>I would suggest to take a look at the BRIN opclass multi-minmax currently in development.
Thank you, this does look like it could help a lot with BRIN performance in this situation!
But again, if index performance alone was the only issue, then I would simply accept the space overhead and switch to btree. However, the disk fragmentation issue still remains and is significant. It is also amplified in my use case due to using ZFS, mostly for compression. But it is worth it: I am currently observing a 13x compression ratio (when comparing disk space reported by du and select sum(octet_length(x)), so this does not include the false gains from compressing padding). But in general, any variable-sized append-only workload suffers from this fragmentation problem. It's just that with filesystem compression, there is no longer a good reason to fill up those holes and accept the fragmentation.
To be clear, the main reason why I even brought my questions to this mailing list is that I don't know how to (correctly) get past the check in heap_getnext (see my first email) when implementing the workaround as a custom table access method. A reloption could theoretically disable free space maps entirely for some added efficiency, but I'm inclined to agree that this is not really needed.
On Fri, Feb 12, 2021 at 6:21 PM Noah Bergbauer <
noah@statshelix.com> wrote:
>
> A btree index on the same column is 700x the size of BRIN, or 10% of relation itself. It does not perform significantly better than BRIN. The issue here is twofold: not only does slotting these tuples into older pages significantly reduce the effectiveness of BRIN, it also causes fragmentation on disk. Ultimately, this is why CLUSTER exists. One way to look at this situation is that my data is inserted exactly in index order, but Postgres keeps un-clustering it for reasons that are valid in general (don't waste disk space) but don't apply at all in this case (the file system uses compression, no space is wasted).
>
> Any alternative ideas would of course be much appreciated! But at the moment HEAP_INSERT_SKIP_FSM seems like the most practical solution to me.
I would suggest to take a look at the BRIN opclass multi-minmax currently in development. It's designed to address that exact situation, and more review would be welcome:
https://commitfest.postgresql.org/32/2523/--
John Naylor
EDB:
http://www.enterprisedb.com