Repository navigation

Stop replicated deletes from echoing across the mesh, with an offline repair for logs already affected - #895

Merged

kriszyp merged 12 commits into

mainfrom

fix/replicated-delete-relog-storm

Sep 25, 2026

Commits on Sep 25, 2026
- Ship an echoed delete run as one entry instead of wedging the leg
  
  kriszyp
  and
  claude
  committed
  69b0f33
- Hold a real copy of the queued delete and compare the framed bytes
  
  kriszyp
  and
  claude
  committed
  a6fafac
- Compare the queued delete in a reusable buffer
  
  kriszyp
  and
  claude
  committed
  a0682ba
- Say why the queued delete is copied, and guard the empty queue
  
  kriszyp
  and
  claude
  committed
  b76ca14
- Repair echoed delete runs offline instead of collapsing them on every send
  
  kriszyp
  and
  claude
  committed
  6b518da
- Guard the repair by the database's RocksDB lock and make every step recoverable
  
  kriszyp
  and
  claude
  committed
  02ed49b
- Serialize repairs, own only recorded backups, and refuse runs too large to compare
  
  kriszyp
  and
  claude
  committed
  835d65b
- Hold the database's RocksDB lock for the whole repair, and carry runs across files
  
  kriszyp
  and
  claude
  committed
  310197f
- Keep the whole repair under the lock, and never re-own files by path
  
  kriszyp
  and
  claude
  committed
  58ab410
- Resolve databases before dropping privileges, and check restore sources up front
  
  kriszyp
  and
  claude
  committed
  5c145e8
- Check restore sources' bytes before the first change, and rewind txn.state first
  
  kriszyp
  and
  claude
  committed
  9b2eb76
- Make restore's rewind durable before it swaps files, and reuse span maps
  
  kriszyp
  and
  claude
  committed
  c207249

Navigation Menu

Repository navigation

Stop replicated deletes from echoing across the mesh, with an offline repair for logs already affected - #895

Stop replicated deletes from echoing across the mesh, with an offline repair for logs already affected#895
kriszyp merged 12 commits into
mainfrom
fix/replicated-delete-relog-storm

Commits on Sep 25, 2026

Ship an echoed delete run as one entry instead of wedging the leg

Hold a real copy of the queued delete and compare the framed bytes

Compare the queued delete in a reusable buffer

Say why the queued delete is copied, and guard the empty queue

Repair echoed delete runs offline instead of collapsing them on every send

Guard the repair by the database's RocksDB lock and make every step recoverable

Serialize repairs, own only recorded backups, and refuse runs too large to compare

Hold the database's RocksDB lock for the whole repair, and carry runs across files

Keep the whole repair under the lock, and never re-own files by path

Resolve databases before dropping privileges, and check restore sources up front

Check restore sources' bytes before the first change, and rewind txn.state first

Make restore's rewind durable before it swaps files, and reuse span maps