vdd 2023q1 notes

Nils Goroll nils.goroll at uplex.de
Mon Feb 13 14:25:57 UTC 2023


Here's the decentralized backup of our etherpad at 
https://etherpad.wikimedia.org/p/VDD2023q1

Dridi has also created a nicely formatted backup at 
https://github.com/varnishcache/varnish-cache/wiki/VDD23Q1

VDD 2023q1 https://etherpad.wikimedia.org/p/VDD2023q1

IRC: #vdd on irc.linpro.no

TOPICS?

happy varnish day
evening programme?

     Feb 7 in town


revisit last year's notes 
https://etherpad.wikimedia.org/p/VDD2022q2_topics_suggestions

vdd2022q2 #1 varnishlog scalability
===================================

- dridi talks about progress on #1 logging
   - keep shared memory, but add option to block varnishd
   - different vsm layout for improved concurrency, more page based
   - -d option would still be kept, idea of running queries in varnishd is not 
on the table any more at this time
- discussion about the nature of the problem
   - phk mentions old idea about split log (complete/incomplete tx)
   - phk idea to suppress transactions from VCL
   - phk idea to use log files only once, open a new one once full
   - phk idea keep req grouping in varnishd (into dynamic memory), flush only 
once per top level request.
   - dridi presents vsl scrubbing idea
     - replace log entries or fields with Xes
     - discussion about oob info from length - should redaction be a single 
fixed token or original length?

   tasks
   - VIP for scrubbing (dridi)
   - VIP batching on toplevel req in varnishd (PHK)

vdd2022q2 #3 Varnish Association
================================

- reiterated over status quo

vdd2022q2 #4 TLS
================

- no progress
- discussion again about the future life of vtest: binary dependency vs. source 
dependency vs. submodule
- VS has green light to contribute TLS support
- slink has talked to CF guys at H3 workshop, no known show stoppers
- h2o also has keyless "neverbleed" https://github.com/h2o/neverbleed MIT licensed
- should be one or several VEXTs (openssl, libressl, ...)

- TODO:
     - vtest
     - VIP the backend interface
     - backend definition like vmod function for vext
        backend newstuff {

     .options = {

     .foo = 123;

     }

     }


     options get passed as string at init time


vdd2022q2 #5 H3
===============

- asad: still working on it with msquic
- also researching alternatives
- haproxy also needs patched openssl

Tenants and VCL files
=====================

- hermunn/dridi: want to add a parameter to vcl.load as the base path for 
includes with fall back on global vcl_path, ban absolute includes from vcl

vdd2022q2 #6 Extensions
=======================

- do we want extension arguments?
- env vars working find for now

BANs and persistent storages
============================

- loading old storage requires a control system
- control system can assign the "ban id" (current vtim_real)
- CLI command to return min(max(silo(ban_id))) (or per silo max(silo(ban_id))) ?
- feature flag to deny bans from vcl
- multi tenancy support

General idea to ponder:  proper VCL "jails" for multi-tenancy

Direction of VCL?
=================

- local, task scoped variables/objects
- vcl functions
- bump vcl version every time? -> NO
- new include version check?
   - "umbrella" vcl x.y requires x.Y with Y <= y
   for (x in vmod.bla(...)) {
   }
- foreach (key[, val[, idx]]) in vmod.something(arguments) {
       key type is defined by vmod.something iterator
       val type is defined by vmod.something iterator
       idx is integer
   }
- foreach x in vmod.something(arguments) {
       x type is defined by vmod.something iterator
       can have x.key x.index, x.whatever
   }
   # built-in structured fields knowledge?
    foreach bla in req.http {  // bla has type "http_header"
       what members does `bla` have ?
       // "literal, string, iterable".fields possible too
       if (bla.name == "something") {
         for fld in bla.val.fields { // comma, fld has type sf-element (?)
             for x in fld.attr { // semi-colon, x has type sf-attr (?)
                 x.key x.val
             }
         }
       }
    }

    jobj = vmod_json(input)
    for x in jobj.somthing_i_really_want(dont_fail_hard=True) {
    }
    if (jobj.failed) {
    }

    foreach bla in req.http {
       if (bla.name ~ "(?i)cache-control") {
           foreach directive in std.split(bla.val, ", ") {
               ...
             }
         }
    }

=> Everyone should re-write their complex vcl how they would
    want it to look

Fastly-Syntax for SF

     resp.http.Cache-Control["max-age"] = 2h;



function foo(myvmod.mytype x) {
}

What about:
for (x in myvmod.bla()) {
     foo(x);
}
Also, what if the above is in foo()? Answer: Maybe we will allow recursive 
functions, maybe not

# foo takes type returned by vmod?

compliance
==========

Dridi gives a very short talk about a minor change. He will share his 49 page 
presentation later.

https://github.com/varnishcache/varnish-cache/files/10707808/vdd23q1-compliance-stuff.pdf

====
DAY2
====

VCL SYNTAX DISCUSSION AROUND FIELD ACCESS
=========================================

if (req.http.Cache-Control[private])  => true if private present
if (req.http.Cache-Control[private] is none) => true if private has no value
if (req.http.Cache-Control[private] == "public") => true if value of private 
folds to

     string "public"


if (req.http.Cache-Control has private)

if (req.http.Cache-Control[private] is flag) => true if private is present
if (req.http.Cache-Control[private] is not bool) => true if private= is present
if ("private=4"[private] is string) => false
if ("private=*true"[private] is not bool) => false
if ("private"[private]) => bool (true)

if (exists(req.http.Cache-Control[private]))
if (nonzero(req.http.Cache-Control[private]))

if (req.http.Cache-Control[private])
if (req.http.Cache-Control[private].exists)
if (req.http.Cache-Control[private].val)
if (req.http.Cache-Control[private].type)
  "private" => True, None, Bool
  "private=" => True, "", String
  "private=X" => True, "X", String
  "private=43" => True, 43, Number
  "" => False, None, None

if (req.http.Cache-Control[private]) { //exist

     if (req.http.Cache-Control[private].value) { //exist && has value


     }

}

==> PHK to write up a strawman

Specifically on issue 3844:

     * HEADER == STRING should be equivalent to STRING == HEADER (it is)
     * implicit HEADER to BOOL conversion conveys existence (it does)

---
varnishtest cmp

varnish v1 -vcl {
         backend be none;

         sub vcl_recv {
                 return (synth(200));
         }

         sub vcl_synth {
                 set resp.http.nohdr-eq-empty = req.http.nohdr == "";
                 set resp.http.empty-eq-nohdr = "" == req.http.nohdr;
         }
} -start

client c1 {
         txreq
         rxresp
         expect resp.http.nohdr-eq-empty == true
         expect resp.http.empty-eq-nohdr == true
} -start
---

BUGWASH & NEXT VDD
==================

Does it work as is?
- basically yes
- but we want to set up IRC & web bridge with simple basic auth
   - should keep logs

@slink to ping @theis about VUG

VUG needs to be a varnish-cache (!) thing
find a sponsor for a venue
VS, UPLEX, fastly, whoever, are welcome as sponsors

10 days after release
idea: 26./27. september

VDD: combine with VUG

sub vcl_connect / vcl_accept / vcl_client_hello
================================================
sub vcl_init {
         mytls = new newtls();
         mytls.load("/etc/tls/.....pem");
         mytls.load("/etc/tls/.....pem");
}
sub vcl_client_hello {
         mytls.dynload(tls.sni, ttl=10m);  # Dynamicly load cert?

         if (tls.sni == "legacysite.asas") {
                 set tls.ciphers = "ssl, tlsv3";
                 tls.hint = mytls;
         }

         set tls.ciphers = "sslv1, tlsv3";
         set tls.alpn = "h1,h2,h3=....";
         tls.hint = fallback;
         return(OK);
}

also: vcl_h[23]_settings, proxy2_tlv
----------------------

VEXT client_ssl has vmod tls

sub vcl_init {
     new something = tls.init(...)
     something.callback(myfunc)
}
func myfunc(tls.type: arg) {
     arg.crytomumblemumble = "bla";
     return(tls.OK)
}
-----------------------

sub vcl_accept { // Probably not needed.
     if (...) {
         filter.use();
     }
}


modifying the response body (thi.js)
================================


sub vcl_backend_response {
     set beresp.filter_list += "gunzip vmodfoo.spot_foo";
}


sub vcl_backend_body_is_here {
     if (vmodfoo.did_you_see_foo) {
         if (! beresp.body.try_filter({"regsuball(...) gzip"})) {
           return (deliver); # keep the "source" object
         }
         beresp.body.filter("...");
     }
     // trailers goes here
}

h2/h3 on the backend, for backend not understanding h1.
=======================================================

we can have h2 in-core, but vext is maybe also viable option.

AGENDA DAY2
===========

h3
add vext_path -> probably just PR it
-- 

** * * UPLEX - Nils Goroll Systemoptimierung

Scheffelstraße 32
22301 Hamburg

tel +49 40 28805731
mob +49 170 2723133
fax +49 40 42949753

xmpp://slink@jabber.int.uplex.de/

http://uplex.de/


More information about the varnish-dev mailing list