Performance benchmarks:

Thanks for the implementation — this is a nice addition to the Altair backend.

I had a small observation regarding the tooltip handling in draw_agents. We currently iterate over arguments["tooltip"] multiple times and perform per-row df.loc updates inside nested loops.

Since the Schelling benchmarks show a slight runtime increase, I’m wondering if the row-wise DataFrame mutation might be contributing.

It might be worth considering pre-aggregating tooltip values column-wise and extending the DataFrame in a more vectorized way instead of mutating individual cells.

Not blocking — just a performance consideration.

The tooltip handling in draw_agents currently does nested loops with per-cell df.loc[i, key] = value writes. With n agents and k tooltip fields per agent, the algorithm is O(n·k). The main issue is the repeated scalar .loc writes, since those are expensive in pandas and add a large constant cost. That likely contributes to the Schelling regression in the benchmarks.

A cleaner approach would be to compute the full set of tooltip keys once, pre-allocate python lists of length n for each key., fill them in one O(n·k) pass and assign each column to the df once.

This avoids repeated df mutation and should reduce the overhead. Given the measurable runtime increase, I think this part is worth restructuring.

There’s also an issue in how tooltip columns are determined.

Currently, the keys are taken from the first non-None tooltip only. If a later agent provides additional tooltip keys, those columns will still be created in the df via df.loc, but they won’t be included in tooltip_list. This can result in a mismatch between the df columns and the configured Altair tooltips.

A safer approach would be to compute the union of tooltip keys across all agents before constructing tooltip_list, for example by collecting all keys from non-None tooltips first and then extending tooltip_list with that full set.

Thanks @souro26 and @SAKSHI-DOT693 for the helpful feedback!

I'll fix:

1. Collect ALL unique tooltip keys (not just the first).
2. Pre-build columns vectorized instead of using df.loc in loops

This will fix both the performance regression and ensure all tooltip fields show up. Updating the PR now.

Since the Schelling benchmarks show a slight runtime increase, I’m wondering if the row-wise DataFrame mutation might be contributing.

Just to clarify: the benchmarks only run the models without the visualization being involved. There is some variability simply due to using github actions.

Thanks for this PR. I tested it locally, and it seems to be working with 1 caveat: why are x,y included? I guess this is some default behavior somewhere. However, with the explicit control this PR adds, I suggest we remove the coordinates so that only what the user specifies is included in the tooltip.

Thanks for this PR. I tested it locally, and it seems to be working with 1 caveat: why are x,y included? I guess this is some default behavior somewhere. However, with the explicit control this PR adds, I suggest we remove the coordinates so that only what the user specifies is included in the tooltip.

Thanks and I’ll remove the default x and y so only specified fields are included.