      59739 |   streamlit.version
  

• String util is slow to import probably due to emoji computation: 27ms

  import time:     26409 |      27205 |       streamlit.string_util
  import time:       168 |      27587 |     streamlit.file_util
  import time:       729 |      58916 |   streamlit.config
  

Proposal

1. For some of the culprits above, we can just move the offending import out of the file's root scope and into the actual scope where it's used.
2. For the files where we import modules to set up Altair and Plotly theming, stop setting up the themes globally at import time and, instead, change the Altair and Plotly figures to use our theme before marshalling them.
3. For the emoji computation: move EMOJI_EXTRACTION_REGEX into a functools.cache'd function.
4. Consider no longer checking the Streamlit version upstream to prompt people to upgrade. We do this with 10% probability on every run, and requires requests (52ms) -- but it's unclear whether this actually causes a nontrivial number of people to actually upgrade.

---

Community voting on feature requests enables the Streamlit team to understand which features are most important to our users.

If you'd like the Streamlit team to prioritize this feature request, please use the 👍 (thumbs up emoji) reaction in response to the initial post.

[cmayoracurzio]

Thanks for analyzing this! Some time ago I opened enhancement request #5798 with the same goal in mind (reduce startup time) but mostly focused on the size and quantity of Streamlit core static files. Sharing here because looking at both (python imports and static files) might be worthwhile.

[LukasMasuch]

Thanks for the investigation 👍 I did a quick check. I think besides the Plotly theme, all other proposed changes are viable.

As far as I remember, in order to apply our Streamlit chart theme in the frontend we need to apply changes to the global Plotly theme before any chart object is created (fyi @willhuang1997).

Regarding Altair theme: removing it might break a few apps (that's why we kept it in), but it was never an official feature anyways. So, I assume this isn't a problem.

[sfc-gh-tteixeira]

We can always do this in parts: first the easy ones, then the themes (if possible).

[LukasMasuch]

Update: the altair dependency was moved into lazy loading in this PR: #6618 and pandas styler, validator, and requests here: #6531. This will be part of the 1.23 release giving us a small boost in startup performance.

[LukasMasuch]

Removed a couple of dependencies:

• Remove dateutil and tzlocal dependencies #8097
• Remove importlib-metadata dependency #8094
• Fix issue using a local path with st.image on windows. #8092 (validators)

And lazy-loaded a couple of other modules:

• Lazy-load emoji module to improve performance #8109
• Lazy-load vendored pympler asizeof module #8124
• Lazy-load the watchdog event watcher module #8111
• Lazy-load pandas and pyarrow to improve performance #8125
• Lazy-load numpy and pillow to improve performance #8134
• Use env variable to configure matplotlib backend #8113
• Refactor typing and imports #8105

This allows to reduce the import time to as low as ~200ms:

And will enable significant speed ups related to the loading time of stlite 🥳 We also now have an e2e test to make sure that we don't accidentally import any of the lazy-loaded packages.