Conversation

Replying to

and

You need the cache to be warmed up by usage. There isn't a way to push something to their CDN. It acts as a caching reverse proxy using their custom nginx fork rather than a traditional CDN that you push to in advance, which I don't think would fit with real-time usage.

1

1

Replying to

I guess you could log & analyze how they request data from your server to figure a lot of it out.

1

Replying to

They add a CF-Cache-Status header showing whether a cache hit occurred too.

1

1

Replying to

If you see the request, how was it not a cache miss?

2

Replying to

I mean they add the header in the response to the client.

1

1

Replying to

and

For example, try this, which should be a HIT: curl -v "www2.coinbase.com/assets/16x16.p" |& grep -i cf-cache Then, the following, which should be a MISS: curl -v "www2.coinbase.com/assets/16x16.p" |& grep -i cf-cache Then, repeat that request a dozen times and it should become a HIT after 1-3.

1

Replying to

and

On that note, I really hate Twitter's link handling... it shouldn't have the ... in the copy-pasted text. They can and should make changes like that with CSS and have it copy exactly as it was entered...

1

Replying to

and

So, from client-side testing, if you do a range request for an uncached file, it ends up pulling the whole file and caching that. It isn't quite what they did before. They still only cache files as a whole, but range requests can pull the whole file from the origin to cache it.

1

Replying to

Is the range request by any chance satisfiable before the whole file finishes transferring?

2

Replying to

and

Even if so, this sounds like Cloudflare would badly pessimize a site where most requests are small range requests.

1

Replying to

Yeah, it still handles this very oddly, because their caching still seems totally incapable of caching a portion of a file. Would need to test with a large file to confirm that it doesn't just want to cache large chunks though.

2:37 AM · Nov 5, 2018Twitter Web Client

Replying to

and

It also matters what happens if requests are made for content that's already being pulled from the origin server. I don't think it's nearly as clever as people would expect at being able to merge new requests with currently pending ones, at least for ranges.

1

1

Replying to

and

They use a fork of nginx and relevant option is nginx.org/en/docs/http/n. I think the range option might also explain the range handling: nginx.org/en/docs/http/n i.e. I think it caches a certain amount of the start of the file, but not arbitrary sub-ranges within it.

1