Files
scylladb/dist/common
Yaron Kaikov b74565e83f dist/common/scripts/scylla_raid_setup: reduce XFS metadata overhead
The block size of 1k is significantly increasing metadata overhead with xfs since it reserves space upfront for btree expansion. With CRC disabled, this reservation doesn't happen. Smaller btree blocks reduce the fanout factor, increasing btree height and the reservation size. So block size implies a trade-off between write amplification and metadata size. Bigger blocks, smaller metadata, more write ampl. Smaller blocks, more metadata, and less write ampl.

Let's disable both `rmapbt` and `relink` since we replicate data, and we can afford to rebuild a replica on local corruption.

Fixes: https://github.com/scylladb/scylladb/issues/22028

Closes scylladb/scylladb#22072
2025-01-07 13:18:21 +02:00
..