GeezerVison

In fixed-iteration mode the participation filter is an absolute threshold on unnormalized deltas. For the uniform case that is harmless: each vertex starts at resetProb, so deltas are around 0.15 regardless of the graph size. In the personalized case, though, the total mass is 1.0 and spreads over the reachable set, so per-vertex deltas scale like 1 / reach.

For example, with d = 0.85 and about 1e8 vertices reachable at hop 4, a vertex receives about 0.85^4 / 1e8, roughly 5e-9, which is below DELTA_EPS. About half of the total mass then stops propagating after hop 4, and the ranks of farther vertices are underestimated by a lot. So "numerically identical to the untruncated run" holds for the uniform case but not for personalized PageRank on large graphs.

Two options: disable the participation filter in maxIter mode (use it only when tol is set), or make the threshold relative, for example scaled by 1 / numVertices in the personalized case.

coalesce(sum, 1.0) covers the empty graph, but not a zero sum. If sourceId is not a vertex of the graph, every rank stays 0.0, and so does every rank with resetProbability(0.0) in the uniform case. Then this divides by zero and every rank becomes NaN. GraphX's normalizeRankSum has the same behaviour, but a require that the source exists, or a guard for rankSum == 0, would give a clear error instead of silent NaNs.

With Int.MaxValue the run relies entirely on the deltas decaying. In the personalized case with resetProbability(0.0), damping = 1, so the mass circulates on any cycle without shrinking, and the loop never stops. (The GraphX runUntilConvergence has the same property.) A require(resetProb > 0) in tol mode, or a large but finite iteration cap, would avoid it.

This warning fires on every g.pageRank call. The Python API (classic and Connect) and the Connect plugin (GraphFramesConnectUtils) still call pageRank, and pageRankV2 isn't exposed there yet. So Python and Connect users get a message they can't act on, although the description says there are no user-facing changes yet. Would it make sense to add the warning in the follow-up that exposes v2 in Python and Connect?