> > Explicitly
> > locking (assuming you stay in your lane) should only need to guard
> > against access from other
> > backends of this type if using shared buffers, so will be use-case dependent.
>
> I'm not sure what you mean here?
I'm mainly pointing out that the specific code that manages this
feature is the only one who has to worry about modifying said page
region.
> > This does have a runtime overhead due to moving some offset
> > calculations from compile time to
> > runtime. It is thought that the utility of this feature will outweigh
> > the costs here.
>
> Have you done some benchmarking to give an idea of how much overhead we're
> talking about?
Not yet, but I am going to work on this. I suspect the current code
could be improved, but will try to get some sort of measurement of the
additional overhead.
> > Candidates for page features include 32-bit or 64-bit checksums,
> > encryption tags, or additional
> > per-page metadata.
> >
> > While we are not currently getting rid of the pd_checksum field, this
> > mechanism could be used to
> > free up that 16 bits for some other purpose.
>
> IIUC there's a hard requirement of initdb-time initialization, as there's
> otherwise no guarantee that you will find enough free space in each page at
> runtime. It seems like a very hard requirement for a full replacement of the
> current checksum approach (even if I agree that the current implementation
> limitations are far from ideal), especially since there's no technical reason
> that would prevent us from dynamically enabling data-checksums without doing
> all the work when the cluster is down.
As implemented, that is correct; we are currently assuming this
specific feature mechanism is set at initdb time only. Checksums are
not the primary motivation here, but were something that I could use
for an immediate illustration of the feature. That said, presumably
you could define a way to set the features per-relation (say with a
template field in pg_class) which would propagate to a relation on
rewrite, so there could be ways to handle things incrementally, were
this an overall goal.
Thanks for looking,
David
