MinIO 30,943 stores objects as plain files on ordinary drives, with no metadata database. Drives form erasure sets of 2 to 16; SipHash of the object name picks the set, and each object part is cut into data shards plus Reed-Solomon parity shards, one per drive. With EC:2 on four drives, any two shards rebuild the data. A per-object xl.meta on every drive records the version, the layout and a HighwayHash bitrot checksum per shard, so silent corruption is detected on read and repaired from the other shards.
The multipart object from Consistency and Versioning shows the layout, and survives losing two drives' shards:
# Where MinIO put orders.jsonl: shards of every part on each of the four drives.
docker exec l1-minio sh -c 'cd /data; for d in disk*; do
echo "$d: $(du -b $d/booknest-demo/orders.jsonl/*/part.* | sed "s|\t.*/| |" | tr "\n" " ")"
done'
# Lose two of the four drives' shards: the object still reads back byte for byte.
docker exec l1-minio sh -c 'rm -rf /data/disk[13]/booknest-demo/orders.jsonl'
mc cat lake/booknest-demo/orders.jsonl | sha256sum | cut -c1-16
sha256sum "/mnt/d/Books/Data Engineering/demos/ch03/data/orders.jsonl" | cut -c1-16
mc admin heal -r lake/booknest-demo >/dev/null 2>&1 # rebuild the missing shards nowdisk1: 8389120 part.1 3399323 part.2 disk2: 8389120 part.1 3399323 part.2 disk3: 8389120 part.1 3399323 part.2 disk4: 8389120 part.1 3399323 part.2 ff51790e08776738 ff51790e08776738
Each drive holds 11,788,443 bytes, so the 23.6 MB object occupies 47.2 MB, the 2x of EC:2 plus bitrot hashes. With two drives' shards gone, the read still matched JSON, Columnar and Binary Formats's SHA-256; mc admin heal then rewrites them. In production the drives sit in four servers; here they share one volume.