ngx_http_proxy_connect_moduleis likely better if you don't want to cache (Polipo2's cache is an integral part of its operation and cannot be disabled; you could also set up Tinyproxy as a non-caching proxy but it might be less reliable than nginx).
Polipo2 is a 2017 fork of Juliusz Chroboczek's "Polipo" [octopus] after he ceased maintaining the original in 2016.
Polipo2 is not intended for general client-side use---as Juliusz said when he stepped down, caching proxies are becoming obsolete for general use on the client side due to the increasing prevalence of encrypted alternatives to HTTP that reduce caching proxies to simple relays. If all you want is a simple relay (for example so your Web traffic originates from a remote IP address), then you can do better by using a VPN or a SOCKS5 proxy.
Polipo2 is now intended as a drop-in caching layer for experimental HTTP proxies such as Web Adjuster (see Adjuster's --upstream-proxy option; I suggest increasing polipo2's serverSlots and serverMaxSlots when used with parentProxy=localhost:8124).
Polipo2 is available on GitHub, and on a suitably well-provisioned GNU/Linux system you can do:
git clone https://github.com/ssb22/polipo2.git cd polipo2 make sudo make installthe last step being optional (you can run polipo2 directly from your user account, e.g. ./polipo2 logFile=/tmp/polipo2.log pidFile=/tmp/polipo2.pid diskCacheRoot=""; if you are using it with Web Adjuster's --upstream-proxy=:8123 you should also add parentProxy=localhost:8124 to the Polipo2 options, and consider serverSlots=256 and serverMaxSlots=256).
poll()system call (like 4.2BSD's
select()without the 1024-socket limit) to monitor its in-progress connections from a single thread. It does not use the more advanced
kqueuemechanism that modern versions of large-scale proxies like Squid can use (and that Tornado and hence Web Adjuster can use) on the GNU/Linux and BSD platforms.
The problem with
poll() is, it can monitor a huge number of connections at once to tell your code when one or more of them needs processing, but it doesn't tell you which ones need this processing. So when
poll() returns, Polipo2 has to spend CPU time looping through all of its open connections to see what needs doing. By contrast, Linux's
epoll and BSD's
kqueue can point your code directly at the connections that need attention, eliminating that loop.
This is not an issue if you have only a few dozen connections going at once, but once you're in the tens of thousands, you will notice a CPU holdup from those tens of thousands of extra checks that have to be done every time anything happens on any connection!
If that's your situation, I'd recommend switching to Squid, which is more scalable. I currently have no plans to upgrade Polipo2's
poll() into an
kqueue, as Polipo2 is intended for small-scale experimental use (so if you're getting big, "bite the bullet" and install Squid).
http_access allow localhost
cache_peer 127.0.0.1 parent 3129 0 no-query no-digest
never_direct allow all
dead_peer_timeout 99 seconds
access_log none localhost
cache_mem 256 MB
cache_dirdirective if you also want a disk cache)
SIGUSR2(discard objects) to the running Polipo2, but if the machine is expected to stay up, it's likely easier to run in RAM+swap (by setting
diskCacheRoot="") and then the size can be constrained more accurately (defaults to 25% of RAM; see note on differences with Polipo below).
On the other hand, a non-expiring disk cache is a useful option if you wish to collect a corpus of material from a site as users browse it (without needing to run a 'crawler' which might annoy the site); Polipo2's file format is quite easy for other programs to read. Obviously you'd have to respect the copyright on the resulting material.
chunkHighMarknow defaults to 25% of the physical RAM even on machines above 96M (its default is not limited to 24M as it is when
objectHighMarkdefaults to 2048 objects for every 24M in
chunkHighMark. You can still override these of course (remember
chunkHighMarkis in bytes).