On 03/12/2025 11:54, Maxim Orlov wrote:
The biggest problem with compression, in my opinion, is that losing
even one byte causes the loss of the entire compressed block in the
worst case scenario. After all, we still don't have checksums for the
SLRU's, which is a shame by itself.
Again, I'm not against the idea of compression, but the risks need to
be considered.
There are plenty of such critical bytes in the system where a single bit
flip renders the whole block unreadable. Actually, if we had checksums
on SLRU pages, a single bit flip anywhere in the page would make the
checksum fail and render the block unreadable.
If things go really bad and you need to open a hex editor and try to fix
the data manually, it shouldn't be too hard to deduce the correct base
offset from surrounding data.
As a software developer, I definitely want to implement compression and
save a few gigabytes. However, given my previous experience using
Postgres in real-world applications, reliability at the cost of several
gigabytes would not have caused me any trouble. Just saying.
+1. If we decide to do some kind of compression here, I want it to be
very simple. Otherwise it's just not worth the code complexity and risk.
Let's do the math of how much disk space we'd save. Let's assume the
worst case that every multixid consists of only one transaction ID.
Currently, every such multixid takes up 4 bytes in the offsets SLRU, and
5 bytes in the members SLRU (one flag byte and 4 bytes for the XID). So
that's 9 bytes. With 64-bit offsets, it becomes 13 bytes. With the
compression, we're back to 9 bytes again (ignoring the one base offset
per page). So in an extreme case that you have 1 billion multixids, with
only one XID per multixid, the difference is between 9 GB and 13 GB.
That seems acceptable.
And having just one XID per multixid is a rare corner case. Much more
commonly, you have at at least two XIDs. With two XIDs per multixid, the
difference is between 14 bytes and 18 bytes.
And having a billion multixids is pretty extreme. Your database is
likely very large too if you reach that point, and a few gigabytes won't
matter.
One could argue that the memory needed for the SLRU cache matters more
than the disk space. That's perhaps true, but I think this is totally
acceptable from that point of view, too.
- Heikki
