Logical Replica ReorderBuffer Size Accounting Issues

We've noticed an odd memory issue with walsenders for logical replication slots - They experience large spikes in memory usage up to ~10x over the baseline from ~500MiB to ~5GiB, exceeding the configured logical_decoding_work_mem.  Since we have ~40 active subscriptions this produces a spike of ~200GiB on the sender, which is quite worrying.

The spikes in memory always slowly ramp up to ~5GB over ~10 minutes, then quickly drop back down to the ~500MB baseline.

logical_decoding_work_mem is configured to 256MB, and streaming is configured on the subscription side, so I would expect the slots to either stream to spill bytes to disk when they get to the 256MB limit, and not get close to 5GiB.  However pg_stat_replication_slots shows 0 spilled or streamed bytes for any slots.

I used GDB to call MemoryContextStats on a walsender process with 5GB usage, which logged this large reorderbuffer context:

--- snip ---

--- snip ---

It's my understanding that the reorder buffer context is the thing that logical_decoding_work_mem specifically constraints, so it's surprising to see that it's holding onto ~4GB of tuples instead of spooling them.  I found the code for that here: https://github.com/postgres/postgres/blob/eb5ad4ff05fd382ac98cab60b82f7fd6ce4cfeb8/src/backend/replication/logical/reorderbuffer.c#L3557 which suggests it's checking rb->size against the configured work_mem.

I then used GDB to break into a high memory walsender and grab rb->size, which was only 73944.  So it looks like the tuple memory isn't being properly accounted for in the total reorderbuffer size, so nothing is getting streamed/spooled?

Not super familiar with this so please let me know if there's something I've missed, otherwise it seems like the reorder buffer size accounting is a bit wrong.

Thanks,

- Alex.