global backaddr is assumed to be static

[yogo1212]

hi :-)

when configuring a hostname for a backend the hostname is only resolved once in init_globals.
=>
in a way, hitch acts as a DNS cache that doesn't respect TTL.

related use case: using amazon ELB as backend.

possible solution: register an event to refresh the cached value of backaddr when the DNS value expires.

[olivierHa]

Any updates about this issue ? Is this also affecting docker internal name resolution ?

[yogo1212]

@olivierHa
the issue i've described only affects hitch internally.
e.g.

start hitch -> A record with TTL 20 received
one minute passes
DNS changes, server goes down
hitch loses connection
hitch retries using the expired address until something stops it.

we've replaced hitch with nginx in our application

[dmatetelki]

I've started to work on a periodic backend IP check fix (with the guidance of @daghf ), let me update the progress here.

[rtkbkish]

@dmatetelki Thanks. This will be very helpful.

[gquintard]

I'm interested by the approach here. First thing is when do we update the pool, and the second one is: how do we spread connections? Last one is what do we do in case of failure?

[dmatetelki]

Hi @rtkbkish, @gquintard, @yogo1212 and @daghf !

The first version is ready to be tested. (thou i did create a testcase)
The patch:

• introduces a new parameter, -R --backendrefresh=USECS which allows periodic backend address lookup. You turn it off / have static backend address with value 0. Default is 1s.
• The main process refreshes (or not) the address, and notify the client processes.
• The client processes do the create socket and connect now in end_handshake, not in 2 functions, for concurrency safety.