QEMU data loss fixes and PostgreSQL replication work

QEMU and PostgreSQL developers pushed correctness fixes for storage races and logical replication, alongside policy and performance proposals. The day centered on data integrity under concurrency, durable slot state, and clearer virt security boundaries.

QEMU preallocate filter data loss fix

A thirteen patch series on qemu-devel addresses data loss in the block layer preallocate filter. The work targets races in multi-AioContext writes and preallocation state. Operators running QEMU storage stacks should track it because concurrent I/O can otherwise lose data silently.

PostgreSQL logical decoding DDL and subtransactions

Patch v5 for logical decoding bug 19555 is under review on pgsql-hackers. The bug triggers an assert around tuplecid handling when DDL mixes with subtransactions. Backport to all stable branches is under discussion, which matters for replication users who combine schema changes with nested transactions.

QEMU AI generated contributions policy

QEMU posted v4 of a docs and AGENTS.md change that relaxes a blanket ban on AI generated contributions. The draft adds attestation trailers and pre approval for large submissions. Contributors need the rules so AI assisted patches can be accepted and audited consistently.

PostgreSQL slot invalidation persistence order

A patch ensures replication slot invalidations are written to disk before shared memory publishes them. Publishing first can leave restart inconsistencies after a crash. Logical replication and HA setups depend on durable slot state surviving process restarts.

QEMU secure marking for virt devices

A twenty eight patch series marks user creatable devices secure for virtualization use cases. The series defines security boundaries for KVM and Xen. Operators who expose devices to guests care which components sit inside the trusted computing base.

PostgreSQL UNDO with constant time recovery

Greg Burd posted v2 of a series adding UNDO for FILEOPS with claimed constant time rollback. Early design and bug comments arrived on pgsql-hackers. The design aims to keep recovery cost independent of the amount of work being undone.

Direct TOAST v2 in PostgreSQL

A v2 patch proposes direct TID addressing for TOAST tuples. It seeks fewer index lookups and lower write amplification without a data migration. Workloads with large attributes would gain most if the approach lands cleanly.

Logical replication apply worker crash on deferrable PK

A patch fixes an assertion crash in the apply worker when a subscriber table has only a deferrable primary key and no REPLICA IDENTITY FULL. Apply fails under that schema layout. Subscribers using deferrable constraints need the fix to keep workers stable.