gh-140677 :Add heatmap visualization to Tachyon sampling profiler #140680
Conversation
pablogsal
force-pushed
the
heatmap
branch
2 times, most recently
from
b1d5307 to
9ee2b80
Compare
|
CC @ambv |
Introduce a new --heatmap output format that provides line-by-line execution visualization. The heatmap shows: - Color-coded execution intensity for each line (cold → warm → hot → very hot) - Inline sample counts and percentages per line - Per-file statistics (total samples, hottest line) - Interactive call graph navigation with caller/callee buttons - Module type badges (stdlib, site-packages, project code) Unlike flamegraphs which show call stacks and time distribution, heatmaps excel at identifying hot code paths within files, understanding line-level execution patterns, and navigating through call relationships.
| </div> | ||
| <div class="stat-card"> | ||
| <span class="stat-value"><!-- TOTAL_SAMPLES --></span> | ||
| <span class="stat-label">Total Samples</span> |
There was a problem hiding this comment.
We're using "samples" twice in the same view here - once for total samples (i.e., number of snapshots taken during profiling) and the other "samples" to mean the number of lines recorded in that file.
This can be a bit confusing, since file samples can exceed the total number of samples. One idea to help disambiguate might be to add a tooltip or subtext on file samples to say something like "Line samples captured (sum across call stacks)", or rename one of them?
| <th onclick="sortTable(5)" style="text-align: right;">% of Total ⇅</th> | ||
| <th style="text-align: left;">Intensity</th> |
There was a problem hiding this comment.
I don't know if I was sleepy or what but I genuinely had to use brainpower ™️ to understand these columns.
IIUC, these two columns show the same data in different formats. It was not initially obvious to me that % of total wasn't intended to sum to 100% and that the % and bar are actually relative to the hottest file (which should always equal 100%). I feel like my brain expected the bars below the top file to be a proportion of a whole.
A thought here might be to merge the columns into a single column called "Intensity" and then show the percentage number with a coloured background using the same blue --> red heatmap. We could also add a note in the header that this is really the % relative to the hottest file.
| module_name = html.escape(stat['module_name']) | ||
| module_type = stat['module_type'] | ||
| badge_class = f"badge-{module_type}" | ||
| bar_width = min(100, stat['percentage'] * 2) # Scale for visibility |
There was a problem hiding this comment.
I think this calculation isn't quite right? It looks like a file with 46.15% intensity is actually like 92% full because of the * 2
| filename = filename or f"flamegraph.{pid}.html" | ||
| case "heatmap": | ||
| collector = HeatmapCollector(skip_idle=skip_idle) | ||
| filename = filename or f"heatmap_{pid}" |
There was a problem hiding this comment.
I wonder if folks will want to have this saved by timestamp to compare historical runs?
2 participants
Introduce a new --heatmap output format that provides line-by-line
execution visualization. The heatmap shows:
Unlike flamegraphs which show call stacks and time distribution, heatmaps
excel at identifying hot code paths within files, understanding line-level
execution patterns, and navigating through call relationships.