Lazy multithread RDataFrame::Snapshot cause unnessary warning and break gDirectory

[karuboniru]

Check duplicate issues.

• Checked for duplicates

Description

I am getting some Warning in <Snapshot>: A lazy Snapshot action was booked but never triggered. warning at the end of the program while found that all output files I expected to be present are present. What make things worse is that snapshot operation in this case can cause TH1D::SaveAs to stop working.

I did some test regarding the issue and found following minimal reproduce.

Reproducer

Create toy.cpp with content:

#include <ROOT/RDataFrame.hxx>
#include <TRandom.h>

void toy() {
  // comment this line out (i.e. single thread) won't reproduce the problem.
  ROOT::EnableImplicitMT();
  ROOT::RDataFrame emptydf{10000};
  auto df = emptydf.Define("x", []() { return gRandom->Gaus(); }, {});
  ROOT::RDF::RSnapshotOptions opts;
  opts.fLazy = true;
  auto act = df.Snapshot("treeout", "test3.root", "", opts);
  // this should triggr the snapshot
  *act;
  return;
}

int main() {
  toy();
  return 0;
}

Run it by rool -l toy.cpp or compile and run by c++ toy.cpp -o t $(root-config --cflags --glibs) && ./t

Result: Warning in <Snapshot>: A lazy Snapshot action was booked but never triggered. warning is given but test3.root is present.

---

Create toy.cpp with content:

#include <ROOT/RDataFrame.hxx>
#include <TRandom.h>

void toy() {
  // comment this line out (i.e. single thread) won't reproduce the problem.
  ROOT::EnableImplicitMT();
  ROOT::RDataFrame emptydf{10000};
  auto df = emptydf.Define("x", []() { return gRandom->Gaus(); }, {});
  ROOT::RDF::RSnapshotOptions opts;
  opts.fLazy = true;
  auto act = df.Snapshot("treeout", "test3.root", "", opts);
  // create a histogram and trigger the event loop
  df.Histo1D("x")->SaveAs("hist.root");
  return;
}

int main() {
  toy();
  return 0;
}

Run it by rool -l toy.cpp or compile and run by c++ toy.cpp -o t $(root-config --cflags --glibs) && ./t

Result: Warning in <Snapshot>: A lazy Snapshot action was booked but never triggered. warning is given, test3.root is present.

But hist.root that should come from TH1D::SaveAs is missing.

GDB suggest that control flow have reached TObject::SaveAs, I have no idea how to dig into the issue, but it looks to me as if

root/core/base/src/TObject.cxx

Line 690 in d245ee7

if (gDirectory) gDirectory->SaveObjectAs(this,filename,"");

is skipped due to gDirectory being casted to false as TDirectory::GetSharedLocalCurrentDirectory().get()._M_b._M_p is nullptr in this case.

ROOT version

Reproduced on

ROOT Version: 6.28/08
Built for linuxx8664gcc on Oct 13 2023, 09:48:14
From tags/v6-28-08@v6-28-08

and

ROOT Version: 6.30/00
Built for linuxx8664gcc on Nov 26 2023, 23:25:40
From heads/v6-30-patches@tags/v6-30-00

Installation method

6.28/08 from Fedora 39, self-compiled 6.30/00

Operating system

Fedora release 39 (Thirty Nine)

Additional context

No response

[dpiparo]

Thanks for this report. I can reproduce. There is something to be figured out here.
The warning is triggered here

root/tree/dataframe/inc/ROOT/RDF/ActionHelpers.hxx

Line 1702 in e46ffe8

std::all_of(fOutputFiles.begin(), fOutputFiles.end(), [](const auto &f) { return !f; }) /* never run */)

.
The condition

std::all_of(fOutputFiles.begin(), fOutputFiles.end(), [](const auto &f) { return !f; })

Is always true if Finalize is called before: fOutputFiles is always empty:

root/tree/dataframe/inc/ROOT/RDF/ActionHelpers.hxx

Line 1817 in e46ffe8

fOutputFiles.clear();

.

[fnechans]

Was this followed up on? It's not a critical issue but the extra warnings are still annoying.

[dpiparo]

Was this followed up on? It's not a critical issue but the extra warnings are still annoying.

I agree, and this is followed up in this PR, being tested now: #16550