StreamerInfo corruption or loss upon TFile update

[sawenzel]

Check duplicate issues.

• Checked for duplicates

Description

The original report is available here:
https://root-forum.cern.ch/t/streamerinfo-is-not-stored-in-tfile-with-update-mode/57383

We have an executable which loops over multiple TFiles (each containing 1 object of different classes not known to this application, though the problem is the same if all includes are added) , and for each file

opens it in the "update" mode (w/o explicitly loading the object stored in the file)
writes some simple struct with metadata
closes the file.
Afterwards, in separate processes, these files are read by other executables (Consumers) which include headers for objects inside these files.

The ClassVersion of the content for some of these files is lower than the current ClassVersion used by Consumers. Automatic schema evolution accounts for differences.

The problem appears when two initial files contain objects of the same class (MyClass in the example below) of version A while the Consumer uses higher class version B because:

the 1st updated file happens to have StreamerInfo of its original object together with the StreamerInfo of the added metadata (as shown by gFile->GetStreamerInfoList()->Print() )
the 2nd updated file has StreamerInfo only for added metadata class, reading the object from these files leads to an error.
I am wondering if there is a way to enforce the presence of the original StreamerInfo in the updated file.

The reproducer is attached:

tar xvzf tst.tar.gz ;  cd tst; chmod +x runTest.sh;
./runTest.sh
It produces (suppressing ACLIC output):

Creating library for MyClass version 1
Storing object of class version 1 in two identical files
Creating library for MyClass version 2
Updating files containing object of old class version 1 (by writing metadata object of unrelated class OtherClass)

Reading 1st (updated) file
Processing testFile.C("outv1.root")...
printing GetStreamerInfoList()->Print() for outv1.root
Collection name='TList', class='TList', size=2
 OBJ: TStreamerInfo	MyClass	
 OBJ: TStreamerInfo	OtherClass	

Reading 2nd (updated) file
Processing testFile.C("outv2.root")...
Error in <TBufferFile::CheckByteCount>: object of class MyClass read too many bytes: 12 instead of 6
Warning in <TBufferFile::CheckByteCount>: MyClass::Streamer() not in sync with data on file outv2.root, fix Streamer()
printing GetStreamerInfoList()->Print() for outv2.root
Collection name='TList', class='TList', size=1
 OBJ: TStreamerInfo	OtherClass	

Reproducer

Unzip attached file bug.zip
and run the shell scriptrunTest.sh

ROOT version

6.28/04
gcc (GCC) 12.2.0

Installation method

build from source

Operating system

MacOS, Linux

Additional context

No response

[sawenzel]

Further discussions in the forum are most likely related https://root-forum.cern.ch/t/streamer-info-disappears-when-opening-a-file-again-for-writing-using-update-option/55140

[pcanal]

As fas as I can tell the problem was solved by IO: Fix StreamerInfo record write during file update. by pcanal · Pull Request #13842 · root-project/root · GitHub which is available in v6.28/06 v6.28/08 and up.

[sawenzel]

I confirm that the issue is fixed in v6.28/08 and above. Closing the ticket.

[github-actions]

Hi @sawenzel, @pcanal,

It appears this issue is closed, but wasn't yet added to a project. Please add upcoming versions that will include the fix, or 'not applicable' otherwise.

Sincerely,
🤖

[github-actions]

Hi @sawenzel, @pcanal,

It appears this issue is closed, but wasn't yet added to a project. Please add upcoming versions that will include the fix, or 'not applicable' otherwise.

Sincerely,
🤖