The 4-node/8-drive erasure-coded cluster matched HAL's topology faithfully,
but real throughput testing against both showed why that costs more than
it's worth here: every write on the sharded cluster fans out across 4
processes over the internal network with erasure-coding overhead, capping
safe throughput around 1.3-2 MiB/s and breaking outright above ~256
concurrent transfers (IncompleteBody errors, confirmed via a controlled
512x test). The identical copy against a single-node instance sustained
23+ MiB/s at the same concurrency with zero errors — over 10x faster,
verified side-by-side, not assumed.
Trades away erasure-coded redundancy (no single-drive fault tolerance) for
that throughput. Deliberate, and reversible if it turns out to matter later
-- the data itself is migrated over the S3 API either way, so the storage
topology underneath isn't locked in by anything upstream of it.
jschoubben
20df40c949