[RDF] Fix the events/s figure in the progress bar's final line

[kutsibalci]

This Pull request:

Changes or fixes:

ProgressHelper::PrintStatsFinal computes the events/s figure from a duration that has already been truncated to whole seconds:

const auto elapsedSeconds =
   std::chrono::duration_cast<std::chrono::seconds>(std::chrono::system_clock::now() - fBeginTime);
...
stream << "  " << std::scientific << std::setprecision(2) << (double)totalEvents / elapsedSeconds.count()
       << " evt/s";

duration_cast truncates toward zero, so elapsedSeconds.count() is an integer. Two consequences:

• an event loop shorter than one second divides by zero and the line reads inf evt/s;
• any longer loop is divided by a value that is up to one second too small, so the reported rate is always too high — by 100% for a run of just under two seconds.

Measured on this branch's parent, one 400-entry tree with a Filter that spins for a fixed time per entry, so the real rate is known:

real event loop	measured rate	master prints	this PR prints
0.447 s	894 evt/s	inf evt/s	8.94e+02 evt/s
2.408 s	166 evt/s	2.00e+02 evt/s (+20%)	1.66e+02 evt/s
5.607 s	71.3 evt/s	8.00e+01 evt/s (+12%)	7.13e+01 evt/s

The fix keeps the truncated value for the h:mm display, which wants whole seconds, and divides by the full-precision duration:

const std::chrono::duration<double> elapsed = std::chrono::system_clock::now() - fBeginTime;
const auto elapsedSeconds = std::chrono::duration_cast<std::chrono::seconds>(elapsed);

prettyPrint(elapsedSeconds) is unchanged, so the elapsed-time field prints exactly as before.

The in-run line is not affected and is not touched. PrintProgressAndStats takes its rate from EvtPerSec(), which averages the per-interval rates recorded in RecordEvtCountAndTime using a duration<double>. Only the final line had the integer division.

This is independent of #15323. That issue is about the numerator on the same line — ComputeTotalEvents() reports the size of the input dataset rather than what was processed. This PR deliberately does not touch the numerator; whichever way #15323 is decided, the denominator has to be a real duration.

Reproducer used for the table

The counts go to a file because the progress bar clears the terminal line with \r plus padding and swallows anything printed beside it.

#include <fstream>
#include <chrono>
void rate_repro()
{
   {
      TFile f("rate.root", "RECREATE");
      TTree t("t", "t");
      int x = 0;
      t.Branch("x", &x);
      for (int i = 0; i < 400; ++i) { x = i; t.Fill(); }
      t.Write();
   }
   std::ofstream out("rate.txt");

   auto run = [&out](const char *label, int spinUs) {
      ROOT::RDataFrame df("t", "rate.root");
      ROOT::RDF::Experimental::AddProgressBar(df);
      auto slow = df.Filter([spinUs](int) {
         auto stop = std::chrono::steady_clock::now() + std::chrono::microseconds(spinUs);
         while (std::chrono::steady_clock::now() < stop) {}
         return true;
      }, {"x"});
      auto t0 = std::chrono::steady_clock::now();
      auto n = *slow.Count();
      std::chrono::duration<double> wall = std::chrono::steady_clock::now() - t0;
      out << label << " entries=" << n << " wall=" << wall.count()
          << " true_evt_per_s=" << n / wall.count() << std::endl;
   };

   run("SHORT", 1000);
   run("MID", 6000);
   run("LONG", 14000);
}

Checklist:

• tested changes locally
• updated the docs (if necessary) — not necessary, no documented behaviour changes

Built from source at 6b57f8fc (ninja, gcc 13, -O3) and exercised through tree/dataframe/test/dataframe_helpers:

• with only the test applied and the source unchanged, RDFHelpers.ProgressBarFinalRateIsFinite fails:

  the events/s figure is not finite: [Total elapsed time: 0:00m  processed files: 1  processed events: 100  inf evt/s]
  [  FAILED  ] RDFHelpers.ProgressBarFinalRateIsFinite
  

• with the source change, it passes, and the whole suite is green: 33 tests from 2 test suites ran. [ PASSED ] 33 tests. — including ProgressBarRestorePrecision, which is the other test that inspects this stream.

This PR fixes # — no issue filed; happy to open one if you would rather track it that way.

AI disclosure

AI-assisted (Claude Code). The tool found the truncation while I was reading PrintStatsFinal for #15323, wrote the reproducer and the regression test, and drafted this description. I built ROOT from source, ran the reproducer and the test suite myself, and checked the before/after numbers in the table against the measured wall-clock rates. I have reviewed and understood the change and take responsibility for it.

[kutsibalci]

Six days in, this PR has had no CI at all — not a red run, no run. That is the outside-contributor
approval gate rather than anything about the change, and it does not surface on the Checks tab,
so I would rather name it than leave it looking like a silent failure.

Four workflow runs exist against 4a11a851c and all four are parked at action_required, waiting
on "Approve and run":

Workflow	Queued
ROOT CI	2026-08-11 23:26
code analysis	2026-08-11 23:26
Test Coverage	2026-08-11 23:26
ROOT Python wheels	2026-08-12 08:14

check-runs for the head commit returns total_count: 0, which is why the PR reads as having no CI
rather than as waiting on one.

Restating the change so nobody has to re-read the diff — two files, +27/-3:

• tree/dataframe/src/RDFHelpers.cxx — PrintStatsFinal divided the event count by
  elapsedSeconds, which had already been duration_cast to whole seconds. Any event loop shorter
  than a second divides by zero and the final line prints inf evt/s. The fix keeps the truncated
  value for the "Total elapsed time" text, so that output is byte-for-byte unchanged, and divides by
  a full-precision std::chrono::duration<double> instead.
• tree/dataframe/test/dataframe_helpers.cxx — a regression test asserting the final line is
  printed and contains neither inf nor nan. I checked it fails on master before it passes with
  the fix, rather than only checking it goes green.

Still applies cleanly to master, and I re-read the original code upstream today to confirm nothing
has changed underneath it.

@vepadulano (assigned) / @hageboeck (last touched the progress bar) — whenever one of you has a
moment for the button.

[github-actions]

Test Results

    23 files      23 suites   3d 15h 12m 33s ⏱️
 3 857 tests  3 855 ✅ 0 💤 2 ❌
78 561 runs  78 558 ✅ 1 💤 2 ❌

For more details on these failures, see this check.

Results for commit 4a11a85.

[vepadulano]

Hi @kutsibalci ,

Thanks for your contribution and apologies for the longer wait caused by the assignees being on leave.

I see the CI was approved and ran, reporting only a few unrelated test failures, so your changes don't seem to break any existing test. I will review your code and comment accordingly.

[vepadulano]

Changes LGTM, but see a minor request for the test.

[vepadulano]

Thank you for the contribution!