To put it into perspective: this isn't a ridiculous amount of storage, but even in 2020 I'm mildly surprised they went all-in on SSD. Tiered storage tends to be more cost-effective. Inkbunny's main server has a 500GB NVMe SSD and 4x4GB HDDs: database and thumbnails reside on the SSD (read latency: ~0.45ms), while submissions are on HDD (~7.4ms). A big difference, but data accessed frequently typically remains in RAM. Let's hope 1.5TB is enough for FA! It's unclear what they meant by a "cluster-based server platform", but I'm inclined towards a 'chonky' box like the ThinkSystem SR950, given that a "three-headed hydra" (application server, database and storage?) shares one body.
There are other reasons to go all-in on SSD, such as power usage and reliability; perhaps those are a bigger concern where it's going. Or maybe they expect to read from storage more frequently – I recall FA's thumbnailing system acts as a cache rather than making immutable copies; scaling tends to be a high-CPU operation, so may drive core-count. Databases can parallelise to many cores, too.
To put it into perspective: this isn't a ridiculous amount of storage, but even in 2020 I'm mildly surprised they went all-in on SSD. Tiered storage tends to be more cost-effective. Inkbunny's main server has a 500GB NVMe SSD and 4x4GB HDDs: database and thumbnails reside on the SSD (read latency: ~0.45ms), while submissions are on HDD (~7.4ms). A big difference, but data accessed frequently typically remains in RAM. Let's hope 1.5TB is enough for FA! It's unclear what they meant by a "cluster-based server platform", but I'm inclined towards a 'chonky' box like the ThinkSystem SR950, given that a "three-headed hydra" (application server, database and storage?) shares one body.
There are other reasons to go all-in on SSD, such as power usage and reliability; perhaps those are a bigger concern where it's going. Or maybe they expect to read from storage more frequently – I recall FA's thumbnailing system acts as a cache rather than making immutable copies; scaling tends to be a high-CPU operation, so may drive core-count. Databases can parallelise to many cores, too.
They'll presumably still have their old hardware, which might be put out to pasture for backup storage or secondary processing. It might not stay in North Virginia, though – I guess they'll end up close to Redwood City, where IMVU is based. Judging by this, their datacenter is in Oakland; but they've been migrating to AWS. So maybe it's not a box at all, but Kurbernetes in the cloud? But then, they wouldn't be talking about servers and switches.