Reza Naghibi
Reza Naghibi
This is interesting. My one problem here is that this pollutes the `beresp`/`resp` namespace. What this means is that anytime we touch a header, we now need to think about:...
> In my mind, not asking for a conditional response should, under all circumstances, be more efficient for a backend than making a useless conditional request. It might be better...
> Not having to run the validation should be faster than running one no matter if indexed or not. So you are talking about using `strcmp()` on 2 headers, right?...
> I asked for a reference why an S3 backend should be faster for 304s than 200 with no body and what I get is an angry response. Didn't give...
I did 5 requests to AWS S3 for an empty file, no etag, average response time of 147ms: ``` > time curl -vs http://localhost:8888/EMPTY_FILE -H "Host: varnish-s3-test.s3-us-west-2.amazonaws.com" -H "profile: s3_west_ro"...
> Also, could you clarify, is AWS pricing different based on the response status? I don't think so.
I did 100 requests with no ETag, and 100 with an ETag: No ETag: 81.0788ms ETag: 80.8085ms https://filebin.varnish-software.com/lAbU0L39gZx3Jxin-reza-x1-g6
Really hard to say much here. Looks like we are measuring network latency.
> the risk of serialization introduced by this change Request serialization can only happen when you set a zero TTL. Having a zero grace value does not affect this, so...
I think this is tricky because there are a handful of situations which will end up with `0s` + `uncacheable`, some either done in error and others possibly done on...