Home | News | Documentation | Download

Snapshot of RVec<RVec<T>> produces broken ROOT file

Hello @eguiraud ,

if I may ask on this topic, somehow related question, I managed to store a ROOT::VecOps::RVec<ROOT::VecOps::RVec<>> in my output files, but even with properly generating the specific dictionary in my LinkDef.h I am never able to read it back with root (v06.22) without a seg fault. Using std::vector<std::vector<>> works just fine as expected. Is this an expected behaviour?
Cheers,
Clement

Hi,
how did you write RVecs out and what does tree->Print() show for that output TTree?

I/O of RVecs if a bit flaky currently, that’s why e.g. RDataFrame does not write out RVecs but always std::vectors, performing an on-the-fly conversion if needed. We are going to fix it in a backward-incompatible way in the next ROOT version, so writing RVecs to disk at the moment is discouraged (v6.24 will print a warning).

Oh, I see. I tried as much as possible use RVec everywhere even writing them in the TTree (I build a dictionary of C++ functions that returns RVecs). I should then change to std::vectors following your statement.

Below is a print of the ttree where you see one output RVec that works: jets_kt_pz my vector<vector> that also works jetconstituents_kt but the RVec of RVec of int jetconstituents2_kt does not.

*............................................................................*
*Br   17 :jets_kt_pz : vector<float,ROOT::Detail::VecOps:                    *
*         | :RAdoptAllocator<float> >                                        *
*Entries :    10000 : Total  Size=     381827 bytes  File Size  =     261844 *
*Baskets :       14 : Basket Size=      32000 bytes  Compression=   1.46     *
*............................................................................*
*Br   18 :jetconstituents_kt : vector<vector<int> >                          *
*Entries :    10000 : Total  Size=    3491564 bytes  File Size  =    1294513 *
*Baskets :      113 : Basket Size=      32000 bytes  Compression=   2.70     *
*............................................................................*
*Br   19 :jetconstituents2_kt : Int_t jetconstituents2_kt_                   *
*Entries :    10000 : Total  Size=      82397 bytes  File Size  =      13133 *
*Baskets :        5 : Basket Size=      32000 bytes  Compression=   6.13     *
*............................................................................*
*Br   20 :jetconstituents2_kt.fData : vector<int,ROOT::Detail::VecOps:       *
*         | :RAdoptAllocator<int> > fData[jetconstituents2_kt_]              *
*Entries :    10000 : Total  Size=     342201 bytes  File Size  =      22440 *
*Baskets :       14 : Basket Size=      32000 bytes  Compression=  15.22     *
*............................................................................*

I have copied the file in my public afs in case you want to take a look at it:
/afs/cern.ch/user/h/helsens/public/4ROOT/events_196309147.root

Cheers,
Clement

The file looks fine, those are actually all std::vectors (with a custom allocator, but ROOT I/O does not care) – probably thanks to that on-the-fly conversion that Snapshot does.

Thank you for sharing the file, can you please also share a snippet of code that crashes when reading it?

Cheers,
Enrico

It crashes just by doing events->Scan("jetconstituents2_kt").

And following up on storing RVec, could you confirm that I should change to std::vector?

Cheers,
Clement

From what you posted above, you are not storing RVecs, you are storing std::vectors (probably thanks to Snapshot performing this on-the-fly conversion I mentioned). I will check the actual file and get back to you.

yes, I understood that, sorry for not being clear enough. I’m wondering if I should avoid using the OTF conversion by directly using std::vectors, or would that have a negligible impact on performance?

In 6.22, the conversion has zero overhead. In 6.24, the conversion will have a little overhead. In 6.26, there will be no conversion and I/O of RVecs will work just fine (even better – you will be able to write std::vectors and read them as RVecs and vice-versa with pure ROOT I/O).

Bottom line: sorry for the trouble, we are making things better, and I will let you know why TTree::Scan crashes :smiley:

Hi @clementhelsens ,
I checked your file and indeed the nested RVecs are causing issues: Snapshot is only converting the outer RVec to a std::vector, and the inner RVec is still being written to file – and then you encounter precisely the kind of RVec I/O issues that we are fixing these days, e.g. in [VecOps] Use collection proxies for RVec I/O by eguiraud · Pull Request #7232 · root-project/root · GitHub .

Indeed in this case I would suggest to use std::vectors instead, as Snapshot is not doing it for you.

Sorry for the trouble, let me know if this “solves” it.

Cheers,
Enrico

Thanks @eguiraud for looking into it and very good to hear that this is already in the pipeline.
I will be using std::vector of std::vectors in the mean time. Issue understood, so I mark it with solution.
Cheers,
Clement

Hi @clementhelsens ,
double-checking, you are generating dictionaries for your nested collection types right?

Yes, I always add to the LinkDef explicitly when it’'s not standard:

1 Like

Alright, I will soon have a test in for I/O of RVec<RVec<T>> and RVec<RVec<RVec<T>>> (just to be sure), where T is a fundamental type. Your LinkDef reminds me that it’s better to also test the case in which T is a user-defined class.

Anyway, I will ping you here when ROOT’s master branch has support and test coverage for these cases, thank you for bringing this up.

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.

Hi @clementhelsens ,
fyi I just merged a test for Snapshot + nested RVecs, i.e. it should work in current master and it should not break in the future.

Cheers,
Enrico

Thanks @eguiraud , will give it a try as soon as I can get it within key4hep stack.