The store benchmark, the runner (bench/store/RESULTS.md, "Size and compression"), the same accounts in each, KiB an account:
as they are
compressed
a table for each account
258
145 with each block compressed (zstd 3), 126 with a dictionary: measured, not built
RocksDB
255
174 (Snappy), 153 (zstd)
fjall
256
A table as the server writes it holds what RocksDB holds in the same bytes; RocksDB as it is usually run is two thirds of that. A compressed block costs a read not in the cache 6 to 8 µs more (8.5 µs now, RocksDB 12.4 with its blocks as they are, 17.4 with Snappy).
What it would take
A table's format has a version in its footer (PDSTAB02); a block's compression would be a third version, the earlier two read as they are. The benchmark's table writer has the compressed variants to compare with (bench/store/src/tables.rs).
The store benchmark, the runner (
bench/store/RESULTS.md, "Size and compression"), the same accounts in each, KiB an account:A table as the server writes it holds what RocksDB holds in the same bytes; RocksDB as it is usually run is two thirds of that. A compressed block costs a read not in the cache 6 to 8 µs more (8.5 µs now, RocksDB 12.4 with its blocks as they are, 17.4 with Snappy).
What it would take
A table's format has a version in its footer (
PDSTAB02); a block's compression would be a third version, the earlier two read as they are. The benchmark's table writer has the compressed variants to compare with (bench/store/src/tables.rs).