Observation (dgx-spark GB10, pyg-bench filter/PageRank harness at pinned PageRank contract, 2 warmups + 5 timed runs, median)
Same harness, two PyGraphistry commits, Twitter (81,306 nodes / 2.4M edges), engine='pandas' + igraph:
| commit |
pipeline median |
filter+PageRank stage |
cutoff+filter stage |
selected nodes |
| 49db91c (2026-07-28 baseline) |
2.344 s |
1.958 s |
0.387 s |
43,065 |
| 3fb216d (0.59.0) |
2.625 s |
2.171 s |
0.506 s |
43,065 |
GPU arm (cuDF + cuGraph) is unchanged (0.231 s vs 0.233 s) and both engines select identical node sets across commits, so this is a CPU-path cost change, not a semantic one. Both stages slowed: the compound GRAPH { MATCH ... WHERE n.degree >= $c } GRAPH { USE g1 CALL graphistry.igraph.pagerank.write(...) } stage by ~11% and the final MATCH (n)-[e]-(m) WHERE n.pagerank >= $c filter by ~31%.
On the 30M-edge GPlus graph the 0.59.0 pandas pipeline measures 34.11 s under the same contract; the locked single-thread lane of 2026-08-30 measured 36.02 s, and the July run 32.10 s under the older contract, so the trend is consistent but not isolated there.
Receipts (private pyg-bench): results/runs/filter-pagerank-diag-49db91cc-20260904/ and results/filter-pagerank-059-20260904/. Candidate: bisect the pandas row/GRAPH pipeline path between the two commits; the 0.59 changelog lists several pandas-path identity/NULL-contract and bag-semantics fixes that may have added per-stage work.
🤖 Generated with Claude Code
https://claude.ai/code/session_01WwMmVFo44ADiRRj5cxh7i1
Observation (dgx-spark GB10, pyg-bench filter/PageRank harness at pinned PageRank contract, 2 warmups + 5 timed runs, median)
Same harness, two PyGraphistry commits, Twitter (81,306 nodes / 2.4M edges),
engine='pandas'+ igraph:GPU arm (cuDF + cuGraph) is unchanged (0.231 s vs 0.233 s) and both engines select identical node sets across commits, so this is a CPU-path cost change, not a semantic one. Both stages slowed: the compound
GRAPH { MATCH ... WHERE n.degree >= $c } GRAPH { USE g1 CALL graphistry.igraph.pagerank.write(...) }stage by ~11% and the finalMATCH (n)-[e]-(m) WHERE n.pagerank >= $cfilter by ~31%.On the 30M-edge GPlus graph the 0.59.0 pandas pipeline measures 34.11 s under the same contract; the locked single-thread lane of 2026-08-30 measured 36.02 s, and the July run 32.10 s under the older contract, so the trend is consistent but not isolated there.
Receipts (private pyg-bench):
results/runs/filter-pagerank-diag-49db91cc-20260904/andresults/filter-pagerank-059-20260904/. Candidate: bisect the pandas row/GRAPH pipeline path between the two commits; the 0.59 changelog lists several pandas-path identity/NULL-contract and bag-semantics fixes that may have added per-stage work.🤖 Generated with Claude Code
https://claude.ai/code/session_01WwMmVFo44ADiRRj5cxh7i1