Re: [BUG] repack --drop-filtered aborts when a filtered blob is referenced by the index
- From
Siddharth Shrimali <r.siddharth.shrimali@gmail.com>
- Date
- Nov 12, 2026, 07:29 UTC
- Message-ID
- <20261112072921.19170-1-r.siddharth.shrimali@gmail.com>
- In-Reply-To
- <tencent_AC7FF23968DCD0078A6B863A42BCB5ED4E05@qq.com>
Hi,
On Thu, Nov 12, 2026 at 12:51:07AM +0800, kongfankaikai wrote:
Show 8 quoted lines
> When a blob selected by the filter for dropping is referenced by the > current index, Git aborts the entire repack with: > > `fatal: cannot drop 'path/to/large-file.bin' (object-id): it is > referenced by the current index` > > This stops other eligible blobs from being dropped, even though those > blobs are not referenced by the index.
Thanks for the report. This is a known limitation and a fix is already planned.
Samuel Bronson ran into exactly the same thing on a real llvm-project clone and sent an RFC patch for it in September:
https://lore.kernel.org/git/s0vqzjavw8p.fsf@gmail.com/
and I agreed with the direction there:
https://lore.kernel.org/git/CAGWgyh9B=re06aofii9VFB1xOwEeTtxYE=7T14m9WFAx1ORpMg@mail.gmail.com/
Show 5 quoted lines
> Is there a specific reason this condition is treated as fatal rather > than skipping the affected objects? > > This behavior would be especially helpful for partial clones with > large working trees.
No correctness reason. The guard only avoids dropping a blob that the next command would lazy-fetch right back, so such a blob is never at risk. Aborting was simply a conservative version of the guard, and as you and Samuel point out it is a poor trade-off.
The fix is to replace the die() with a warning and drop the remaining eligible blobs. I am working on this and will post it to the list.
Thanks, Siddharth