Re: On markers of changed data

On Sat, Oct 7, 2017 at 6:34 AM, Stephen Frost <sfrost(at)snowman(dot)net> wrote:
>> > That’s actually what pg_rman is doing for what it calls incremental
>> > backups (perhaps that would be differential backup in PG
>> > terminology?), and the performance is bad as you can imagine. We could
>> > have a dedicated LSN map to do such things with 4 bytes per page. I am
>> > still not convinced that this much facility and the potential bug
>> > risks are worth it though, Postgres already knows about differential
>> > backups if you shape it as a delta of WAL segments. I think that, in
>> > order to find a LSN map more convincing, we should find first other
>> > use cases where it could become useful. Some use cases may pop up with
>> > VACUUM, but I have not studied the question hard enough...
>>
>> The case I've discussed with barman developers is a large database
>> (couple dozen of TBs should be enough) where a large fraction (say 95%)
>> is read-only but there are many changes to the active part of the data,
>> so that WAL is more massive than size of active data.
>
> Yes, we've seen environments like that also.

I'm pretty sure that those cases are cases where there are many more
FPIs than might be expected, due to a lack of locality. (UUID PKs can
make the size of WAL balloon, for example.)

--
Peter Geoghegan