Look at optimization

[ehmatthes]

The app did not handle a little Show HN attention well (429), so look at a little optimization. But for a bunch of users, it might need a different deployment approach.

• [x] Start with a focus on size of dataframe. (192.0kb to start) (57.0kb after dropping to two preselected platforms, and only using visible platforms)
• [x] Run a memory profiler?
• [x] Drop to 500 data points? 250? (256)
• [x] Start with just two platforms showing?
• [ ] Deploy elsewhere, and get better metrics as well?
• [x] Run cProfile against this code.
  • [x] data processing code (pricer) (python -m cProfile pricer.py)
  • [x] plotting code
  • [x] overall streamlit app
• [ ] Cache initial pricer call.
• [ ] Cache plots
  • [ ] cost chart
  • [ ] percent chart
  • [ ] profit chart
• [x] Remove all references to matplotlib! (Do this before sharing anywhere else!)
• [x] Simplify slider options. Right now the matrix of possible options is huge. Minimize that complexity. See #24.

[ehmatthes]

Only use needed df columns

Looking at size of pricer.df, it's ~190kB no matter how many platforms are deselected. Only define, fill, and use parts of the df that are going to be visible on the page.

• [x] pricer
• [x] cost chart
• [x] por chart
• [x] profit chart

After this work, dashboard starts with only two platforms selected. Uses ~92kB, instead of 190kb. Dropping to one platform -> 57kb. No platforms -> 23.0kb.

[ehmatthes]

Other tasks

• [x] Rearrange ordering of platform references to be consistent throughout codebase and app copy.
  • [x] pricer
  • [x] nl_analyzer
  • [x] nl_compare
  • [x] nl_config
  • [x] cost_chart
  • [x] por_chart
  • [x] profit_chart

[ehmatthes]

Open questions

• [ ] Does streamlit have a paid hosting option, that's like CC but scalable? No, they do not.
• [ ] Start multiple instances from the local server. What happens to the memory footprint?
• [ ] If a user opens a tab, moves some sliders, and then doesn't interact for ten minutes, is their dataframe still in memory? If so, can we clear it at the end of every page load?
• [ ] 1GB of resources, but how much is available for data?
• [ ] Is pyinstrument helpful?

[ehmatthes]

Further optimizations

• [ ] Cache the initial dataset. How much does that affect concurrent users who don't modify the dataframe? (Probably won't work, as that user modifies the cached copy I think it affects all other users.)
• [ ] Cache the initial dataset for all platforms selected?
• [ ] Cache all of the default plotly charts?
• [x] Refactor Pricer, so it's easier to identify sections that might be optimized.
• [ ] Suggestion to cache fig.to_json().
• [x] From 6 tips, they suggest "at most" a couple dozen slider options. I have 100 large scale, 100 small scale for num_users, 1000 large scale, 100 small scale percent options, and 500 revenue options. That's a lot of options, but they're kind of necessary. See #24.

[ehmatthes]

Working notes

• [x] Add info box to pro chart explaining why y_max jumps a bit.
• [x] In summary, percentage is showing 5.8999999999999995%. Fix that!

[ehmatthes]

Observations

• Pushing first optimization to main, I'm seeing the same numbers on the deployed version as I'm seeing locally. That makes sense.

[ehmatthes]

Reduce number of data points

• [x] Make number of data points a setting in NLConfig.
• [x] Reduce this number until it degrades charts. Note change in memory.
• [x] pts, kb:
  • 1280: 97
  • 640: 41
  • 320: 21
  • 256: 17 Best balance of quality and size.
  • 160: 10

[ehmatthes]

Optimize and profile Pricer

• [x] Refactor methods to fill df. (293 lines -> 255)
  • [x] Refactor por, profit calculations.
  • [x] Fix bug with blank profit chart.
  • [x] Rename cost methods.
  • [x] Run through cost calculations, and refactor/ move to utils.
  • [x] Any other cleanup?
• [x] Simplify percent_rev_data(), if ss comes out to 0.1 anyway.
• [x] Run cProfile against this, perhaps just by running pricer.py.

[ehmatthes]

Caching

• [x] I'm not seeing a .streamlit/cache directory appear. Does that mean nothing is actually getting cached? Or is it cached in RAM? (It was in RAM. Adding persist=True makes cache dir appear.)
• [ ] Conclusion: Use cache_data(ttl=180) on each poorly get_plot() function. This works, as evidenced by cache directory showing up.
• [ ] Try some rudimentary testing on a staging deployment?
• [ ] Does caching only cache the most common charts, or does someone playing around wipe out the caches of the most common charts?
• [ ] I've seen this asked elsewhere, but can I define the values of parameters I want cached?
  • [ ] Could certainly enforce this in code: get_plot() checks num_users, if 10k, call cached function, if not call uncached function. Both of these call the same plotting function.

[ehmatthes]

Remove matplotlib

• [x] Remove profit comparison chart for now.
• [x] Remove mpl import.
• [x] grep mpl
• [x] Drops about 10MB on a simple page load.