HashBackup performance test: Storj, B2, Wasabi, iDrive, S3

[littleskunk]

@hashbackup

One SNO I talked with said they are getting hundreds of thousands of requests every day for 256 bytes, sometimes 512 bytes (about 10-20%), with an offset into a piece. For that to happen, a customer must be doing very small ranged gets. As a customer, I would expect that if I read 10 bytes, I get billed for 10 bytes; not 9K. My actual concern was that the customer was getting billed only for the bytes he read, but Storj was paying for all the overhead, so Storj was paying 900x more to SNOs than it was receiving for this download.

Customer and SNO usage tracking is based on the same signed orders that the SNO collects and sends back to the satellite. That has to line up.

[Toyoo]

@hashbackup

If he’s reading 90 bytes per request for example, he is paying Storj about 100x what he would pay on competing services.

I’m not sure about that. Hyperscalers also charge per request. A billion GET requests on AWS S3 costs 400 USD just in per-request fees; and bandwidth fees would probably have to include a few kilobytes in headers/metadata/TLS on top of actual data transfered, back-of-a-napkin estimate would be another 50–300 USD depending on setup.

Same with CloudFlare R2, 360 USD for per-request fees.

I’m deliberately not comparing this to Wasabi (their egress policy prohibits this kind of traffic) or Backblaze (performance/durability concerns).

A billion 90-byte requests counted as 9kB requests in Storj costs 63 USD.

[hashbackup]

Good points, thanks. I was lazy and didn’t do the full comparison. I guess this customer is smarter than me!

[hashbackup]

@Toyoo

Same with CloudFlare R2, 360 USD for per-request fees.

There is a twist here: with CF’s CDN, the object gets cached on an edge server, even if a ranged get is used. The fetch from R2 is an operation charge but subsequent fetches from the CDN edge server have no egress or operation fees. There are conditions on caching, for example, the object has to be 512MB or less. A larger file would have to be split into 512MB objects.

[Toyoo]

Yeah, this bit is more difficult to estimate. It seems like for a ranged query missing in cache, CloudFlare CDN would fetch the whole object, so this depends a lot on how many edge servers would have to do it, and how often. We don’t have that data.