Benchmark filesystems with lots of changes

[matheus23]

One of my accounts has lots of filesystem public files history entries. Maybe it might make sense to benchmark this. (Just wanted to write this down somewhere, this came up with of fission-codes/fission#489)

[expede]

We should 100% totally check this, but I'll be pretty surprised if it turned out to be the actual bottleneck. Here's my thinking:

• Deep structures can require a lot of round trips with bitswap
• But all (or most) of files (i.e. the heavy stuff) will be available at the same level, even if they're in lower layers
  • i.e. the "heavy" stuff most likely already available in the top layer
• I've been able to reproduce >60s updates with a single small image

Which isn't to say that trying to sync a new really deep structure won't have a lot of round trips, but each of those should be pretty lightweight because it's just pointers. That of course compounds, so if you need to do 10k round trips because of bitswap, yeah that's pretty rough.

I can absolutely be wrong — the above is pure theory — and we should definitely test this empirically 🔬🧪👍

[matheus23]

I've been able to reproduce >60s updates with a single small image

That sounds interesting! I'd love to hear more.

I'll be pretty surprised if it turned out to be the actual bottleneck.

Yeah I had this issue in my notes and I talked to James about it. I just wanted this to be persisted somewhere. It's all theory of course. Maybe this would've been more appropriate at another place 🤔

[expede]

I've been able to reproduce >60s updates with a single small image

That sounds interesting! I'd love to hear more.

With the known-effected account, I opened the public directory in Drive, and upload a ~400kb image. It took over a minute to complete.

Maybe this would've been more appropriate at another place 🤔

I think that this a great place to record this, and thanks for writing it down! We absolutely need to check this assumption.

My comments above were mainly 1. stating my assumptions, 2. getting the conversation going, and 3. when we test this, we should account for the factors in my (and others) existing assumptions, for example:

• How many files in a directory (number of flat / horizontal links)
• Time to upload a single small file vs single large file (i.e. latency vs bandwidth)
• Time for syncing deeply nested structures, when:
  • File are present at surface layer
  • Files are not present at surface layer

[matheus23]

With the known-effected account, I opened the public directory in Drive, and upload a ~400kb image. It took over a minute to complete.

Ah right, I thought you reproduced this on a new account 😅