Hi,
On 2026-04-06 18:10:56 -0700, Noah Misch wrote:
> On Mon, Apr 06, 2026 at 05:11:30PM -0400, Andres Freund wrote:
> > heap_insert()
> > ->CacheInvalidateHeapTuple()
> > ->CacheInvalidateHeapTupleCommon()
> > ->AssertCouldGetRelation()
> > not being cheap and running a *lot*.
> >
> > Admittedly it's way worse if you build with -O0, which I tend to do to make
> > debugging easier.
> >
> > In that config, the assert single-handled increases the time for a repack by
> > 35% or so.
> >
> >
> > Noah, is there any reason we need to do the AssertCouldGetRelation() before
> > the !IsCatalogRelation(relation)? Given that the goal is to make
> > RelationGetRelid() safe, it doesn't seem there is?
>
> By running AssertCouldGetRelation() during every INSERT statement, this
> detects cases that would be unsafe when the target of the INSERT happens to be
> a system catalog.
I see.
> Little of our INSERT/UPDATE coverage targets a system catalog.
Sure. We do have plenty DML doing heap_insert/update however.
> Hence, the current position is better for detection.
What if we returned early in AssertBufferLocksPermitCatalogRead() if
InterruptHoldoffCount == 0? That'd only fail if some code manually did a
RESUME_INTERRUPTS() to balance the one acquired as part of the content lock?
> I wonder if this got slower in v19. In v14-v18, the assert's cost is
> proportional to the number of held lwlocks, often 0 or 1. In v19, it's
> proportional to PrivateRefCountHash cardinality.
Yea, plausible. It will only scan PrivateRefCountHash if
PrivateRefCountOverflowed overflowed, but it did overflow in the case I was
testing...
Greetings,
Andres Freund
