Subject Re: [Firebird-Architect] Compression
Author Thomas Steinmaurer
>> A starting point:
> Wow! Just got into the history of compression and was stunned t learn that
> DB2 invented record level compression as early as 2006. "Standing on each
> other's feet", indeed.

Oh, yes.

> OK, here's my summary.
> Record level compression. We do it, InterBase did it in 1985, Rdb/ELN did
> it in 1982. The algorithm is primitive, but fast. Jim's got another
> algorithm that compresses 40% more which was described on this list a couple
> years ago. Snappy may be another choice, but may depend on having larger
> and more consistent sample than a typical record.

Yes, probably. In case of Hadoop/HBase, we are talking about chunks of
64MB uncompressed data or in production systems even larger.

> Page level compression. Oracle does it, InnoDB does it. It would appear to
> make random access harder. Certainly for a disk system where you need to
> read and write fixed size blocks, compressing pages to odd lengths seems
> like an odd choice. Questions: how do you find a compressed page on disk?

Good question. I don't know. Perhaps storing meta information in an
uncompressed area?

> How do you prevent fragmentation when you need to write back a page that
> has grown or shrunk?

Another good question I don't have an answer. ;-) Possibly some kind of
(automatic) reorganisation/compaction.