[JainAyush-01]

This Pull request:

Fixes silent data loss when merging ROOT files containing nested subdirectories (depth >= 2) with chunking / batching enabled (TFileMerger::SetMaxOpenedFiles or hadd -n).

Changes or fixes:

Root Cause:

When TFileMerger merges files incrementally in batches (type & kIncremental), TFileMerger::MergeOne cleans up in-memory directory objects at lines 830–834:

if (ownobj && (!(type & kIncremental) || dirobj->GetFile() != target)) {
   dirobj->ResetBit(kMustCleanup);
   delete dirobj;
}

For top-level subdirectories (depth 1), target is the TFile* itself, so dirobj->GetFile() == target evaluates to true and dirobj is preserved across batches until the final fOutputFile->Write() flushes all accumulated objects.

However, for nested subdirectories at depth $\ge 2$ (e.g., nested/dir):

• target is the parent TDirectoryFile* (nested)
• dirobj->GetFile() returns the root TFile* (fOutputFile)

Because a TDirectoryFile* does not equal a TFile*, dirobj->GetFile() != target always evaluated to true. This prematurely deleted the in-memory subdirectory dirobj before subsequent batches could write and flush into it, causing silent loss of objects merged in later batches.

Fix:

Change dirobj->GetFile() != target to dirobj->GetFile() != fOutputFile. This correctly checks whether dirobj belongs to the output file being accumulated across incremental batches, ensuring nested directories and their contents remain intact in memory until written.

Testing:

• Added GoogleTest regression test TEST(TFileMerger, NestedDirMaxOpenedFiles) in io/io/test/TFileMergerTests.cxx.
• Validated with multi-level directory depths (depth 2 and depth 3) and batch configurations (-n 2, -n 3).
• Verified 100% pass on full gtest-io-io-TFileMerger test suite via ctest.

Checklist:

• tested changes locally
• updated the docs (if necessary)

This PR fixes #23666

---

AI-assisted coding disclosure: I used an AI assistant to help navigate the codebase, structure the unit test, and format this PR description. I have personally reviewed, debugged, and tested the C++ changes locally and take full responsibility for them.

[github-actions]

Test Results

    24 files      24 suites   3d 23h 28m 7s ⏱️
 3 900 tests  3 900 ✅ 0 💤 0 ❌
83 506 runs  83 506 ✅ 0 💤 0 ❌

Results for commit 971a552.

♻️ This comment has been updated with latest results.

[JainAyush-01]

Thanks @vepadulano and @enirolf for improving the test and adding the comprehensive roottest coverage!

I noticed in the CI run for commit 65ae899 that the Alma10 build failed in gtest-io-io-TFileMerger due to histogram auto-registration being off in that configuration. Please let me know if you would like me to adjust the test or if you prefer to handle it directly. Happy to help!

[enirolf]

Hi @JainAyush-01, thank you for the proposed fix. I've changed the test to also run correctly without object auto-registration, let's see if the CI is happy now.

[enirolf]

Hi @JainAyush-01, thank you again for pointing us in the right direction for the fix.
However, on our side, we still had to invest quite some time into understanding how exactly this proposed fix solved the underlying issue, and take care into validating it. We concluded that this PR only partially addresses the problem, both in terms of the fix and the testing thereof. As such, we have decided to instead open #23878 and work from there, and close this PR.